Ben Athiwaratkun

Ben Athiwaratkun, Sanjay Krishna Gouda, Zijian Wang et al. · amazon-science, ibm-research

We present new benchmarks on evaluation code generation models: MBXP and Multilingual HumanEval, and MathQA-X. These datasets cover over 10 programming languages and are generated using a scalable conversion framework that transpiles prompts and test cases from the original Python datasets into the corresponding data in the target language. Using these benchmarks, we are able to assess the performance of code generation models in a multi-lingual fashion, and discovered generalization ability of language models on out-of-domain languages, advantages of multi-lingual models over mono-lingual, the ability of few-shot prompting to teach the model new languages, and zero-shot translation abilities even on mono-lingual settings. Furthermore, we use our code generation model to perform large-scale bootstrapping to obtain synthetic canonical solutions in several languages, which can be used for other code-related evaluations such as code insertion, robustness, or summarization tasks. Overall, our benchmarks represents a significant step towards a deeper understanding of language models' code generation abilities. We publicly release our code and datasets at https://github.com/amazon-research/mxeval.

Xiaokai Wei, Sujan Gonugondla, Wasi Ahmad et al. · amazon-science, ibm-research

ML-powered code generation aims to assist developers to write code in a more productive manner, by intelligently generating code blocks based on natural language prompts. Recently, large pretrained deep learning models have substantially pushed the boundary of code generation and achieved impressive performance. Despite their great power, the huge number of model parameters poses a significant threat to adapting them in a regular software development environment, where a developer might use a standard laptop or mid-size server to develop her code. Such large models incur significant resource usage (in terms of memory, latency, and dollars) as well as carbon footprint. Model compression is a promising approach to address these challenges. Several techniques are proposed to compress large pretrained models typically used for vision or textual data. Out of many available compression techniques, we identified that quantization is mostly applicable for code generation task as it does not require significant retraining cost. As quantization represents model parameters with lower-bit integer (e.g., int8), the model size and runtime latency would both benefit from such int representation. We extensively study the impact of quantized model on code generation tasks across different dimension: (i) resource usage and carbon footprint, (ii) accuracy, and (iii) robustness. To this end, through systematic experiments we find a recipe of quantization technique that could run even a $6$B model in a regular laptop without significant accuracy or robustness degradation. We further found the recipe is readily applicable to code summarization task as well.

Harman Singh, Xiuyu Li, Kusha Sareen et al. · berkeley

Test-time scaling for complex reasoning tasks shows that leveraging inference-time compute, by methods such as independently sampling and aggregating multiple solutions, results in significantly better task outcomes. However, a critical bottleneck is verification: sampling is only effective if correct solutions can be reliably identified among candidates. While existing approaches typically evaluate candidates independently via scalar scoring, we demonstrate that models are substantially stronger at pairwise self-verification. Leveraging this insight, we introduce $V_1$, a framework that unifies generation and verification through efficient pairwise ranking. $V_1$ comprises two components: $V_1$-Infer, an uncertainty-guided algorithm using a tournament-based ranking that dynamically allocates self-verification compute to candidate pairs whose relative correctness is most uncertain; and $V_1$-PairRL, an RL framework that jointly trains a single model as both generator and pairwise self-verifier, ensuring the verifier adapts to the generator's evolving distribution. On code generation (LiveCodeBench, CodeContests, SWE-Bench) and math reasoning (AIME, HMMT) benchmarks, $V_1$-Infer improves Pass@1 by up to $10%$ over pointwise verification and outperforms recent test-time scaling methods while being significantly more efficient. Furthermore, $V_1$-PairRL achieves $7$--$9%$ test-time scaling gains over standard RL and pointwise joint training, and improves base Pass@1 by up to 8.7% over standard RL in a code-generation setting.

Monishwaran Maheswaran, Leon Lakhani, Zhongzhu Zhou et al.

