arxiv_id
stringlengths 9
12
| abstract
stringlengths 431
13.8k
|
|---|---|
2406.09415 | This work does not introduce a new method. Instead, we present an interesting finding that questions the necessity of the inductive bias –localityin modern computer vision architectures. Concretely, we find that vanilla Transformers can operate by directly treating each individual pixel as a token and achieve highly performant results. This is substantially different from the popular design in Vision Transformer, which maintains the inductive bias from ConvNets towards local neighborhoods (e.g.by treating each 1616 patch as a token) . We mainly showcase the effectiveness of pixels-as-tokens across three well-studied tasks in computer vision: supervised learning for object classification, self-supervised learning via masked autoencoding, and image generation with diffusion models. Although directly operating on individual pixels is less computationally practical, we believe the community must be aware of this surprising piece of knowledge when devising the next generation of neural architectures for computer vision. |
2406.07550 | Recent advancements in generative models have highlighted the crucial role of image tokenization in the efficient synthesis of high-resolution images. Tokenization, which transforms images into latent representations, reduces computational demands compared to directly processing pixels and enhances the effectiveness and efficiency of the generation process. Prior methods, such as VQGAN, typically utilize 2D latent grids with fixed downsampling factors. However, these 2D tokenizations face challenges in managing the inherent redundancies present in images, where adjacent regions frequently display similarities.
To overcome this issue, we introduceTransformer-based 1-DimensionalTokenizer (TiTok) , an innovative approach that tokenizes images into 1D latent sequences.
TiTok provides a more compact latent representation, yielding substantially more efficient and effective representations than conventional techniques. For example, aimage can be reduced to just32discrete tokens, a significant reduction from the 256 or 1024 tokens obtained by prior methods. Despite its compact nature, TiTok achieves competitive performance to state-of-the-art approaches. Specifically, using the same generator framework, TiTok attains1.97gFID, outperforming MaskGIT baseline significantly by 4.21 at ImageNetbenchmark. The advantages of TiTok become even more significant when it comes to higher resolution. At ImageNetbenchmark, TiTok not only outperforms state-of-the-art diffusion model DiT-XL/2 (gFID 2.74vs.3.04) , but also reduces the image tokens by, leading tofastergeneration process. Our best-performing variant can significantly surpasses DiT-XL/2 (gFID2.13vs.3.04) while still generating high-quality samplesfaster. |
2405.18870 | This paper examines the extent to which large language models (LLMs) have developed higher-order theory of mind (ToM) ; the human ability to reason about multiple mental and emotional states in a recursive manner (e.g. Ithinkthat youbelievethat sheknows) . This paper builds on prior work by introducing a handwritten test suite – Multi-Order Theory of Mind Q&A – and using it to compare the performance of five LLMs to a newly gathered adult human benchmark. We find that GPT-4 and Flan-PaLM reach adult-level and near adult-level performance on ToM tasks overall, and that GPT-4 exceeds adult performance on 6th order inferences. Our results suggest that there is an interplay between model size and finetuning for the realisation of ToM abilities, and that the best-performing LLMs have developed a generalised capacity for ToM. Given the role that higher-order ToM plays in a wide range of cooperative and competitive human behaviours, these findings have significant implications for user-facing LLM applications. |
2405.20519 | Large language models generate code one token at a time. Their autoregressive generation process lacks the feedback of observing the program’s output. Training LLMs to suggest edits directly can be challenging due to the scarcity of rich edit data. To address these problems, we propose neural diffusion models that operate on syntax trees of any context-free grammar. Similar to image diffusion models, our method also inverts “noise” applied to syntax trees. Rather than generating code sequentially, we iteratively edit it while preserving syntactic validity, which makes it easy to combine this neural model with search. We apply our approach to inverse graphics tasks, where our model learns to convert images into programs that produce those images. Combined with search, our model is able to write graphics programs, see the execution result, and debug them to meet the required specifications. We additionally show how our system can write graphics programs for hand-drawn sketches. Video results can be found athttps://tree-diffusion.github.io. |
2406.09279 | Learning from preference feedback has emerged as an essential step for improving the generation quality and performance of modern language models (LMs) .
Despite its widespread use, the way preference-based learning is applied varies wildly, with differing data, learning algorithms, and evaluations used, making disentangling the impact of each aspect difficult.
In this work, we identify four core aspects of preference-based learning:preference data,learning algorithm,reward model, andpolicy training prompts, systematically investigate the impact of these components on downstream model performance, and suggest a recipe for strong learning for preference feedback. Our findings indicate that all aspects are important for performance, with better preference data leading to the largest improvements, followed by the choice of learning algorithm, the use of improved reward models, and finally the use of additional unlabeled prompts for policy training. Notably, PPO outperforms DPO by up to 2.5% in math and 1.2% in general domains. High-quality preference data leads to improvements of up to 8% in instruction following and truthfulness. Despite significant gains of up to 5% in mathematical evaluation when scaling up reward models, we surprisingly observe marginal improvements in other categories. We publicly release the code used for training111https://github.com/hamishivi/EasyLMand evaluating222https://github.com/allenai/open-instructour models, along with the models and datasets themselves333https://huggingface.co/collections/allenai/tulu-v25-suite-66676520fd578080e126f618. |
2406.07522v1 | Efficiently modeling sequences with infinite context length has been a long-standing problem. Past works suffer from either the quadratic computation complexity or the limited extrapolation ability on length generalization. In this work, we presentSamba, a simple hybrid architecture that layer-wise combines Mamba, a selective State Space Model (SSM) , with Sliding Window Attention (SWA) .Sambaselectively compresses a given sequence into recurrent hidden states while still maintaining the ability to precisely recall memories with the attention mechanism.
We scaleSambaup to 3.8B parameters with 3.2T training tokens and show thatSambasubstantially outperforms the state-of-the-art models based on pure attention or SSMs on a wide range of benchmarks. When trained on 4K length sequences,Sambacan be efficiently extrapolated to 256K context length with perfect memory recall and show improved token predictions up to 1M context length. As a linear-time sequence model,Sambaenjoys ahigher throughput compared to Transformers with grouped-query attention when processing user prompts of 128K length, andspeedup when generating 64K tokens with unlimited streaming. A sample implementation ofSambais publicly available inhttps://github.com/microsoft/Samba. |
2310.01801v3 | In this study, we introduce adaptive KV cache compression, a plug-and-play method that reduces the memory footprint of generative inference for Large Language Models (LLMs) .
Different from the conventional KV cache that retains key and value vectors for all context tokens, we conduct targeted profiling to discern the intrinsic structure of attention modules.
Based on the recognized structure, we propose FastGen, which constructs the KV cache in an adaptive manner: evicting long-range contexts on attention heads emphasizing local contexts, discarding non-special tokens on attention heads centered on special tokens, and only employing the standard KV cache for attention heads that broadly attend to all tokens.
Moreover, with the lightweight attention profiling used to guide the construction of the adaptive KV cache, FastGen can be deployed without resource-intensive fine-tuning or re-training.
In our experiments across various asks, FastGen demonstrates substantial reduction on GPU memory consumption with negligible generation quality loss.
We will release our code and the compatible CUDA kernel for reproducibility. |
2406.08414 | Offline preference optimization is a key method for enhancing and controlling the quality of Large Language Model (LLM) outputs.
Typically, preference optimization is approached as an offline supervised learning task using manually-crafted convex loss functions. While these methods are based on theoretical insights, they are inherently constrained by human creativity, so the large search space of possible loss functions remains under explored.
We address this by performing LLM-drivenobjective discoveryto automatically discover new state-of-the-art preference optimization algorithms without (expert) human intervention.
Specifically, we iteratively prompt an LLM to propose and implement new preference optimization loss functions based on previously-evaluated performance metrics. This process leads to the discovery of previously-unknown and performant preference optimization algorithms.
The best performing of these we callDiscovered Preference Optimization(DiscoPOP) 111Code:https://github.com/luchris429/DiscoPOP., a novel algorithm that adaptively blends logistic and exponential losses. Experiments demonstrate the state-of-the-art performance of DiscoPOP and its successful transfer to held-out tasks. |
2403.02241 | Our understanding of the generalization capabilities of neural networks (NNs) is still incomplete.
Prevailing explanations are based on implicit biases of gradient descent (GD)
but they cannot account for
the capabilities of models from gradient-free methodsnor the simplicity bias recently observed inuntrainednetworks.
This paper seeks other sources of generalization in NNs. Findings.To understand the inductive biases provided by architectures independently from GD, we examine untrained, random-weight networks.
Even simple MLPs show strong inductive biases:
uniform sampling in weight space yields a very biased distribution of functions in terms of complexity.
But unlike common wisdom, NNs do not have an inherent “simplicity bias”.
This property depends on components such as ReLUs, residual connections, and layer normalizations.
Alternative architectures can be built with a bias for any level of complexity.
Transformers also inherit all these properties from their building blocks. Implications.We provide a fresh explanation for the success of deep learning independent from gradient-based training.
It points at promising avenues for controlling the solutions implemented by trained models. |
2406.07887 | Selective state-space models (SSMs) like Mambaovercome some of the shortcomings of Transformers, such as quadratic computational complexity with sequence length and large inference-time memory requirements from the key-value cache.
Moreover, recent studies have shown that SSMs can match or exceed the language modeling capabilities of Transformers, making them an attractive alternative.
In a controlled setting (e.g., same training data) , however, studies so far have only presented small scale experiments (training with3B parameters and1T tokens) comparing SSMs to equivalent Transformers.
To understand the strengths and weaknesses of these architectures at larger scales, we present a direct comparison between 8B-parameter Mamba, Mamba-2, and Transformer models trained on the same datasets of up to 3.5T tokens.
We also compare these models to an 8B-parameter hybrid architecture consisting of 43% Mamba-2, 7% self-attention, and 50% MLP layers (Mamba-2-Hybrid) .
Using a diverse set of natural language tasks, we answer the important question of whether Mamba models can match their Transformer counterparts at larger training budgets.
Our results show that while pure SSM-based models match or exceed Transformers on many tasks, both Mamba and Mamba-2 models lag behind Transformer models on tasks which require strong copying or in-context learning abilities (e.g., five-shot MMLU, Phonebook Lookup) or long-context reasoning.
In contrast, we find that the 8B-parameter Mamba-2-Hybrid exceeds the 8B-parameter Transformer on all 12 standard tasks we evaluated (+2.65 points on average) and is predicted to be up to 8faster when generating tokens at inference time.
To validate long-context capabilities, we provide additional experiments evaluating variants of the Mamba-2-Hybrid and Transformer extended to support 16K, 32K, and 128K sequence lengths.
On an additional 23 long-context tasks, the hybrid model continues to closely match or exceed the Transformer on average.
To enable further study, we release the checkpoints as well as the code used to train our SSM-based models as part of NVIDIA’s Megatron-LM project (https://github.com/NVIDIA/Megatron-LM) 111A fixed snapshot of the code used in this technical report is available athttps://github.com/NVIDIA/Megatron-LM/tree/ssm/examples/mamba.. |
2406.08391 | When using large language models (LLMs) in high-stakes applications, we need to know when we can trust their predictions. Some works argue that prompting high-performance LLMs is sufficient to produce calibrated uncertainties, while others introduce sampling methods that can be prohibitively expensive. In this work, we first argue that prompting on its own is insufficient to achieve good calibration and then show that fine-tuning on a small dataset of correct and incorrect answers can create an uncertainty estimate with good generalization and small computational overhead. We show that a thousand graded examples are sufficient to outperform baseline methods and that training through the features of a model is necessary for good performance and tractable for large open-source models when using LoRA. We also investigate the mechanisms that enable reliable LLM uncertainty estimation, finding that many models can be used as general-purpose uncertainty estimators, applicable not just to their own uncertainties but also the uncertainty of other models. Lastly, we show that uncertainty estimates inform human use of LLMs in human-AI collaborative settings through a user study. |
2311.04254v3 | Recent advancements in Large Language Models (LLMs) have revolutionized decision-making by breaking down complex problems into more manageable language sequences referred to as “thoughts”. An effective thought design should consider three key perspectives: performance, efficiency, and flexibility. However, existing thought can at most exhibit two of these attributes. To address these limitations, we introduce a novel thought prompting approach called “Everything of Thoughts” (XoT) to defy the law of “Penrose triangle” of existing thought paradigms.XoTleverages pretrained reinforcement learning and Monte Carlo Tree Search (MCTS) to incorporate external domain knowledge into thoughts, thereby enhancing LLMs’ capabilities and enabling them to generalize to unseen problems efficiently. Through the utilization of the MCTS-LLM collaborative thought revision framework, this approach autonomously produces high-quality comprehensive cognitive mappings with minimal LLM interactions. Additionally,XoTempowers LLMs to engage in unconstrained thinking, allowing for flexible cognitive mappings for problems with multiple solutions. We evaluateXoTon several challenging problem-solving tasks, including Game of 24, 8-Puzzle, and Pocket Cube. Our results demonstrate thatXoTsignificantly outperforms existing approaches in various dimensions, showcasing its remarkable proficiency in addressing complex problems across diverse domains. |
2406.07394 | This paper introduces the MCT Self-Refine (MCTSr) algorithm, an innovative integration of Large Language Models (LLMs) with Monte Carlo Tree Search (MCTS) , designed to enhance performance in complex mathematical reasoning tasks. Addressing the challenges of accuracy and reliability in LLMs, particularly in strategic and mathematical reasoning, MCTSr leverages systematic exploration and heuristic self-refine mechanisms to improve decision-making frameworks within LLMs. The algorithm constructs a Monte Carlo search tree through iterative processes of Selection, self-refine, self-evaluation, and Backpropagation, utilizing an improved Upper Confidence Bound (UCB) formula to optimize the exploration-exploitation balance. Extensive experiments demonstrate MCTSr’s efficacy in solving Olympiad-level mathematical problems, significantly improving success rates across multiple datasets, including GSM8K, GSM Hard, MATH, and Olympiad-level benchmarks, including Math Odyssey, AIME, and OlympiadBench. The study advances the application of LLMs in complex reasoning tasks and sets a foundation for future AI integration, enhancing decision-making accuracy and reliability in LLM-driven applications. Codes publicly accessible atgithub.com/trotsky1997/MathBlackBox. |
2406.04329 | Masked (or absorbing) diffusion is actively explored as an alternative to autoregressive models for generative modeling of discrete data.
However, existing work in this area has been hindered by unnecessarily complex model formulations and unclear relationships between different perspectives, leading to suboptimal parameterization, training objectives, and ad hoc adjustments to counteract these issues.
In this work, we aim to provide a simple and general framework that unlocks the full potential of masked diffusion models.
We show that the continuous-time variational objective of masked diffusion models is a simple weighted integral of cross-entropy losses. Our framework also enables training generalized masked diffusion models with state-dependent masking schedules. When evaluated by perplexity, our models trained on OpenWebText surpass prior diffusion language models at GPT-2 scale and demonstrate superior performance on 4 out of 5 zero-shot language modeling tasks.
Furthermore, our models vastly outperform previous discrete diffusion models on pixel-level image modeling, achieving 2.78 (CIFAR-10) and 3.42 (ImageNet 6464) bits per dimension that are comparable or better than autoregressive models of similar sizes. |
2406.06563 | In this technical report, we introduce the training methodologies implemented in the development of Skywork-MoE, a high-performance mixture-of-experts (MoE) large language model (LLM) with 146 billion parameters and 16 experts. It is initialized from the pre-existing dense checkpoints of our Skywork-13B model. We explore the comparative effectiveness of upcycling versus training from scratch initializations. Our findings suggest that the choice between these two approaches should consider both the performance of the existing dense checkpoints and the MoE training budget.
We highlight two innovative techniques: gating logit normalization, which improves expert diversification, and adaptive auxiliary loss coefficients, allowing for layer-specific adjustment of auxiliary loss coefficients. Our experimental results validate the effectiveness of these methods.
Leveraging these techniques and insights, we trained our upcycled Skywork-MoE on a condensed subset of our SkyPile corpus. The evaluation results demonstrate that our model delivers strong performance across a wide range of benchmarks. |
2406.07524 | While diffusion models excel at generating high-quality images, prior work reports a significant performance gap between diffusion and autoregressive (AR) methods in language modeling.
In this work, we show that simple masked discrete diffusion is more performant than previously thought.
We apply an effective training recipe that improves the performance of masked diffusion models and derive a simplified, Rao-Blackwellized objective that results in additional improvements.
Our objective has a simple form—it is a mixture of classical masked language modeling losses—and can be used to train encoder-only language models that admit efficient samplers, including ones that can generate arbitrary lengths of text semi-autoregressively like a traditional language model.
On language modeling benchmarks, a range of masked diffusion models trained with modern engineering practices achieves a new state-of-the-art among diffusion models, and approaches AR perplexity. We release our code at:https://github.com/kuleshov-group/mdlm |
2405.14782 | Effective evaluation of language models remains an open challenge in NLP. Researchers and engineers face methodological issues such as the sensitivity of models to evaluation setup, difficulty of proper comparisons across methods, and the lack of reproducibility and transparency. In this paper we draw on three years of experience in evaluating large language models to provide guidance and lessons for researchers. First, we provide an overview of common challenges faced in language model evaluation. Second, we delineate best practices for addressing or lessening the impact of these challenges on research. Third, we present theLanguage Model Evaluation Harness(lm-eval) : an open source library for independent, reproducible, and extensible evaluation of language models that seeks to address these issues. We describe the features of the library as well as case studies in which the library has been used to alleviate these methodological concerns. |
2310.04799 | With the advancements in conversational AI, such as ChatGPT, this paper focuses on exploring developing Large Language Models (LLMs) for non-English languages, especially emphasizing alignment with human preferences. We introduce a computationally efficient method, leveraging “chat vector,” to synergize pre-existing knowledge and behaviors in LLMs, restructuring the conventional training paradigm from continual pre-trainSFTRLHF to continual pre-train + chat. Our empirical studies, primarily focused on Traditional Chinese, employ LLaMA2 as the base model and acquire the chat vector by subtracting the pre-trained weights, LLaMA2, from the weights of LLaMA2-chat. Evaluating from three distinct facets, which are toxicity, ability of instruction following, and multi-turn dialogue demonstrates the chat vector’s superior efficacy in “chatting”. To confirm the adaptability of our approach, we extend our experiments to include models pre-trained in both Korean and Simplified Chinese, illustrating the versatility of our methodology. Overall, we present a significant solution in aligning LLMs with human preferences efficiently across various languages, accomplished by the chat vector. |
2406.04313 | AI systems can take harmful actions and are highly vulnerable to adversarial attacks.
We present an approach, inspired by recent advances in representation engineering, that interrupts the models as they respond with harmful outputs with “circuit breakers.”
Existing techniques aimed at improving alignment, such as refusal training, are often bypassed. Techniques such as adversarial training try to plug these holes by countering specific attacks. As an alternative to refusal training and adversarial training, circuit-breaking directly controls the representations that are responsible for harmful outputs in the first place.
Our technique can be applied to both text-only and multimodal language models to prevent the generation of harmful outputs without sacrificing utility—even in the presence of powerful unseen attacks.
Notably, while adversarial robustness in standalone image recognition remains an open challenge, circuit breakers allow the larger multimodal system to reliably withstand image “hijacks” that aim to produce harmful content.
Finally, we extend our approach to AI agents, demonstrating considerable reductions in the rate of harmful actions when they are under attack.
Our approach represents a significant step forward in the development of reliable safeguards to harmful behavior and adversarial attacks. Code is available atgithub.com/blackswan-ai/circuit-breakers. |
2310.20624 | Content Warning:This document contains content that some may find disturbing or offensive, including content that is hateful or violent in nature.AI developers often apply safety alignment procedures to prevent the misuse of their AI systems. For example, before Meta released Llama 2-Chat—a collection of instruction fine-tuned large language models—they invested heavily in safety training, incorporating extensive red-teaming and reinforcement learning from human feedback. However, it remains unclear how well safety training guards against model misuse when attackers have access to model weights.
We explore the robustness of safety training in language models by subversively fine-tuning the public weights of Llama 2-Chat.
We employ low-rank adaptation (LoRA) as an efficient fine-tuning method. With a budget of less than $200 per model and using only one GPU, we successfully undo the safety training of Llama 2-Chat models of sizes 7B, 13B, and 70B. Specifically, our fine-tuning technique significantly reduces the rate at which the model refuses to follow harmful instructions. We achieve arefusal ratebelow 1% for our 70B Llama 2-Chat model on two refusal benchmarks. Our fine-tuning method retains general performance, which we validate by comparing our fine-tuned models against Llama 2-Chat across two benchmarks. Additionally, we present a selection of harmful outputs produced by our models.
While there is considerable uncertainty about the scope of risks from current models, it is likely that future models will have significantly more dangerous capabilities, including the ability to hack into critical infrastructure, create dangerous bio-weapons, or autonomously replicate and adapt to new environments. We show that subversive fine-tuning is practical and effective, and hence argue that evaluating risks from fine-tuning should be a core part of risk assessments for releasing model weights. |
2406.03618 | Large Language Models (LLMs) often do not perform well on queries that require the aggregation of information across texts. To better evaluate this setting and facilitate modeling efforts, we introduce TACT—Text And Calculations through Tables, a dataset crafted to evaluate LLMs’ reasoning and computational abilities using complex instructions. TACT contains challenging instructions that demand stitching information scattered across one or more texts, and performing complex integration on this information to generate the answer. We construct this dataset by leveraging an existing dataset of texts and their associated tables. For each such tables, we formulate new queries, and gather their respective answers. We demonstrate that all contemporary LLMs perform poorly on this dataset, achieving an accuracy below 38%. To pinpoint the difficulties and thoroughly dissect the problem, we analyze model performance across three components: table-generation, Pandas command-generation, and execution. Unexpectedly, we discover that each component presents substantial challenges for current LLMs. These insights lead us to propose a focused modeling framework, which we refer to asIE as a tool. Specifically, we propose to add “tools” for each of the above steps, and implement each such tool with few-shot prompting. This approach shows an improvement over existing prompting techniques, offering a promising direction for enhancing model capabilities in these tasks. The dataset is available attact-benchmark.github.io. |
2406.04391 | Predictable behavior from scaling advanced AI systems is an extremely desirable property. Although a well-established literature exists on how pretraining performance scales, the literature on how particular downstream capabilities scale is significantly muddier. In this work, we take a step back and ask: *why has predicting specific downstream capabilities with scale remained elusive?* While many factors are certainly responsible, we identify a new factor that makes modeling scaling behavior on widely used multiple-choice question-answering benchmarks challenging. Using five model families and twelve well-established multiple-choice benchmarks, we show that downstream performance is computed from negative log likelihoods via a sequence of transformations that progressively degrade the statistical relationship between performance and scale. We then reveal the mechanism causing this degradation: downstream metrics require comparing the correct choice against a small number of specific incorrect choices, meaning accurately predicting downstream capabilities requires predicting not just how probability mass concentrates on the correct choice with scale, but also how probability mass fluctuates on specific incorrect choices with scale. We empirically study how probability mass on the correct choice co-varies with probability mass on incorrect choices with increasing compute, suggesting that scaling laws for *incorrect* choices might be achievable. Our work also explains why pretraining scaling laws are commonly regarded as more predictable than downstream capabilities and contributes towards establishing scaling-predictable evaluations of frontier AI models. |
2406.04823 | This paper explores the in-context learning capabilities of masked language models, challenging the common view that this ability does not ‘emerge’ in them. We present anembarrassingly simpleinference technique that enables DeBERTa to operate as a generative model without any additional training. Our findings demonstrate that DeBERTa can match and even surpass GPT-3, its contemporary that famously introduced the paradigm of in-context learning. The comparative analysis reveals that the masked and causal language models behave very differently, as they clearly outperform each other on different categories of tasks. This suggests that there is great potential for a hybrid training approach that takes advantage of the strengths of both training objectives. |
2406.04770 | We introduce WildBench, an automated evaluation framework designed to benchmark large language models (LLMs) using challenging, real-world user queries.WildBenchconsists of 1,024 tasks carefully selected from over one million human-chatbot conversation logs.
For automated evaluation withWildBench, we have developed two metrics, WB-Reward and WB-Score, which are computable using advanced LLMs such as GPT-4-turbo.WildBenchevaluation uses task-specific checklists to evaluate model outputs systematically and provides structured explanations that justify the scores and comparisons, resulting in more reliable and interpretable automatic judgments.
WB-Reward employs fine-grained pairwise comparisons between model responses, generating five potential outcomes: much better, slightly better, slightly worse, much worse, or a tie.
Unlike previous evaluations that employed a single baseline model, we selected three baseline models at varying performance levels to ensure a comprehensive pairwise evaluation.
Additionally, we propose a simple method to mitigate length bias, by converting outcomes of “slightly better/worse” to “tie” if the winner response exceeds the loser one by more thancharacters.