We show that verifier-free evolution is bottlenecked by both diversity and efficiency: without external correction, repeated evolution accelerates collapse toward narrow modes, while the uniform use of a high-cost model wastes compute and quickly becomes economically impractical. We introduce Squeeze Evolve, a unified multi-model orchestration framework for verifier-free evolutionary inference. Our approach is guided by a simple principle: allocate model capability where it has the highest marginal utility. Stronger models are reserved for high-impact stages, while cheaper models handle the other stages at much lower costs. This principle addresses diversity and cost-efficiency jointly while remaining lightweight. Squeeze Evolve naturally supports open-source, closed-source, and mixed-model deployments. Across AIME 2025, HMMT 2025, LiveCodeBench V6, GPQA-Diamond, ARC-AGI-V2, and multimodal vision benchmarks, such as MMMU-Pro and BabyVision, Squeeze Evolve consistently improves the cost-capability frontier over single-model evolution and achieves new state-of-the-art results on several tasks. Empirically, Squeeze Evolve reduces API cost by up to $\sim$3$\times$ and increases fixed-budget serving throughput by up to $\sim$10$\times$. Moreover, on discovery tasks, Squeeze Evolve is the first verifier-free evolutionary method to match, and in some cases exceed, the performance of verifier-based evolutionary methods.

Maurice Weber, Daniel Fu, Quentin Anthony et al.

Large language models are increasingly becoming a cornerstone technology in artificial intelligence, the sciences, and society as a whole, yet the optimal strategies for dataset composition and filtering remain largely elusive. Many of the top-performing models lack transparency in their dataset curation and model development processes, posing an obstacle to the development of fully open language models. In this paper, we identify three core data-related challenges that must be addressed to advance open-source language models. These include (1) transparency in model development, including the data curation process, (2) access to large quantities of high-quality data, and (3) availability of artifacts and metadata for dataset curation and analysis. To address these challenges, we release RedPajama-V1, an open reproduction of the LLaMA training dataset. In addition, we release RedPajama-V2, a massive web-only dataset consisting of raw, unfiltered text data together with quality signals and metadata. Together, the RedPajama datasets comprise over 100 trillion tokens spanning multiple domains and with their quality signals facilitate the filtering of data, aiming to inspire the development of numerous new datasets. To date, these datasets have already been used in the training of strong language models used in production, such as Snowflake Arctic, Salesforce's XGen and AI2's OLMo. To provide insight into the quality of RedPajama, we present a series of analyses and ablation studies with decoder-only language models with up to 1.6B parameters. Our findings demonstrate how quality signals for web data can be effectively leveraged to curate high-quality subsets of the dataset, underscoring the potential of RedPajama to advance the development of transparent and high-performing language models at scale.

James Liu, Pragaash Ponnusamy, Tianle Cai et al.

Activation sparsity can enable practical inference speedups in large language models (LLMs) by reducing the compute and memory-movement required for matrix multiplications during the forward pass. However, existing methods face limitations that inhibit widespread adoption. Some approaches are tailored towards older models with ReLU-based sparsity, while others require extensive continued pre-training on up to hundreds of billions of tokens. This paper describes TEAL, a simple training-free method that applies magnitude-based activation sparsity to hidden states throughout the entire model. TEAL achieves 40-50% model-wide sparsity with minimal performance degradation across Llama-2, Llama-3, and Mistral families, with sizes varying from 7B to 70B. We improve existing sparse kernels and demonstrate wall-clock decoding speed-ups of up to 1.53$\times$ and 1.8$\times$ at 40% and 50% model-wide sparsity. TEAL is compatible with weight quantization, enabling further efficiency gains.

Jon Saad-Falcon, Avanika Narayan, Hakki Orhun Akengin et al.