WB-Score evaluates the quality of model outputs individually, making it a fast and cost-efficient evaluation metric.WildBenchresults demonstrate a strong correlation with the human-voted Elo ratings from Chatbot Arena on hard tasks. Specifically, WB-Reward achieves a Pearson correlation of 0.98 with top-ranking models. Additionally, WB-Score reaches 0.95, surpassing both ArenaHard’s 0.91 and AlpacaEval2.0’s 0.89 for length-controlled win rates, as well as the 0.87 for regular win rates. |
2311.08596 | The interactive nature of Large Language Models (LLMs) theoretically allows models to refine and improve their answers, yet systematic analysis of the multi-turn behavior of LLMs remains limited. In this paper, we propose the FlipFlop experiment: in the first round of the conversation, an LLM responds to a prompt containing a classification task. In a second round, the LLM is challenged with a follow-up phrase like “Are you sure?”, offering an opportunity for the model to reflect on its initial answer, and decide whether to confirm or flip its answer. A systematic study of nine LLMs on seven classification tasks reveals that models flip their answers on average 46% of the time and that all models see a deterioration of accuracy between their first and final prediction, with an average drop of 17%. The FlipFlop experiment illustrates the universality of sycophantic behavior in LLMs and provides a robust framework to analyze model behavior and evaluate potential solutions. |
2406.04692 | Recent advances in large language models (LLMs) demonstrate substantial capabilities in natural language understanding and generation tasks.
With the growing number of LLMs, how to harness the collective expertise of multiple LLMs is an exciting open direction.
Toward this goal, we propose a new approach that leverages the collective strengths of multiple LLMs through a Mixture-of-Agents (MoA) methodology.
In our approach, we construct a layered MoA architecture wherein each layer comprises multiple LLM agents.
Each agent takes all the outputs from agents in the previous layer as auxiliary information in generating its response.
MoA models achieves state-of-art performance on AlpacaEval 2.0, MT-Bench and FLASK, surpassing GPT-4 Omni. For example, our MoA using only open-source LLMs is the leader of AlpacaEval 2.0 by a substantial gap, achieving a score of 65.1% compared to 57.5% by GPT-4 Omni.111Our code can be found in:https://github.com/togethercomputer/moa. |
2402.08733 | Identifying how much a modelknows about the stochastic real-world processit was trained on is important to ensure it avoids producing incorrect or “hallucinated” answers or taking unsafe actions.
But this is difficult
for generative models
because
probabilistic predictions do not distinguish between per-response noise (aleatoric uncertainty) and lack of knowledge about the process (epistemic uncertainty) , and existing epistemic uncertainty quantification techniques tend to
be overconfident
when the model underfits.
We propose a general strategy for teaching a
model to both approximateand also estimate
the remaining gaps betweenand:
train it to predictpairsof independent responses drawn from the true conditional distribution, allow it to “cheat” by observing one response while predicting the other, then measure how much it cheats.
Remarkably, we prove that being good at cheating (i.e. cheating whenever it improves your prediction) is
equivalent to beingsecond-order calibrated, a principled extension of ordinary calibration that allows us to construct provably-correct frequentist confidence intervals forand detect incorrect responses with high probability.
We demonstrate empirically that our approach accurately estimates how much models don’t know across ambiguous image classification, (synthetic) language modeling, and partially-observable navigation tasks, outperforming existing techniques. |
2305.17190 | Multiplications are responsible for most of the computational cost involved in neural network training and inference.
Recent research has thus looked for ways to reduce the cost associated with them.
Inspired by() , we replace multiplication with a cheap piecewise affine approximation that is achieved by adding the bit representation of the floating point numbers together as integers.
We show that transformers can be trained with the resulting modified matrix multiplications on both vision and language tasks with little to no performance impact, and without changes to the training hyperparameters.
We further replace all non-linearities in the networks making them fully and jointly piecewise affine in both inputs and weights.
Finally, we show that we can eliminate all multiplications in the entire training process, including operations in the forward pass, backward pass and optimizer update, demonstrating the first successful training of modern neural network architectures in a fully multiplication-free fashion. |
2310.13548 | Human feedback is commonly utilized to finetune AI assistants. But human feedback may also encourage model responses that match user beliefs over truthful ones, a behaviour known as sycophancy. We investigate the prevalence of sycophancy in models whose finetuning procedure made use of human feedback, and the potential role of human preference judgments in such behavior. We first demonstrate that five state-of-the-art AI assistants consistently exhibit sycophancy across four varied free-form text-generation tasks. To understand if human preferences drive this broadly observed behavior, we analyze existing human preference data. We find that when a response matches a user’s views, it is more likely to be preferred. Moreover, both humans and preference models (PMs) prefer convincingly-written sycophantic responses over correct ones a non-negligible fraction of the time. Optimizing model outputs against PMs also sometimes sacrifices truthfulness in favor of sycophancy. Overall, our results indicate that sycophancy is a general behavior of state-of-the-art AI assistants, likely driven in part by human preference judgments favoring sycophantic responses. |
2406.04267 | We study how information propagates in decoder-only Transformers, which are the architectural backbone of most existing frontier large language models (LLMs) . We rely on a theoretical signal propagation analysis—specifically, we analyse the representations of the last token in the final layer of the Transformer, as this is the representation used for next-token prediction. Our analysis reveals arepresentational collapsephenomenon: we prove that certain distinct sequences of inputs to the Transformer can yield arbitrarily close representations in the final token. This effect is exacerbated by the low-precision floating-point formats frequently used in modern LLMs. As a result, the model is provably unable to respond to these sequences in different ways—leading to errors in, e.g., tasks involving counting or copying. Further, we show that decoder-only Transformer language models can lose sensitivity to specific tokens in the input, which relates to the well-known phenomenon ofover-squashingin graph neural networks. We provide empirical evidence supporting our claims on contemporary LLMs. Our theory also points to simple solutions towards ameliorating these issues. |
2406.02528 | Matrix multiplication (MatMul) typically dominates the overall computational cost of large language models (LLMs) . This cost only grows as LLMs scale to larger embedding dimensions and context lengths. In this work, we show that MatMul operations can be completely eliminated from LLMs while maintaining strong performance at billion-parameter scales. Our experiments show that our proposed MatMul-free models achieve performance on-par with state-of-the-art Transformers that require far more memory during inference at a scale up to at least 2.7B parameters. We investigate the scaling laws and find that the performance gap between our MatMul-free models and full precision Transformers narrows as the model size increases. We also provide a GPU-efficient implementation of this model which reduces memory usage by up to 61% over an unoptimized baseline during training. By utilizing an optimized kernel during inference, our model’s memory consumption can be reduced by more than 10compared to unoptimized models. To properly quantify the efficiency of our architecture, we build a custom hardware solution on an FPGA which exploits lightweight operations beyond what GPUs are capable of. We processed billion-parameter scale models at 13W beyond human readable throughput, moving LLMs closer to brain-like efficiency. This work not only shows how far LLMs can be stripped back while still performing effectively, but also points at the types of operations future accelerators should be optimized for in processing the next generation of lightweight LLMs. Our code implementation is available athttps://github.com/ridgerchu/matmulfreellm. |
2406.03476 | Pretraining datasets for large language models (LLMs) have grown to trillions of tokens composed of large amounts of CommonCrawl (CC) web scrape along with smaller, domain-specific datasets.
It is expensive to understand the impact of these domain-specific datasets on model capabilities as training at large FLOP scales is required to reveal significant changes to difficult and emergent benchmarks.
Given the increasing cost of experimenting with pretraining data, how does one determine the optimal balance between the diversity in general web scrapes and the information density of domain specific data?
In this work, we show how to leverage the smaller domain specific datasets by upsampling them relative to CC at the end of training to drive performance improvements on difficult benchmarks.
This simple technique allows us to improve up to 6.90 pp on MMLU, 8.26 pp on GSM8K, and 6.17 pp on HumanEval relative to the base data mix for a 7B model trained for 1 trillion (T) tokens, thus rivaling Llama-2 (7B) —a model trained for twice as long.
We experiment with ablating the duration of domain upsampling from 5% to 30% of training and find that 10% to 20% percent is optimal for navigating the tradeoff between general language modeling capabilities and targeted benchmarks.
We also use domain upsampling to characterize at scale the utility of individual datasets for improving various benchmarks by removing them during this final phase of training.
This tool opens up the ability to experiment with the impact of different pretraining datasets at scale, but at an order of magnitude lower cost compared to full pretraining runs. |
2405.16712 | In this technical report, we present Zamba, a novel 7B SSM-transformer hybrid model which achieves competitive performance against leading open-weight models at a comparable scale. Zamba is trained on 1T tokens from openly available datasets and is the best non-transformer model at this scale. Zamba pioneers a unique architecture combining a Mamba backbone with a single shared attention module, thus obtaining the benefits of attention at minimal parameter cost. Due to its architecture, Zamba is significantly faster at inference than comparable transformer models and requires substantially less memory for generation of long sequences. Zamba is pretrained in two phases: the first phase is based on existing web datasets, while the second one consists of annealing the model over high-quality instruct and synthetic datasets, and is characterized by a rapid learning rate decay. We open-source the weights and all checkpoints for Zamba, through both phase 1 and annealing phases. |
2404.19737 | Large language models such as GPT and Llama are trained with a next-token prediction loss.
In this work, we suggest that training language models to predictmultiplefuture tokens at once results in higher sample efficiency.
More specifically, at each position in the training corpus, we ask the model to predict the followingtokens usingindependent output heads, operating on top of a shared model trunk.
Considering multi-token prediction as an auxiliary training task, we measure improved downstream capabilities with no overhead in training time for both code and natural language models.
The method is increasingly useful for larger model sizes, and keeps its appeal when training for multiple epochs. Gains are especially pronounced ongenerativebenchmarks like coding, where our models consistently outperform strong baselines by several percentage points. Our 13B parameter models solves 12 % more problems on HumanEval and 17 % more on MBPP than comparable next-token models. Experiments on small algorithmic tasks demonstrate that multi-token prediction is favorable for the development of induction heads and algorithmic reasoning capabilities.
As an additional benefit, models trained with 4-token prediction are up tofaster at inference, even with large batch sizes. |
2405.18380 | The rapid advancements in Large Language Models (LLMs) have revolutionized various natural language processing tasks. However, the substantial size of LLMs presents significant challenges in training or fine-tuning. While parameter-efficient approaches such as low-rank adaptation (LoRA) have gained popularity, they often compromise performance compared to full-rank fine-tuning. In this paper, we proposeOutlier-weighed Layerwise Sampled Low-Rank Projection(OwLore) , a new memory-efficient fine-tuning approach, inspired by the layerwise outlier distribution of LLMs, which dynamically samples pre-trained layers to fine-tune instead of adding additional adaptors. We first interpret the outlier phenomenon through the lens of Heavy-Tailed Self-Regularization theory (HT-SR) , discovering that layers with more outliers tend to be more heavy-tailed and consequently better trained. Inspired by this finding, OwLore strategically assigns higher sampling probabilities to layers with more outliers to better leverage the knowledge stored in pre-trained LLMs. To further mitigate the memory demands of fine-tuning, we integrate gradient low-rank projection into our approach, which facilitates each layer to be efficiently trained in a low-rank manner. By incorporating the efficient characteristics of low-rank and optimal layerwise sampling, OwLore significantly improves the memory-performance trade-off in LLM pruning. Our extensive experiments across various architectures, including LLaMa2, LLaMa3, and Mistral, demonstrate that OwLore consistently outperforms baseline approaches, including full fine-tuning. Specifically, it achieves up to a 1.1% average accuracy gain on the Commonsense Reasoning benchmark, a 3.0% improvement on MMLU, and a notable 10% boost on MT-Bench, while being more memory efficient. OwLore allows us to fine-tune LLaMa2-7B with only 21GB of memory. Code is available athttps://github.com/pixeli99/OwLore. |
2405.21060v1 | While Transformers have been the main architecture behind deep learning's success in language modeling, state-space models (SSMs) such as Mamba have recently been shown to match or outperform Transformers at small to medium scale. We show that these families of models are actually quite closely related, and develop a rich framework of theoretical connections between SSMs and variants of attention, connected through various decompositions of a well-studied class of structured *semiseparable matrices*. Our state space duality (SSD) framework allows us to design a new architecture (**Mamba-2**) whose core layer is an a refinement of Mamba's selective SSM that is 2-8× faster, while continuing to be competitive with Transformers on language modeling. |
2402.04553 | We present a novel approach to accelerate stochastic gradient descent (SGD) by utilizing curvature information obtained from Hessian-vector products or finite differences of parameters and gradients, similar to the BFGS algorithm. Our approach involves two preconditioners: a matrix-free preconditioner and a low-rank approximation preconditioner. We update both preconditioners online using a criterion that is robust to stochastic gradient noise and does not require line search or damping. To preserve the corresponding symmetry or invariance, our preconditioners are constrained to certain connected Lie groups. The Lie group’s equivariance property simplifies the preconditioner fitting process, while its invariance property eliminates the need for damping, which is commonly required in second-order optimizers. As a result, the learning rate for parameter updating and the step size for preconditioner fitting are naturally normalized, and their default values work well in most scenarios. Our proposed approach offers a promising direction for improving the convergence ofSGDwith low computational overhead. We demonstrate that Preconditioned SGD (PSGD) outperforms SoTA on Vision, NLP, and RL tasks across multiple modern deep-learning architectures. We have provided code for reproducingtoyandlarge scale experimentsin this paper. |
2404.08819 | State-space models (SSMs) have emerged as a potential alternative architecture for building large language models (LLMs) compared to the previously ubiquitous transformer architecture. One theoretical weakness of transformers is that they cannot express certain kinds of sequential computation and state tracking, which SSMs are explicitly designed to address via their close architectural similarity to recurrent neural networks (RNNs) .But do SSMs truly have an advantage (over transformers) in expressive power for state tracking?Surprisingly, the answer is no. Our analysis reveals that the expressive power of SSMs is limited very similarly to transformers: SSMs cannot express computation outside the complexity class. In particular, this means they cannot solve simple state-tracking problems like permutation composition. It follows that SSMs are provably unable to accurately track chess moves with certain notation, evaluate code, or track entities in a long narrative. To supplement our formal analysis, we report experiments showing that Mamba-style SSMs indeed struggle with state tracking. Thus, despite its recurrent formulation, the “state” in an SSM is an illusion: SSMs have similar expressiveness limitations to non-recurrent models like transformers, which may fundamentally limit their ability to solve real-world state-tracking problems. |
2404.19553 | We extend the context length of Llama-3-8B-Instruct from 8K to 80K via QLoRA fine-tuning111The model is noted as Llama-3-8B-Instruct-80K-QLoRA given its max context length during fine-tuning. However, users could apply the model for even longer contexts via extrapolation.. The entire training cycle is super efficient, which takes 8 hours on one 8xA800 (80G) GPU machine. The resulted model exhibits superior performances across a broad range of evaluation tasks, such as NIHS, topic retrieval, and long-context language understanding; meanwhile, it also well preserves the original capability over short contexts. The dramatic context extension is mainly attributed to merely 3.5K synthetic training samples generated by GPT-4 , which indicates the LLMs’ inherent (yet largely underestimated) potential to extend its original context length. In fact, the context length could be extended far beyond 80K with more computation resources. Therefore, the team will publicly release the entire resources (including data, model, data generation pipeline, training code) so as to facilitate the future research from the community:https://github.com/FlagOpen/FlagEmbedding. |
2404.18424 | The current use of large language models (LLMs) for zero-shot document ranking follows one of two ways: 1) prompt-based re-ranking methods, which require no further training but are feasible for only re-ranking a handful of candidate documents due to the associated computational costs; and 2) unsupervised contrastive trained dense retrieval methods, which can retrieve relevant documents from the entire corpus but require a large amount of paired text data for contrastive training.
In this paper, we propose PromptReps, which combines the advantages of both categories: no need for training and the ability to retrieve from the whole corpus. Our method only requires prompts to guide an LLM to generate query and document representations for effective document retrieval. Specifically, we prompt the LLMs to represent a given text using a single word, and then use the last token’s hidden states and the corresponding logits associated to the prediction of the next token to construct a hybrid document retrieval system. The retrieval system harnesses both dense text embedding and sparse bag-of-words representations given by the LLM. Our experimental evaluation on the BEIR zero-shot document retrieval datasets illustrates that this simple prompt-based LLM retrieval method can achieve a similar or higher retrieval effectiveness than state-of-the-art LLM embedding methods that are trained with large amounts of unsupervised data, especially when using a larger LLM.111Code for fully reproducing the results is available athttps://github.com/ielab/PromptReps. |
2405.18144 | Second-order optimizers, maintaining a matrix termed a preconditioner, are superior to first-order optimizers in both theory and practice.
The states forming the preconditioner and its inverse root restrict the maximum size of models trained by second-order optimizers.
To address this, compressing 32-bit optimizer states to lower bitwidths has shown promise in reducing memory usage.
However, current approaches only pertain to first-order optimizers.
In this paper, we propose the first 4-bit second-order optimizers, exemplified by 4-bit Shampoo, maintaining performance similar to that of 32-bit ones.
We show that quantizing the eigenvector matrix of the preconditioner in 4-bit Shampoo is remarkably better than quantizing the preconditioner itself both theoretically and experimentally.
By rectifying the orthogonality of the quantized eigenvector matrix, we enhance the approximation of the preconditioner’s eigenvector matrix, which also benefits the computation of its inverse 4-th root.
Besides, we find that linear square quantization slightly outperforms dynamic tree quantization when quantizing second-order optimizer states.
Evaluation on various networks for image classification demonstrates that our 4-bit Shampoo achieves comparable test accuracy to its 32-bit counterpart while being more memory-efficient.
The source code will be made available. |
2405.15743 | Several challenges make it difficult for sparse neural networks to compete with dense models. First, setting a large fraction of weights
to zero impairs forward and gradient signal propagation.
Second, sparse studies often need to test multiple sparsity levels, while also introducing new hyperparameters (HPs) , leading to prohibitive tuning costs.
Indeed, the standard practice is to
re-use the learning HPs originally crafted for dense models. Unfortunately, we show sparse and
dense networks do not share the same optimal HPs.
Without stable dynamics and effective training recipes, it is
costly to test sparsity at scale, which is key to surpassing dense networks and making the business case for
sparsity acceleration in hardware. A holistic approach is needed to tackle these
challenges and we proposesparse maximal update parameterization(SµPar)
as one such approach.SµParensures activations, gradients, and weight updates all scale independently of sparsity level.
Further, by reparameterizing the HPs,SµParenables the same HP values to be optimal as we vary both sparsity level and model width.
HPs can be tuned on small dense networks and transferred to large sparse models, greatly reducing tuning costs.
On large-scale language modeling,SµPartraining improves loss by up to 8.2% over the common approach of using the dense model standard parameterization. |
2405.16661 | In recent years, large language models (LLMs) have had a dramatic impact on various sub-fields of AI, most notably on natural language understanding tasks. However, there is widespread agreement that the logical reasoning capabilities of contemporary LLMs are, at bestfragmentary(i.e., may work well on some problem instances but fail dramatically on others) . While traditional LLM fine-tuning approaches (e.g., those that use human feedback) do address this problem to some degree, they suffer from many issues, including unsound black-box reward models, difficulties in collecting preference data, and sparse scalar reward values. To address these challenges, we propose a new training/fine-tuning paradigm we refer to as Reinforcement Learning via Symbolic Feedback (RLSF) , which is aimed at enhancing the reasoning capabilities of LLMs. In the RLSF setting, the LLM that is being trained/fine-tuned is considered as the RL agent, while the environment is allowed access to reasoning or domain knowledge tools (e.g., solvers, algebra systems) . Crucially, in RLSF, these reasoning tools can provide feedback to the LLMs viapoly-sized certificates(e.g., proofs) , that characterize errors in the LLM-generated object with respect to some correctness specification. The ability of RLSF-based training/fine-tuning to leverage certificate-generating symbolic tools enables sound fine-grained (token-level) reward signals to LLMs, and thus addresses the limitations of traditional reward models mentioned above. Via extensive evaluations, we show that our RLSF-based fine-tuning of LLMs outperforms traditional approaches on two different applications, namely, program synthesis from natural language pseudo-code to programming language (C++) , and solving the Game of 24. For the former, RLSF-tuned LLMs show a significant improvement (e.g., +52.64% in compilation accuracy and +31.43% in functional correctness for Google’s CodeGemma-2b) compared to supervised fine-tuning. Further, CodeGemma-2b gets +34.82% in compilation accuracy and +17.01% in functional correctness compared to ChatGPT (100larger) . Similarly, for the Game of 24, we observed +25% success rate using Meta’s Llama2-7b compared to traditional methods, and +7% success rate compared to ChatGPT (25larger) . A takeaway is that fine-tuning via RLSF enables relatively smaller LLMs to significantly outperform orders of magnitude larger models (e.g., ChatGPT) . |
2106.05237 | There is a growing discrepancy in computer vision between large-scale models that achieve state-of-the-art performance and models that are affordable in practical applications. In this paper we address this issue and significantly bridge the gap between these two types of models. Throughout our empirical investigation we do not aim to necessarily propose a new method, but strive to identify a robust and effective recipe for making state-of-the-art large scale models affordable in practice. We demonstrate that, when performed correctly, knowledge distillation can be a powerful tool for reducing the size of large models without compromising their performance. In particular, we uncover that there are certain implicit design choices, which may drastically affect the effectiveness of distillation. Our key contribution is the explicit identification of these design choices, which were not previously articulated in the literature. We back up our findings by a comprehensive empirical study, demonstrate compelling results on a wide range of vision datasets and, in particular, obtain a state-of-the-art ResNet-50 model for ImageNet, which achieves 82.8% top-1 accuracy.222We provide code and models in thebig_visioncodebase. |
2312.05934 | Large language models (LLMs) encapsulate a vast amount of factual information within their pre-trained weights, as evidenced by their ability to answer diverse questions across different domains. However, this knowledge is inherently limited, relying heavily on the characteristics of the training data. Consequently, using external datasets to incorporate new information or refine the capabilities of LLMs on previously seen information poses a significant challenge. In this study, we compare two common approaches: unsupervised fine-tuning and retrieval-augmented generation (RAG) . We evaluate both approaches on a variety of knowledge-intensive tasks across different topics. Our findings reveal that while unsupervised fine-tuning offers some improvement, RAG consistently outperforms it, both for existing knowledge encountered during training and entirely new knowledge. Moreover, we find that LLMs struggle to learn new factual information through unsupervised fine-tuning, and that exposing them to numerous variations of the same fact during training could alleviate this problem. |
2405.18428 | Diffusion models with large-scale pre-training have achieved significant success in the field of visual content generation, particularly exemplified by Diffusion Transformers (DiT) . However, DiT models have faced challenges with scalability and quadratic complexity efficiency.
In this paper, we aim to leverage the long sequence modeling capability of Gated Linear Attention (GLA) Transformers, expanding its applicability to diffusion models.
We introduce Diffusion Gated Linear Attention Transformers (DiG) , a simple, adoptable solution with minimal parameter overhead, following the DiT design, but offering superior efficiency and effectiveness.
In addition to better performance than DiT, DiG-S/2 exhibitshigher training speed than DiT-S/2 and savesGPU memory at a resolution of.
Moreover, we analyze the scalability of DiG across a variety of computational complexity. DiG models, with increased depth/width or augmentation of input tokens, consistently exhibit decreasing FID.
We further compare DiG with other subquadratic-time diffusion models.
With the same model size, DiG-XL/2 isfaster than the recent Mamba-based diffusion model at aresolution, and isfaster than DiT with CUDA-optimized FlashAttention-2 under theresolution.
All these results demonstrate its superior efficiency among the latest diffusion models. |
2405.15613 | Self-supervised features are the cornerstone of modern machine learning systems. They are typically pre-trained on data collections whose construction and curation typically require extensive human effort. This manual process has some limitations similar to those encountered in supervised learning, e.g.,
the crowd-sourced selection of data is costly and time-consuming, preventing scaling the dataset size. In this work, we consider the problem of automatic curation of high-quality datasets for self-supervised pre-training. We posit that such datasets should be large, diverse and balanced, and propose a clustering-based approach for building ones satisfying all these criteria. Our method involves successive and hierarchical applications of-means on a large and diverse data repository to obtain clusters that distribute uniformly among data concepts, followed by a hierarchical, balanced sampling step from these clusters. Extensive experiments on three different data domains including web-based images, satellite images and text show that features trained on our automatically curated datasets outperform those trained on uncurated data while being on par or better than ones trained on manually curated data. |
2405.19325 | Large language models (LLMs) often hallucinate and lack the ability to provide attribution for their generations. Semi-parametric LMs, such asNN-LM, approach these limitations by refining the output of an LM for a given prompt using its nearest neighbor matches in a non-parametric data store. However, these models often exhibit slow inference speeds and produce non-fluent texts. In this paper, we introduce Nearest Neighbor Speculative Decoding (Nest) , a novel semi-parametric language modeling approach that is capable of incorporating real-world text spans of arbitrary length into the LM generations and providing attribution to their sources.Nestperforms token-level retrieval at each inference step to compute a semi-parametric mixture distribution and identify promising span continuations in a corpus. It then uses an approximate speculative decoding procedure that accepts a prefix of the retrieved span or generates a new token.Nestsignificantly enhances the generation quality and attribution rate of the base LM across a variety of knowledge-intensive tasks, surpassing the conventionalNN-LM method and performing competitively with in-context retrieval augmentation. In addition,Nestsubstantially improves the generation speed, achieving a 1.8speedup in inference time when applied to Llama-2-Chat 70B. |
2405.13326 | Finetuning large language models with a variety of instruction-response pairs has enhanced their capability to understand and follow instructions. Current instruction tuning primarily relies on teacher models or human intervention to generate and refine the instructions and responses, which are costly, non-sustainable, and may lack diversity. In this paper, we introduce Mosaic Instruction Tuning (Mosaic-IT) , a human/model-free method that can efficiently create rich and diverse augmentations from existing instruction tuning data to enhance the finetuned LLM.Mosaic-IT randomly concatenates multiple instruction data into one and trains the model to produce the corresponding responses with predefined higher-level meta-instructions to strengthen its multi-step instruction-following and format-following skills. Our extensive evaluations demonstrate a superior performance and training efficiency of Mosaic-IT, which achieves consistent performance improvements over various benchmarks and areduction in training costs compared with original instruction tuning. Our codes and data are available athttps://github.com/tianyi-lab/Mosaic-IT. |
2310.15421 | Theory of mind(ToM) evaluations currently focus on testing models using passive narratives that inherently lack interactivity.