Large language model (LLM) queries are predominantly processed by frontier models in centralized cloud infrastructure. Rapidly growing demand strains this paradigm, and cloud providers struggle to scale infrastructure at pace. Two advances enable us to rethink this paradigm: small LMs (<=20B active parameters) now achieve competitive performance to frontier models on many tasks, and local accelerators (e.g., Apple M4 Max) run these models at interactive latencies. This raises the question: can local inference viably redistribute demand from centralized infrastructure? Answering this requires measuring whether local LMs can accurately answer real-world queries and whether they can do so efficiently enough to be practical on power-constrained devices (i.e., laptops). We propose intelligence per watt (IPW), task accuracy divided by unit of power, as a metric for assessing capability and efficiency of local inference across model-accelerator pairs. We conduct a large-scale empirical study across 20+ state-of-the-art local LMs, 8 accelerators, and a representative subset of LLM traffic: 1M real-world single-turn chat and reasoning queries. For each query, we measure accuracy, energy, latency, and power. Our analysis reveals $3$ findings. First, local LMs can accurately answer 88.7% of single-turn chat and reasoning queries with accuracy varying by domain. Second, from 2023-2025, IPW improved 5.3x and local query coverage rose from 23.2% to 71.3%. Third, local accelerators achieve at least 1.4x lower IPW than cloud accelerators running identical models, revealing significant headroom for optimization. These findings demonstrate that local inference can meaningfully redistribute demand from centralized infrastructure, with IPW serving as the critical metric for tracking this transition. We release our IPW profiling harness for systematic intelligence-per-watt benchmarking.

Yifan Yu, Yuqing Jian, Junxiong Wang et al.

Diffusion language models promise parallel generation, yet still lag behind autoregressive (AR) models in quality. We stem this gap to a failure of introspective consistency: AR models agree with their own generations, while DLMs often do not. We define the introspective acceptance rate, which measures whether a model accepts its previously generated tokens. This reveals why AR training has a structural advantage: causal masking and logit shifting implicitly enforce introspective consistency. Motivated by this observation, we introduce Introspective Diffusion Language Model (I-DLM), a paradigm that retains diffusion-style parallel decoding while inheriting the introspective consistency of AR training. I-DLM uses a novel introspective strided decoding (ISD) algorithm, which enables the model to verify previously generated tokens while advancing new ones in the same forward pass. From a systems standpoint, we build I-DLM inference engine on AR-inherited optimizations and further customize it with a stationary-batch scheduler. To the best of our knowledge, I-DLM is the first DLM to match the quality of its same-scale AR counterpart while outperforming prior DLMs in both model quality and practical serving efficiency across 15 benchmarks. It reaches 69.6 on AIME-24 and 45.7 on LiveCodeBench-v6, exceeding LLaDA-2.1-mini (16B) by more than 26 and 15 points, respectively. Beyond quality, I-DLM is designed for the growing demand of large-concurrency serving, delivering about 3x higher throughput than prior state-of-the-art DLMs.

Jinda Jia, Jisen Li, Zhongzhu Zhou et al.

KV-cache memory is a major bottleneck in real-world LLM serving, where systems must simultaneously support latency-sensitive small-batch requests and high-throughput concurrent workloads. Although many KV-cache compression methods improve offline accuracy or compression ratio, they often violate practical serving constraints such as paged memory layouts, regular memory access, and fused attention execution, limiting their effectiveness in deployment. In this work, we identify the minimal set of 4-bit KV-cache quantization methods that remain viable under these constraints. Our central finding is that a simple design--token-wise INT4 quantization with block-diagonal Hadamard rotation--consistently achieves the best accuracy-efficiency trade-off. Across multiple models and benchmarks, this approach recovers nearly all of the accuracy lost by naive INT4, while more complex methods such as vector quantization and Hessian-aware quantization provide only marginal additional gains once serving compatibility is taken into account. To make this practical, we implement a fused rotation-quantization kernel that integrates directly into paged KV-cache layouts and introduces zero measurable end-to-end overhead, matching plain INT4 throughput across concurrency levels. Our results show that effective KV-cache compression is fundamentally a systems co-design problem: under real serving constraints, lightweight block-diagonal Hadamard rotation is a viable method that delivers near-lossless accuracy without sacrificing serving efficiency.

Zhongzhu Zhou, Donglin Zhuang, Jisen Li et al.