We introduceFANToM, a new benchmark designed to stress-test ToM within information-asymmetric conversational contexts via question answering.
Our benchmark draws upon important theoretical requisites from psychology and necessary empirical considerations when evaluating large language models (LLMs) .
In particular, we formulate multiple types of questions that demand the same underlying reasoning to identifyillusoryor false sense of ToM capabilities in LLMs.
We show thatFANToMis challenging for state-of-the-art LLMs, which perform significantly worse than humans even with chain-of-thought reasoning or fine-tuning.111https://hyunw.kim/fantom |
2205.06175 | Inspired by progress in large-scale language modeling, we apply a similar approach towards building a single generalist agent beyond the realm of text outputs.
The agent, which we refer to as Gato, works as a multi-modal, multi-task, multi-embodiment generalist policy.
The same network with the same weights can play Atari, caption images, chat, stack blocks with a real robot arm and much more, deciding based on its context whether to output text, joint torques, button presses, or other tokens.
In this report we describe the model and the data, and document the current capabilities of Gato. |
2405.20204 | Contrastive Language-Image Pretraining (CLIP) is widely used to train models to align images and texts in a common embedding space by mapping them to fixed-sized vectors. These models are key to multimodal information retrieval and related tasks. However, CLIP models generally underperform in text-only tasks compared to specialized text models. This creates inefficiencies for information retrieval systems that keep separate embeddings and models for text-only and multimodal tasks. We propose a novel, multi-task contrastive training method to address this issue, which we use to train thejina-clip-v1model to achieve the state-of-the-art performance on both text-image and text-text retrieval tasks. |
2405.19316 | Language model (LM) post-training (or alignment) involves maximizing a reward function that is derived from preference annotations. Direct Preference Optimization (DPO) is a popular offline alignment method that trains a policy directly on preference data without the need to train a reward model or apply reinforcement learning. However, typical preference datasets have only a single, or at most a few, annotation per preference pair, which causes DPO to overconfidently assign rewards that trend towards infinite magnitude. This frequently leads to degenerate policies, sometimes causing even the probabilities of thepreferredgenerations to go to zero. In this work, we analyze this phenomenon and proposedistillationto get a better proxy for the true preference distribution over generation pairs: we train the LM to produce probabilities that match the distribution induced by a reward model trained on the preference data. Moreover, to account for uncertainty in the reward model we are distilling from, we optimize against afamily of reward modelsthat, as a whole, is likely to include at least one reasonable proxy for the preference distribution. Our results show that distilling from such a family of reward models leads to improved robustness to distribution shift in preference annotations, while preserving the simple supervised nature of DPO. |
2405.14862 | We introduceBitune, a method that improves instruction-tuning of pretrained decoder-only large language models, leading to consistent gains on downstream tasks.Bituneapplies both causalandbidirectional attention to the prompt, to obtain a better representation of the query or instruction. We realize this by introducing two sets of parameters, for which we apply parameter-efficient finetuning techniques.
These causal and bidirectional features are then combined into a weighted average with trainable coefficients, which is subsequently used to generate new tokens. We demonstrate significant improvements in zero-shot performance on commonsense reasoning, arithmetic, and language understanding tasks, while extensive ablation studies validate the role of each component and demonstrate the method’s agnosticism to different PEFT techniques. |
2306.06101 | We consider the problem of estimating the learning rate in adaptive methods, such as Adagrad and Adam. We describe two techniques, Prodigy and Resetting, to provably estimate the distance to the solution, which is needed to set the learning rate optimally. Our techniques are modifications of the D-Adaptation method for learning-rate-free learning. Our methods improve upon the convergence rate of D-Adaptation by a factor of, whereis the initial estimate of. We test our methods on 12 common logistic-regression benchmark datasets, VGG11 and ResNet-50 training on CIFAR10, ViT training on Imagenet, LSTM training on IWSLT14, DLRM training on Criteo dataset, VarNet on Knee MRI dataset, as well as RoBERTa and GPT transformer training on BookWiki. Our experimental results show that our approaches consistently outperform D-Adaptation and reach test accuracy values close to that of hand-tuned Adam. |
2301.07733 | D-Adaptation is an approach to automatically setting the learning rate which asymptotically achieves the optimal rate of convergence for minimizing convex Lipschitz functions, with no back-tracking or line searches, and no additional function value or gradient evaluations per step. Our approach is the first hyper-parameter free method for this class without additional multiplicative log factors in the convergence rate. We present extensive experiments for SGD and Adam variants of our method, where the method automatically matches hand-tuned learning rates across more than a dozen diverse machine learning problems, including large-scale vision and language problems. An open-source implementation is available1. |
2405.18719v1 | The attention mechanism is a critical component of Large Language Models (LLMs) that allows tokens in a sequence to interact with each other, but isorder-invariant.
Incorporating position encoding (PE) makes it possible to address by position,
such as attending to the-th token.
However, current PE methods use token counts to derive position, and thus cannot generalize to higher levels of abstraction, such as attending to the-th sentence.
In this paper, we propose a new position encoding method,Contextual Position Encoding(CoPE) , that allows positions to be conditioned on context by incrementing position only on certain tokens determined by the model.
This allows more general position addressing such as attending to the-th particular word, noun, or sentence.
We show that CoPE can solve the selective copy, counting and Flip-Flop tasks where popular position embeddings fail, and improves perplexity on language modeling and coding tasks. |
2405.18400 | Many applications today provide users with multiple auto-complete drafts as they type, including GitHub’s code completion, Gmail’s smart compose, and Apple’s messaging auto-suggestions.
Under the hood, language models support this by running an autoregressive inference pass to provide a draft. Consequently, providingdrafts to the user requires running an expensive language modeltimes. To alleviate the computation cost of runninginference passes, we propose Superposed Decoding, a new decoding algorithm that generatesdrafts at the computation cost of one autoregressive inference pass. We achieve this by feeding a superposition of the most recent token embeddings from thedrafts as input to the next decoding step of the language model. At every inference step we combine thedrafts with the top-tokens to getnew drafts and cache themost likely options, using an n-gram interpolation with minimal compute overhead to filter out incoherent generations. Our experiments show thatdrafts from Superposed Decoding are at least as coherent and factual as Nucleus Sampling and Greedy Decoding respectively, while being at leastfaster for. In a compute-normalized setting, user evaluations demonstrably favor text generated by Superposed Decoding over Nucleus Sampling. Code and more examples open-sourced athttps://github.com/RAIVNLab/SuperposedDecoding. |
2405.05938 | Experts in various fields routinely perform methodical writing tasks
to plan, organize, and report their work. From a clinician writing a
differential diagnosis for a patient, to a teacher writing a lesson
plan for students, these tasks are pervasive, requiring tomethodicallygenerate structured long-form output for a given
input. We develop a typology of methodical tasks structured in the
form of a task objective, procedure, input, and output, and
introduce DoLoMiTes, a novel benchmark with specifications for
519 such tasks elicited from hundreds of experts from across
25 fields. Our benchmark further contains specific instantiations of
methodical tasks with concrete input and output examples (1,857 in
total) which we obtain by collecting expert revisions of up to 10
model-generated examples of each task. We use these examples to
evaluate contemporary language models highlighting that automating
methodical tasks is a challenging long-form generation problem, as
it requires performing complex inferences, while drawing upon the
given context as well as domain knowledge.
Our dataset is available athttps://dolomites-benchmark.github.io/. |
2405.15682 | Existing learning rate schedules that do not require specification of the optimization stopping stepare greatly out-performed by learning rate schedules that depend on. We propose an approach that avoids the need for this stopping time by eschewing the use of schedules entirely, while exhibiting state-of-the-art performance compared to schedules across a wide family of problems ranging from convex problems to large-scale deep learning problems. OurSchedule-Freeapproach introduces no additional hyper-parameters over standard optimizers with momentum. Our method is a direct consequence of a new theory we develop that unifies scheduling and iterate averaging. An open source implementation of our method is available111https://github.com/facebookresearch/schedule_free. |
2405.15319 | LLMs are computationally expensive to pre-train due to their large scale.
Model growth emerges as a promising approach by leveraging smaller models to accelerate the training of larger ones.
However, the viability of these model growth methods in efficient LLM pre-training remains underexplored.
This work identifies three criticalObstacles: (O1) lack of comprehensive evaluation, (O2) untested viability for scaling, and (O3) lack of empirical guidelines.
To tackleO1, we summarize existing approaches into four atomic growth operators and systematically evaluate them in a standardized LLM pre-training setting.
Our findings reveal that a depthwise stacking operator, called, exhibits remarkable acceleration in training, leading to decreased loss and improved overall performance on eight standard NLP benchmarks compared to strong baselines.
Motivated by these promising results, we conduct extensive experiments to delve deeper intoto addressO2 andO3.
ForO2 (untested scalability) , our study shows thatis scalable and consistently performs well, with experiments up to 7B LLMs after growth and pre-training LLMs with 750B tokens.
For example, compared to a conventionally trained 7B model using 300B tokens, ourmodel converges to the same loss with 194B tokens, resulting in a 54.6% speedup.
We further addressO3 (lack of empirical guidelines) by formalizing guidelines to determine growth timing and growth factor for, making it practical in general LLM pre-training.
We also provide in-depth discussions and comprehensive ablation studies of.
Our code and pre-trained model are available athttps://llm-stacking.github.io/. |
2405.16039 | Previous work on Universal Transformers (UTs) has demonstrated the importance of parameter sharing across layers. By allowing recurrence in depth, UTs have advantages over standard Transformers in learning compositional generalizations, but layer-sharing comes with a practical limitation of parameter-compute ratio:
it drastically reduces the parameter count compared to the non-shared model with the same dimensionality. Naively scaling up the layer size to compensate for the loss of parameters makes its computational resource requirements prohibitive. In practice, no previous work has succeeded in proposing a shared-layer Transformer design that is competitive in parameter count-dominated tasks such as language modeling. Here we propose MoEUT (pronounced “moot”) , an effective mixture-of-experts (MoE) -based shared-layer Transformer architecture, which combines several recent advances in MoEs for both feedforward and attention layers of standard Transformers together with novel layer-normalization and grouping schemes that are specific and crucial to UTs. The resulting UT model, for the first time, slightly outperforms standard Transformers on language modeling tasks such as BLiMP and PIQA, while using significantly less compute and memory.111Our code is public:https://github.com/robertcsordas/moeut |
2204.03458 | Generating temporally coherent high fidelity video is an important milestone in generative modeling research. We make progress towards this milestone by proposing a diffusion model for video generation that shows very promising initial results. Our model is a natural extension of the standard image diffusion architecture, and it enables jointly training from image and video data, which we find to reduce the variance of minibatch gradients and speed up optimization.
To generate long and higher resolution videos we introduce a new
conditional sampling technique for spatial and temporal video extension that performs better than previously proposed methods.
We present the first results on a large text-conditioned video generation task, as well as state-of-the-art results on established benchmarks for video prediction and unconditional video generation. Supplementary material is available athttps://video-diffusion.github.io/. |
2403.04690 | Neighborhood attentionreduces the cost of self attention by restricting each token’s attention span
to its nearest neighbors.
This restriction, parameterized by a window size and dilation factor, draws a spectrum of possible attention patterns between
linear projection and self attention.
Neighborhood attention, and more generally sliding window attention patterns, have long been bounded by infrastructure,
particularly in higher-rank spaces (2-D and 3-D) ,
calling for the development of custom kernels, which have been limited in either functionality, or performance, if not both.
In this work, we aim to massively improve upon existing infrastructure by providing two new methods for implementing
neighborhood attention.
We first show that neighborhood attention can be represented as a batched GEMM problem, similar to standard attention,
and implement it for 1-D and 2-D neighborhood attention.
These kernels on average provide 895% and 272% improvement in full precision latency compared to existing naive CUDA kernels
for 1-D and 2-D neighborhood attention respectively.
We find certain inherent inefficiencies in all unfused implementations of neighborhood attention that bound their performance
and hinder their lower-precision scalability.
Motivated by the progress made into fused dot-product attention kernels, we developed fused neighborhood attention;
an adaptation of fused dot-product attention kernels that allow fine-grained control over attention across different
spatial axes.
Known for reducing the quadratic time complexity of self attention to a linear complexity, neighborhood attention can now enjoy
a reduced and constant memory footprint, and record-breaking half precision latency.
We observe that our fused implementation successfully circumvents some of the unavoidable inefficiencies in unfused
implementations.
While our unfused GEMM-based kernels only improve half precision performance compared to naive kernels by an average of 496% and 113% in
1-D and 2-D problems respectively,
our fused kernels improve naive kernels by an average of 1607% and 581% in 1-D and 2-D problems respectively.
These translate into up to 97% improvement in inference speed in existing models based on neighborhood attention,
and additionally extend its applicability in vision (image and video) as well as other modalities.
We will open source our kernels atgithub.com/SHI-Labs/NATTEN. |
2311.01906 | A simple design recipe for deep Transformers is to compose identical building blocks. But standard transformer blocks are far from simple, interweaving attention and MLP sub-blocks with skip connections & normalisation layers in precise arrangements. This complexity leads to brittle architectures, where seemingly minor changes can significantly reduce training speed, or render models untrainable. In this work, we ask to what extent the standard transformer block can be simplified? Combining signal propagation theory and empirical observations, we motivate modifications that allow many block components to be removed with no loss of training speed, including skip connections, projection or value parameters, sequential sub-blocks and normalisation layers. In experiments on both autoregressive decoder-only and BERT encoder-only models, our
simplified transformers emulate the per-update training speed and performance of standard transformers, while enjoying 15% faster training throughput, and using 15% fewer parameters. |
2405.10637 | Huge memory consumption has been a major bottleneck for deploying high-throughput large language models in real-world applications. In addition to the large number of parameters, the key-value (KV) cache for the attention mechanism in the transformer architecture consumes a significant amount of memory, especially when the number of layers is large for deep language models. In this paper, we propose a novel method that only computes and caches the KVs of a small number of layers, thus significantly saving memory consumption and improving inference throughput. Our experiments on large language models show that our method achieves up to 26higher throughput than standard transformers and competitive performance in language modeling and downstream tasks. In addition, our method is orthogonal to existing transformer memory-saving techniques, so it is straightforward to integrate them with our model, achieving further improvement in inference efficiency.
Our code is available athttps://github.com/whyNLP/LCKV. |
2405.12532 | Large Language Models (LLMs) have shown remarkable comprehension abilities but face challenges in GPU memory usage during inference, hindering their scalability for real-time applications like chatbots. To accelerate inference, we store computed keys and values (KV cache) in the GPU memory. Existing methods study the KV cache compression to reduce memory by pruning the pre-computed KV cache. However, they neglect the inter-layer dependency between layers and huge memory consumption in pre-computation. To explore these deficiencies, we find that the number of crucial keys and values that influence future generations decreases layer by layer and we can extract them by the consistency in attention weights. Based on the findings, we propose PyramidInfer, a method that compresses the KV cache by layer-wise retaining crucial context. PyramidInfer saves significant memory by computing fewer keys and values without sacrificing performance. Experimental results show PyramidInfer improves 2.2x throughput compared to Accelerate with over 54% GPU memory reduction in KV cache. Our code is available inhttps://github.com/mutonix/pyramidinfer. |
2405.00732 | Low Rank Adaptation (LoRA) has emerged as one of the most widely adopted methods for Parameter Efficient Fine-Tuning (PEFT) of Large Language Models (LLMs) . LoRA reduces the number of trainable parameters and memory usage while achieving comparable performance to full fine-tuning. We aim to assess the viability of training and serving LLMs fine-tuned with LoRA in real-world applications. First, we measure the quality of LLMs fine-tuned with quantized low rank adapters across 10 base models and 31 tasks for a total of 310 models. We find that 4-bit LoRA fine-tuned models outperform base models by 34 points and GPT-4 by 10 points on average. Second, we investigate the most effective base models for fine-tuning and assess the correlative and predictive capacities of task complexity heuristics in forecasting the outcomes of fine-tuning. Finally, we evaluate the latency and concurrency capabilities of LoRAX, an open-source Multi-LoRA inference server that facilitates the deployment of multiple LoRA fine-tuned models on a single GPU using shared base model weights and dynamic adapter loading. LoRAX powersLoRA Land, a web application that hosts 25 LoRA fine-tuned Mistral-7B LLMs on a single NVIDIA AGPU withGB memory. LoRA Land highlights the quality and cost-effectiveness of employing multiple specialized LLMs over a single, general-purpose LLM. |
2405.08448 | Reinforcement learning from human feedback (RLHF) is the canonical framework for large language model alignment. However, rising popularity in offline alignment algorithms challenge the need for on-policy sampling in RLHF. Within the context of reward over-optimization, we start with an opening set of experiments that demonstrate the clear advantage of online methods over offline methods. This prompts us to investigate the causes to the performance discrepancy through a series of carefully designed experimental ablations. We show empirically that hypotheses such as offline data coverage and data quality by itself cannot
convincingly explain the performance difference. We also find that while offline algorithms train policy to become good at pairwise classification, it is worse at generations; in the meantime the policies trained by online algorithms are good at generations while worse at pairwise classification. This hints at a unique interplay between discriminative and generative capabilities, which is greatly impacted by the sampling process. Lastly, we observe that the performance discrepancy persists for both contrastive and non-contrastive loss functions, and appears not to be addressed by simply scaling up policy networks. Taken together, our study sheds light on the pivotal role of on-policy sampling in AI alignment, and hints at certain fundamental challenges of offline alignment algorithms. |
2405.09818 | We present Chameleon, a family of early-fusion token-based mixed-modal models capable of understanding and generating images and text in any arbitrary sequence. We outline a stable training approach from inception, an alignment recipe, and an architectural parameterization tailored for the early-fusion, token-based, mixed-modal setting. The models are evaluated on a comprehensive range of tasks, including visual question answering, image captioning, text generation, image generation, and long-form mixed modal generation. Chameleon demonstrates broad and general capabilities, including state-of-the-art performance in image captioning tasks, outperforms Llama-2 in text-only tasks while being competitive with models such as Mixtral 8x7B and Gemini-Pro, and performs non-trivial image generation, all in a single model. It also matches or exceeds the performance of much larger models, including Gemini Pro and GPT-4V, according to human judgments on a new long-form mixed-modal generation evaluation, where either the prompt or outputs contain mixed sequences of both images and text. Chameleon marks a significant step forward in a unified modeling of full multimodal documents. |
2308.03188 | Large language models (LLMs) have demonstrated remarkable performance across a wide array of NLP tasks. However, their efficacy is undermined by undesired and inconsistent behaviors, including hallucination, unfaithful reasoning, and toxic content. A promising approach to rectify these flaws isself-correction,
where the LLM itself is prompted or guided to fix problems in its own output.
Techniques leveragingautomated feedback—either produced by the LLM itself or some external system—are of particular interest as they are a promising way to make LLM-based solutions more practical and deployable with minimal human feedback.
This paper presents a comprehensive review of this emerging class of techniques.
We analyze and taxonomize a wide array of recent work utilizing these strategies, including training-time, generation-time, and post-hoc correction. We also summarize the major applications of this strategy and conclude by discussing future directions and challenges. |
2211.00053 | Sequence generation applications require satisfying semantic constraints, such as ensuring that programs are correct,
using certain keywords,
or avoiding undesirable content.
Language models, whether fine-tuned or prompted with few-shot demonstrations, frequently violate these constraints, and lack a mechanism to iteratively revise their outputs.
Moreover, some powerful language models are of extreme scale or inaccessible, making it inefficient, if not infeasible, to update their parameters for task-specific adaptation.
We presentself-correction,
an approach that decouples an imperfect base generator (an off-the-shelf language model or supervised sequence-to-sequence model) from a separate corrector that learns to iteratively correct imperfect generations.
To train the corrector, we propose an online training procedure that can use either scalar or natural language feedback on intermediate imperfect generations.
We show thatself-correctionimproves upon the base generator in three diverse generation tasks– mathematical program synthesis, lexically-constrained generation, and toxicity control– even
when the corrector is much smaller than the base generator. |
2008.08927 | We describe a novel approach for generating music using a self-correcting, non-chronological, autoregressive model.
We represent music as a sequence of edit events, each of which denotes either the addition or removal of a note—even a note previously generated by the model.
During inference, we generate one edit event at a time using direct ancestral sampling.
Our approach allows the model to fix previous mistakes such as incorrectly sampled notes and prevent accumulation of errors which autoregressive models are prone to have.
Another benefit is a finer, note-by-note control during human and AI collaborative composition.
We show through quantitative metrics and human survey evaluation that our approach generates better results than orderless NADE and Gibbs sampling approaches. |
1812.06855 | During the development of AlphaGo, its many hyper-parameters were tuned with Bayesian optimization multiple times. This automatic tuning process resulted in substantial improvements in playing strength. For example,
prior to the match with Lee Sedol, we tuned the latest AlphaGo agent and this improved its
win-rate from 50% to 66.5% in self-play games. This tuned version was deployed in the final match. Of course, since we tuned AlphaGo many times during its development cycle, the compounded contribution was even higher than this percentage.
It is our hope that this brief case study will be of interest to Go fans, and also provide Bayesian optimization practitioners with some insights and inspiration. |
2405.09673 | Low-Rank Adaptation (LoRA) is a widely-used parameter-efficient finetuning method for large language models. LoRA saves memory by training only low rank perturbations to selected weight matrices. In this work, we compare the performance of LoRA and full finetuning on two target domains, programming and mathematics. We consider both the instruction finetuning (100K prompt-response pairs) and continued pretraining (10B unstructured tokens) data regimes. Our results show that, in most settings, LoRA substantially underperforms full finetuning.
Nevertheless, LoRA exhibits a desirable form of regularization: it better maintains the base model’s performance on tasks outside the target domain.
We show that LoRA provides stronger regularization compared to common techniques such as weight decay and dropout; it also helps maintain more diverse generations. We show that full finetuning learns perturbations with a rank that is 10-100X greater than typical LoRA configurations, possibly explaining some of the reported gaps. We conclude by proposing best practices for finetuning with LoRA. |
1503.04069v2 | Several variants of the Long Short-Term Memory (LSTM) architecture for recurrent neural networks have been proposed since its inception in 1995.
In recent years, these networks have become the state-of-the-art models for a variety of machine learning problems.
This has led to a renewed interest in understanding the role and utility of various computational components of typical LSTM variants.
In this paper, we present the first large-scale analysis of eight LSTM variants on three representative tasks: speech recognition, handwriting recognition, and polyphonic music modeling.
The hyperparameters of all LSTM variants for each task were optimized separately using random search, and their importance was assessed using the powerful fANOVA framework.
In total, we summarize the results of 5400 experimental runs (years of CPU time) , which makes our study the largest of its kind on LSTM networks.
Our results show that none of the variants can improve upon the standard LSTM architecture significantly, and demonstrate the forget gate and the output activation function to be its most critical components.
We further observe that the studied hyperparameters are virtually independent and derive guidelines for their efficient adjustment. |
1805.02867v2 | The Softmax function is ubiquitous in machine learning, multiple previous works suggested faster alternatives for it. In this paper we propose a way to compute classical Softmax with fewer memory accesses and hypothesize that this reduction in memory accesses should improve Softmax performance on actual hardware. The benchmarks confirm this hypothesis: Softmax accelerates by up tox and Softmax+TopK combined and fused by up tox. |
2405.08944 | Transformer-based long context generative models power emerging AI applications like
hour-long video understanding and project-level coding agent.
Deploying long context transformers (e.g., 100K to 10M tokens) is prohibitively expensive compared to short context (e.g., 4K tokens) model variants.