INT2 KV-cache quantization is attractive for long-context LLM serving, but it remains difficult to make both accurate and deployable. Simple rotations such as Hadamard transforms reduce outliers, but still degrade at INT2 because they are not aligned with downstream attention. We propose OSCAR, an Ultra-low-bit KV Cache quantization method that estimates attention-aware covariance structures offline and uses them to derive fixed rotations and clipping thresholds for quantization. In this way, it aligns KV quantization with the covariance structures that attention actually consumes. More importantly, we not only provide theoretical justification but also develop a fully deployable OSCAR system with a custom INT2 attention kernel that remains compatible with paged KV-cache serving and fused kernel pipelines, enabling seamless integration into modern LLM serving frameworks such as SGLang and vLLM. We evaluate our methods on recent reasoning models with reasoning traces of up to 32k tokens across 5 tasks. On Qwen3-4B-Thinking-2507 and Qwen3-8B, OSCAR reduces the BF16 accuracy gap to 3.78 and 1.42 points, respectively, while naive rotation INT2 collapses to nearly zero. We further scale OSCAR to Qwen3-32B and GLM-4.7 (358B params), where it remains effectively on par with BF16. On long context - RULER-NIAH up to 128K, OSCAR remains robust on both Qwen3 models, while naive rotation INT2 collapses. System-wise, OSCAR reduces KV-cache memory by approximately 8x, improves throughput by up to 7x at large batch sizes under the same memory budget, and accelerates batch-size-1 decoding by up to 3x over BF16 due to reduced memory bandwidth overhead.

Zhongzhu Zhou, Fengxiang Bie, Ziyan Chen et al.

Converting pretrained attention modules such as grouped-query attention (GQA) into multi-head latent attention (MLA) can improve expressivity without increasing KV-cache cost, making it attractive for efficient inference. However, many practical conversion baselines rely on weight-only low-rank approximations (e.g., SVD-style initializations) and uniform rank allocation. They focus on minimizing the difference between weight matrices rather than on how those weights affect input activations, ignore the covariance structure of activations, and enforce uniform rank across layers, causing activation drift and degraded attention fidelity. To address these issues, we propose CARE, a Covariance-Aware, Rank-Enhanced MLA conversion pipeline under a fixed KV width. CARE introduces three key steps: (i) activation-preserving factorization, which aligns the approximation with the actual input activations rather than just the weights; (ii) adjusted-rank allocation, which spreads a fixed KV budget across layers by giving more capacity to layers that need it most; and (iii) KV-parity mapping, which reparameterizes the converted K and V to fit the MLA format while keeping the KV-cache size unchanged. Our method outperforms a uniform-rank SVD baseline on Qwen3-4B/30B-A3B-Instruct-2507 and Llama-3.1-8B/70B-Instruct, reducing one-shot perplexity by up to 215x and improving mean accuracy by up to 1.70x at matched KV budgets. With a brief post-SVD healing fine-tune, we fully recover the original model's accuracy.

Junlin Wang, Roy Xie, Shang Zhu et al.

Building helpful and harmless large language models (LLMs) requires effective model alignment approach based on human instructions and feedback, which necessitates high-quality human-labeled data. Constructing such datasets is often expensive and hard to scale, and may face potential limitations on diversity and generalization. To address these challenges, we introduce Mixture of Agents Alignment (MoAA), that leverages the collective strengths of various language models to provide high-quality data for model alignment. By employing MoAA, we enhance both supervised fine-tuning and preference optimization, leading to improved performance compared to using a single model alone to generate alignment data (e.g. using GPT-4o alone). Evaluation results show that our approach can improve win rate of LLaMA-3.1-8B-Instruct from 19.5 to 48.3 on Arena-Hard and from 22.33 to 57.23 on AlpacaEval2, highlighting a promising direction for model alignment through this new scalable and diverse synthetic data recipe. Furthermore, we demonstrate that MoAA enables a self-improvement pipeline, where models finetuned on MoA-generated data surpass their own initial capabilities, providing evidence that our approach can push the frontier of open-source LLMs without reliance on stronger external supervision. Data and code will be released.