Reducing the cost of long-context transformers is becoming a pressing research and engineering challenge starting from the year of 2024.
This work describes a concurrent programming framework for quantitatively analyzing the efficiency challenges in serving multiple long-context requests under limited size of GPU high-bandwidth memory (HBM) regime.
We give a detailed analysis of how all additional computational costs, compared to 4K context, trace back toone single source: the large size of the KV cache.
We use a 34B GPT-3.5 level model of 50K context on A100 NVLink as a running example, and describe how its large KV cache
causes four types of deployment challenges:
(1) prefilling long inputs takes much longer compute time and GPU memory than short inputs;
(2) after prefilling, the large KV cache residing on the GPU HBM substantially restricts the number of concurrent users being served;
(3) during decoding, repeatedly reading the KV cache from HBM to SM largely increases latency;
(4) when KV cache memory overflows, swapping it from HBM to DDR causes significant context switching latency.
We use this framework to analyze existing works
and identify possibilities of combining them to build end-to-end systems.
Overall, this work offers a foundational framework for analyzing long context transformer deployment and identifies directions towards reducing the inference cost of 1M context to be as cheap as 4K. |
2404.06654 | The needle-in-a-haystack (NIAH) test, which examines the ability to retrieve a piece of information (the “needle”) from long distractor texts (the “haystack”) , has been widely adopted to evaluate long-context language models (LMs) . However, this simple retrieval-based test is indicative of only a superficial form of long-context understanding. To provide a more comprehensive evaluation of long-context LMs, we create a new synthetic benchmarkRulerwith flexible configurations for customized sequence length and task complexity.Rulerexpands upon the vanilla NIAH test to encompass variations with diverse types and quantities of needles. Moreover,Rulerintroduces new task categoriesmulti-hop tracingandaggregationto test behaviors beyond searching from context. We evaluate ten long-context LMs with 13 representative tasks inRuler. Despite achieving nearly perfect accuracy in the vanilla NIAH test, all models exhibit large performance drops as the context length increases. While these models all claim context sizes of 32K tokens or greater, only four models (GPT-4, Command-R, Yi-34B, and Mixtral) can maintain satisfactory performance at the length of 32K. Our analysis of Yi-34B, which supports context length of 200K, reveals large room for improvement as we increase input length and task complexity. We open sourceRulerto spur comprehensive evaluation of long-context LMs.††* Authors contributed equally. |
2309.17133 | Knowledge-based Visual Question Answering (KB-VQA) requires VQA systems to utilize knowledge from external knowledge bases to answer visually-grounded questions. Retrieval-Augmented Visual Question Answering (RA-VQA) , a strong framework to tackle KB-VQA, first retrieves related documents with Dense Passage Retrieval (DPR) and then uses them to answer questions. This paper proposes Fine-grained Late-interaction Multi-modal Retrieval (FLMR) which significantly improves knowledge retrieval in RA-VQA. FLMR addresses two major limitations in RA-VQA’s retriever: (1) the image representations obtained via image-to-text transforms can be incomplete and inaccurate and (2) relevance scores between queries and documents are computed with one-dimensional embeddings, which can be insensitive to finer-grained relevance.
FLMR overcomes these limitations by obtaining image representations that complement those from the image-to-text transforms using a vision model aligned with an existing text-based retriever through a simple alignment network. FLMR also encodes images and questions using multi-dimensional embeddings to capture finer-grained relevance between queries and documents.
FLMR significantly improves the original RA-VQA retriever’s PRRecall@5 by approximately 8%. Finally, we equipped RA-VQA with two state-of-the-art large multi-modal/language models to achieveVQA score in the OK-VQA dataset. |
2303.09859 | While modern masked language models (LMs) are trained on ever larger corpora, we here explore the effects of down-scaling training to a modestly-sized but representative, well-balanced, and publicly available English text source – the British National Corpus. We show that pre-training on this carefully curated corpus can reach better performance than the original BERT model. We argue that this type of corpora has great potential as a language modeling benchmark. To showcase this potential, we present fair, reproducible and data-efficient comparative studies of LMs, in which we evaluate several training objectives and model architectures and replicate previous empirical results in a systematic way. We propose an optimized LM architecture called LTG-BERT. |
2405.05254 | We introduce a decoder-decoder architecture, YOCO, for large language models, which only caches key-value pairs once. It consists of two components, i.e., across-decoderstacked upon aself-decoder. The self-decoder efficiently encodes global key-value (KV) caches that are reused by the cross-decoder via cross-attention. The overall model behaves like a decoder-only Transformer, although YOCO only caches once. The design substantially reduces GPU memory demands, yet retains global attention capability. Additionally, the computation flow enables prefilling to early exit without changing the final output, thereby significantly speeding up the prefill stage. Experimental results demonstrate that YOCO achieves favorable performance compared to Transformer in various settings of scaling up model size and number of training tokens. We also extend YOCO to 1M context length with near-perfect needle retrieval accuracy. The profiling results show that YOCO improves inference memory, prefill latency, and throughput by orders of magnitude across context lengths and model sizes. Code is available athttps://aka.ms/YOCO. |
2405.07863 | We present the workflow of Online Iterative Reinforcement Learning from Human Feedback (RLHF) in this technical report, which is widely reported to outperform its offline counterpart by a large margin in the recent large language model (LLM) literature. However, existing open-source RLHF projects are still largely confined to the offline learning setting. In this technical report, we aim to fill in this gap and provide a detailed recipe that is easy to reproduce for online iterative RLHF. In particular, since online human feedback is usually infeasible for open-source communities with limited resources, we start by constructing preference models using a diverse set of open-source datasets and use the constructed proxy preference model to approximate human feedback. Then, we discuss the theoretical insights and algorithmic principles behind online iterative RLHF, followed by a detailed practical implementation.
Our trained LLM,SFR-Iterative-DPO-LLaMA-3-8B-R, achieves impressive performance on LLM chatbot benchmarks, including AlpacaEval-2, Arena-Hard, and MT-Bench, as well as other academic benchmarks such as HumanEval and TruthfulQA.
We have shown that supervised fine-tuning (SFT) and iterative RLHF can obtain state-of-the-art performance with fully open-source datasets.
Further, we have made our models, curated datasets, and comprehensive step-by-step code guidebooks publicly available. Please refer tohttps://github.com/RLHFlow/RLHF-Reward-Modelingandhttps://github.com/RLHFlow/Online-RLHFfor more detailed information. |
2402.18668 | Recent work has shown that attention-based language models excel atrecall, the ability to ground generations in tokens previously seen in context. However, the efficiency of attention-based models is bottle-necked during inference by the KV-cache’s aggressive memory consumption.
In this work, we explore whether we can improve language model efficiency (e.g.by reducing memory consumption) without compromising on recall.
By applying experiments and theory to a broad set of architectures, we identify a key tradeoff between a model’sstate sizeand recall ability.
We show that efficient alternatives to attention (e.g.H3, Mamba, RWKV) maintain a fixed-size recurrent state, but struggle at recall.
We proposeBaseda simple architecture combining linear and sliding window attention.
By varyingBasedwindow size and linear attention feature dimension, we can dial the state size and traverse the pareto frontier of the recall-memory tradeoff curve, recovering the full quality of attention on one end and the small state size of attention-alternatives on the other.
We train language models up tob parameters and show thatBasedmatches the strongest sub-quadratic models (e.g.Mamba) in perplexity and outperforms them on real-world recall-intensive tasks by 6.22 accuracy points.
Implementations of linear attention are often less efficient than optimized standard attention implementations. To makeBasedcompetitive, we develop IO-aware algorithms that enablehigher throughput on language generation than FlashAttention-2, when generating 1024 tokens usingb parameter models. |
2402.04347 | Linear attentions have shown potential for improving Transformer efficiency, reducing attention’s quadratic complexity to linear in sequence length. This holds exciting promise for (1) training linear Transformers from scratch, (2) “finetuned-conversion” of task-specific Transformers into linear versions that recover task performance, and (3) “pretrained-conversion” of Transformers such as large language models into linear versions finetunable on downstream tasks. However, linear attentions often underperform standard softmax attention in quality. To close this performance gap, we find prior linear attentions lack key properties of softmax attention tied to good performance: low-entropy (or “spiky”) weights and dot-product monotonicity. We further observe surprisingly simple feature maps that retain these properties and match softmax performance, but are inefficient to compute in linear attention. We thus propose Hedgehog, a learnable linear attention that retains the spiky and monotonic properties of softmax attention while maintaining linear complexity. Hedgehog uses simple trainable MLPs to produce attention weights mimicking softmax attention. Experiments show Hedgehog recovers over 99% of standard Transformer quality in train-from-scratch and finetuned-conversion settings, outperforming prior linear attentions up to 6 perplexity points on WikiText-103 with causal GPTs, and up to 8.7 GLUE score points on finetuned bidirectional BERTs. Hedgehog also enables pretrained-conversion. Converting a pretrained GPT-2 into a linear attention variant achieves state-of-the-art 16.7 perplexity on WikiText-103 for 125M subquadratic decoder models. We finally turn a pretrained Llama-2 7B into a viable linear attention Llama. With low-rank adaptation, Hedgehog-Llama2 7B achieves 28.1 higher ROUGE-1 points over the base standard attention model, where prior linear attentions lead to 16.5 point drops. |
2312.04927 | Attention-free language models that combinegatingandconvolutionsare growing in popularity due to their efficiency and increasingly competitive performance.
To better understand these architectures, we pretrain a suite of 17 attention andgated-convolutionlanguage models, finding that SoTA gated-convolution architectures still underperform attention by up to 2.1 perplexity points on the Pile.
In fine-grained analysis, we find 82% of the gap is explained by each model’s ability to recall information that is previously mentioned in-context,e.g.Hakuna Matata means no worries Hakuna Matata it means no??.
On this task, termedassociative recall, we find that attention outperforms gated-convolutions by a large margin: a 70M parameter attention model outperforms a 1.4 billion parameter gated-convolution model on associative recall.
This is surprising because prior work shows gated convolutions can perfectly solve synthetic tests for AR capability.
To close the gap between synthetics and real language, we develop a new formalization of the task called multi-query associative recall (Mqar) that better reflects actual language.
We perform an empirical and theoretical study ofMqarthat elucidates differences in the parameter-efficiency of attention and gated-convolution recall.
Informed by our analysis, we evaluate simple convolution-attention hybrids and show that hybrids with input-dependent sparse attention patterns can close 97.4% of the gap to attention, while maintaining sub-quadratic scaling. Our code is accessible at:https://github.com/HazyResearch/zoology. |
2302.09178 | Recommender systems play an important role in many content platforms. While most recommendation research is dedicated to designing better models to improve user experience, we found that research on stabilizing the training for such models is severely under-explored. As recommendation models become larger and more sophisticated, they are more susceptible to training instability issues,i.e., loss divergence, which can make the model unusable, waste significant resources and block model developments. In this paper, we share our findings and best practices we learned for improving the training stability of a real-world multitask ranking model for YouTube recommendations. We show some properties of the model that lead to unstable training and conjecture on the causes. Furthermore, based on our observations of training dynamics near the point of training instability, we hypothesize why existing solutions would fail, and propose a new algorithm to mitigate the limitations of existing solutions. Our experiments on YouTube production dataset show the proposed algorithm can significantly improve training stability while not compromising convergence, comparing with several commonly used baseline methods. We open source our implementation athttps://github.com/tensorflow/recommenders/tree/main/tensorflow_recommenders/experimental/optimizers/clippy_adagrad.py. |
2405.04434 | We present DeepSeek-V2, a strong Mixture-of-Experts (MoE) language model characterized by economical training and efficient inference.
It comprises 236B total parameters, of which 21B are activated for each token, and supports a context length of 128K tokens.
DeepSeek-V2 adopts innovative architectures including Multi-head Latent Attention (MLA) and DeepSeekMoE.
MLA guarantees efficient inference through significantly compressing the Key-Value (KV) cache into a latent vector, while DeepSeekMoE enables training strong models at an economical cost through sparse computation.
Compared with DeepSeek 67B, DeepSeek-V2 achieves significantly stronger performance, and meanwhile saves 42.5% of training costs, reduces the KV cache by 93.3%, and boosts the maximum generation throughput to 5.76 times.
We pretrain DeepSeek-V2 on a high-quality and multi-source corpus consisting of 8.1T tokens, and further perform Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) to fully unlock its potential.
Evaluation results show that, even with only 21B activated parameters, DeepSeek-V2 and its chat versions still achieve top-tier performance among open-source models.
The model checkpoints are available athttps://github.com/deepseek-ai/DeepSeek-V2. |
2405.05904 | When large language models are aligned via supervised fine-tuning, they may encounter new factual information that was not acquired through pre-training.
It is often conjectured that this can teach the model the behavior ofhallucinatingfactually incorrect responses, as the model is trained to generate facts that are not grounded in its pre-existing knowledge.In this work,we study the impact of such exposure to new knowledge on the capability of the fine-tuned model to utilize its pre-existing knowledge.
To this end, we design a controlled setup, focused on closed-book QA, where we vary the proportion of the fine-tuning examples that introduce new knowledge.
We demonstrate that large language models struggle to acquire new factual knowledge through fine-tuning, as fine-tuning examples that introduce new knowledge are learned significantly slower than those consistent with the model’s knowledge.
However, we also find that as the examples with new knowledge are eventually learned, they linearly increase the model’s tendency to hallucinate.
Taken together, our results highlight the risk in introducing new factual knowledge through fine-tuning, and support the view that
large language models mostly acquire factual knowledge through pre-training, whereas fine-tuning teaches them to use it more efficiently. |
2404.08763 | Large Language Models (LLMs) have dramatically advanced AI applications, yet their deployment remains challenging due to their immense inference costs.
Recent studies ameliorate the computational costs of LLMs by increasing their activation sparsity but suffer from significant performance degradation on downstream tasks.
In this work, we introduce a new framework for sparsifying the activations of base LLMs and reducing inference costs, dubbedContextuallyAwareThresholding forSparsity (CATS) .
CATS is relatively simple, easy to implement, and highly effective.
At the heart of our framework is a new non-linear activation function.
We demonstrate that CATS can be applied to various base models, including Mistral-7B and Llama2-7B, and outperforms existing sparsification techniques in downstream task performance.
More precisely, CATS-based models often achieve downstream task performance within 1-2% of their base models without any fine-tuning and even at activation sparsity levels of 50%.
Furthermore, CATS-based models converge faster and display better task performance than competing techniques when fine-tuning is applied.
Finally, we develop a custom GPU kernel for efficient implementation of CATS that translates the activation of sparsity of CATS to real wall-clock time speedups.
Our custom kernel implementation of CATS results in a15% improvement in wall-clock inference latency of token
generation on both Llama-7B and Mistral-7B. |
2405.04065 | Retrieval-Augmented Language Modeling (RALM) by integrating large language models (LLM) with relevant documents from an external corpus is a proven method for enabling the LLM to generate information beyond the scope of its pre-training corpus. Previous work utilizing retrieved content by simply prepending it to the input poses a high runtime issue, which degrades the inference efficiency of the LLMs because they fail to use the Key-Value (KV) cache efficiently.
In this paper, we proposeFlashBack, a modular RALM designed to improve the inference efficiency of RALM with appending
context pattern while maintaining decent performance after fine-tuning by Low-Rank Adaption.FlashBackappends retrieved documents at the end of the context for efficiently utilizing the KV cache instead of prepending them. And we introduce Marking Token as two special prompt tokens for marking the boundary of the appending context during fine-tuning. Our experiments on testing generation quality show thatFlashBackcan remain decent generation quality in perplexity. And the inference speed ofFlashBackis up tofaster than the prepending counterpart on a 7B LLM (Llama 2) in the runtime test. Via bypassing unnecessary re-computation, it demonstrates an advancement by achieving significantly faster inference speed, and this heightened efficiency will substantially reduce inferential cost. |
2405.04517 | In the 1990s, the constant error carousel and gating were introduced as the central ideas of the Long Short-Term Memory (LSTM). Since then, LSTMs have stood the test of time and contributed to numerous deep learning success stories, in particular they constituted the first Large Language Models (LLMs). However, the advent of the Transformer technology with parallelizable self-attention at its core marked the dawn of a new era, outpacing LSTMs at scale. We now raise a simple question: How far do we get in language modeling when scaling LSTMs to billions of parameters, leveraging the latest techniques from modern LLMs, but mitigating known limitations of LSTMs? Firstly, we introduce exponential gating with appropriate normalization and stabilization techniques. Secondly, we modify the LSTM memory structure, obtaining: (i) sLSTM with a scalar memory, a scalar update, and new memory mixing, (ii) mLSTM that is fully parallelizable with a matrix memory and a covariance update rule. Integrating these LSTM extensions into residual block backbones yields xLSTM blocks that are then residually stacked into xLSTM architectures. Exponential gating and modified memory structures boost xLSTM capabilities to perform favorably when compared to state-of-the-art Transformers and State Space Models, both in performance and scaling.  Figure 1: The extended LSTM (xLSTM) family. From left to right: 1. The original LSTM memory cell with constant error carousel and gating. 2. New sLSTM and mLSTM memory cells that introduce exponential gating. sLSTM offers a new memory mixing technique. mLSTM is fully parallelizable with a novel matrix memory cell state and new covariance update rule. 3. mLSTM and sLSTM in residual blocks yield xLSTM blocks. 4. Stacked xLSTM blocks give an xLSTM architecture. arXiv:2405.04517v1 [cs.LG] 7 May 2024 |
2405.03548 | Instruction tuning improves the reasoning abilities of large language models (LLMs) , with data quality and scalability being the crucial factors. Most instruction tuning data come from human crowd-sourcing or GPT-4 distillation. We propose a paradigm to efficiently harvest 10 million
naturally existing instruction data from the pre-training web corpus to enhance LLM reasoning. Our approach involves (1) recalling relevant documents, (2) extracting instruction-response pairs, and (3) refining the extracted pairs using open-source LLMs.
Fine-tuning base LLMs on this dataset, we buildMAmmoTH2models, which significantly boost performance on reasoning benchmarks. Notably,MAmmoTH2-7B’s (Mistral) performance increases from 11% to 36.7% on MATH and from 36% to 68.4% on GSM8K without training on any in-domain data. Further trainingMAmmoTH2on public instruction tuning datasets yieldsMAmmoTH2-Plus, achieving state-of-the-art performance on several reasoning and chatbot benchmarks.
Our work demonstrates how to harvest large-scale, high-quality instruction data without costly human annotation or GPT-4 distillation, providing a new paradigm for building better instruction tuning data. |
2405.00200 | As model context lengths continue to increase, the number of demonstrations that can be provided in-context approaches the size of entire training datasets. We study the behavior of in-context learning (ICL) at this extreme scale on multiple datasets and models. We show that, for many datasets with large label spaces, performance continues to increase with hundreds or thousands of demonstrations.
We contrast this with example retrieval and finetuning: example retrieval shows excellent performance at low context lengths but has diminished gains with more demonstrations; finetuning is more data hungry than ICL but can sometimes exceed long-context ICL performance with additional data.
We use this ICL setting as a testbed to study several properties of both in-context learning and long-context models.
We show that long-context ICL is less sensitive to random input shuffling than short-context ICL, that grouping of same-label examples can negatively impact performance, and that the performance boosts we see do not arise from cumulative gain from encoding many examples together.
We conclude that although long-context ICL can be surprisingly effective, most of this gain comes from attending back to similar examples rather than task learning.111Data and code are available athttps://github.com/abertsch72/long-context-icl |
1312.5663 | Recently, it has been observed that when representations are learnt in a way that encourages sparsity, improved performance is obtained on classification tasks. These methods involve combinations of activation functions, sampling steps and different kinds of penalties. To investigate the effectiveness of sparsity by itself, we propose the “k-sparse autoencoder”, which is an autoencoder with linear activation function, where in hidden layers only thehighest activities are kept. When applied to the MNIST and NORB datasets, we find that this method achieves better classification results than denoising autoencoders, networks trained with dropout, and RBMs.k-sparse autoencoders are simple to train and the encoding stage is very fast, making them well-suited to large problem sizes, where conventional sparse coding algorithms cannot be applied. |
2312.05821 | This paper explores a new post-hoc training-free compression paradigm for compressing Large Language Models (LLMs) to facilitate their wider adoption in various computing environments. We delve into the challenges of LLM compression, notably their dependency on extensive training data and computational resources. We propose a training-free approach dubbed Activation-aware Singular Value Decomposition (ASVD) to address these limitations. ASVD effectively manages activation outliers by adjusting the weight matrix based on the activation distribution, improving decomposition accuracy and efficiency. Our method also addresses the varying sensitivity of different LLM layers to decomposition, with an iterative calibration process for optimal layer-specific decomposition. Experiments demonstrate that ASVD can compress network by 10%-20% without losing reasoning capacities. Additionally, it can be seamlessly integrated with other LLM compression paradigms, showcasing its flexible compatibility. Code and compressed models are available atASVD4LLM. |
2405.03553 | Recent advancements in large language models (LLMs) have substantially enhanced their mathematical reasoning abilities. However, these models still struggle with complex problems that require multiple reasoning steps, frequently leading to logical or numerical errors. While numerical mistakes can be largely addressed by integrating a code interpreter, identifying logical errors within intermediate steps is more challenging. Moreover, manually annotating these steps for training is not only expensive but also labor-intensive, requiring the expertise of professional annotators. In our study, we introduce an innovative approach that bypasses the need for process annotations (from human or GPTs) by utilizing the Monte Carlo Tree Search (MCTS) framework. This technique automatically generates both the process supervision and the step-level evaluation signals. Our method iteratively trains the policy and value models, leveraging the capabilities of a well-pretrained LLM to progressively enhance its mathematical reasoning skills. Furthermore, we propose an efficient inference strategy—step-level beam search, where the value model is crafted to assist the policy model (*i.e.*, LLM) in navigating more effective reasoning paths, rather than solely relying on prior probabilities. The experimental results on both in-domain and out-of-domain datasets demonstrate that even without GPT-4 or human-annotated process supervision, our AlphaMath framework achieves comparable or superior results to previous state-of-the-art methods. The code for our method will be made available at https://github.com/MARIO-Math-Reasoning/Super_MARIO. |
2403.09636 | Transformers have emerged as the backbone of large language models (LLMs) . However, generation remains inefficient due to the need to store in memory a cache of key–value representations for past tokens, whose size scales linearly with the input sequence length and batch size.
As a solution, we propose Dynamic Memory Compression (DMC) , a method for on-line key–value cache compression at inference time.
Most importantly, the model learns to apply different compression rates in different heads and layers.
We retrofit pre-trained LLMs such as Llama 2 (7B, 13B and 70B) into DMC Transformers,
achieving up to ~3.7throughput increase during auto-regressive inference on an NVIDIA H100 GPU.
DMC is applied via continued pre-training on a negligible percentage of the original data without adding any extra parameters.
We find that DMC preserves the original downstream performance with up to 4cache compression, outperforming up-trained grouped-query attention (GQA) . GQA and DMC can be even combined to obtain compounded gains. As a result DMC fits longer contexts and larger batches within any given memory budget. |
2405.00982 | Large Language Models (LLMs) have enabled new ways to satisfy information needs.
Although great strides have been made in applying them to settings like document ranking and short-form text generation,
they still struggle to compose complete, accurate, and verifiable long-form reports.
Reports with these qualities are necessary to satisfy the complex, nuanced, or multi-faceted information needs of users.
In this perspective paper, we draw together opinions from industry and academia,
and from a variety of related research areas, to present our vision for automatic report generation, and—critically—a flexible framework by which such reports can be evaluated.
In contrast with other summarization tasks,
automatic report generation starts with a detailed description of an information need,
stating the necessary background, requirements, and scope of the report.
Further, the generated reports should be complete, accurate, and verifiable.
These qualities, which are desirable—if not required—in many analytic report-writing settings,
require rethinking how to build and evaluate systems that exhibit these qualities.
To foster new efforts in building these systems,
we present an evaluation framework that draws on ideas
found in various evaluations. To test completeness and accuracy, the framework uses nuggets of information, expressed as questions and answers, that need to be part of any high-quality generated report.
Additionally, evaluation of citations that map claims made in the report to their source documents ensures verifiability. |
2306.17806 | Classifier-Free Guidance (CFG) has recently emerged in text-to-image generation as a lightweight technique to encourage prompt-adherence in generations. In this work, we demonstrate that CFG can be used broadly as an inference-time technique in pure language modeling. We show that CFG (1) improves the performance of Pythia, GPT-2 and LLaMA-family models across an array of tasks: Q&A, reasoning, code generation, and machine translation, achieving SOTA on LAMBADA with LLaMA-7B over PaLM-540B; (2) brings improvements equivalent to a model with twice the parameter-count; (3) can stack alongside other inference-time methods like Chain-of-Thought and Self-Consistency, yielding further improvements in difficult tasks; (4) can be used to increase the faithfulness and coherence of assistants in challenging form-driven and content-driven prompts: in a human evaluation we show a 75% preference for GPT4All using CFG over baseline. |