Zhenyu Zhang, Xiaoxia Wu, Zhongzhu Zhou et al.

Large Language Models (LLMs) often rely on long chain-of-thought (CoT) reasoning to solve complex tasks. While effective, these trajectories are frequently inefficient, leading to high latency from excessive token generation, or unstable reasoning that alternates between underthinking (shallow, inconsistent steps) and overthinking (repetitive, verbose reasoning). In this work, we study the structure of reasoning trajectories and uncover specialized attention heads that correlate with distinct cognitive behaviors such as verification and backtracking. By lightly intervening on these heads at inference time, we can steer the model away from inefficient modes. Building on this insight, we propose CREST, a training-free method for Cognitive REasoning Steering at Test-time. CREST has two components: (1) an offline calibration step that identifies cognitive heads and derives head-specific steering vectors, and (2) an inference-time procedure that rotates hidden representations to suppress components along those vectors. CREST adaptively suppresses unproductive reasoning behaviors, yielding both higher accuracy and lower computational cost. Across diverse reasoning benchmarks and models, CREST improves accuracy by up to 17.5% while reducing token usage by 37.6%, offering a simple and effective pathway to faster, more reliable LLM reasoning.

Junxiong Wang, Fengxiang Bie, Jisen Li et al.

Speculative decoding can significantly accelerate LLM serving, yet most deployments today disentangle speculator training from serving, treating speculator training as a standalone offline modeling problem. We show that this decoupled formulation introduces substantial deployment and adaptation lag: (1) high time-to-serve, since a speculator must be trained offline for a considerable period before deployment; (2) delayed utility feedback, since the true end-to-end decoding speedup is only known after training and cannot be inferred reliably from acceptance rate alone due to model-architecture and system-level overheads; and (3) domain-drift degradation, as the target model is repurposed to new domains and the speculator becomes stale and less effective. To address these issues, we present Aurora, a unified training-serving system that closes the loop by continuously learning a speculator directly from live inference traces. Aurora reframes online speculator learning as an asynchronous reinforcement-learning problem: accepted tokens provide positive feedback, while rejected speculator proposals provide implicit negative feedback that we exploit to improve sample efficiency. Our design integrates an SGLang-based inference server with an asynchronous training server, enabling hot-swapped speculator updates without service interruption. Crucially, Aurora supports day-0 deployment: a speculator can be served immediately and rapidly adapted to live traffic, improving system performance while providing immediate utility feedback. Across experiments, Aurora achieves a 1.5x day-0 speedup on recently released frontier models (e.g., MiniMax M2.1 229B and Qwen3-Coder-Next 80B). Aurora also adapts effectively to distribution shifts in user traffic, delivering an additional 1.25x speedup over a well-trained but static speculator on widely used models (e.g., Qwen3 and Llama3).

Muru Zhang, Mayank Mishra, Zhongzhu Zhou et al.

Large language model inference is both memory-intensive and time-consuming, often requiring distributed algorithms to efficiently scale. Various model parallelism strategies are used in multi-gpu training and inference to partition computation across multiple devices, reducing memory load and computation time. However, using model parallelism necessitates communication of information between GPUs, which has been a major bottleneck and limits the gains obtained by scaling up the number of devices. We introduce Ladder Residual, a simple architectural modification applicable to all residual-based models that enables straightforward overlapping that effectively hides the latency of communication. Our insight is that in addition to systems optimization, one can also redesign the model architecture to decouple communication from computation. While Ladder Residual can allow communication-computation decoupling in conventional parallelism patterns, we focus on Tensor Parallelism in this paper, which is particularly bottlenecked by its heavy communication. For a Transformer model with 70B parameters, applying Ladder Residual to all its layers can achieve 29% end-to-end wall clock speed up at inference time with TP sharding over 8 devices. We refer the resulting Transformer model as the Ladder Transformer. We train a 1B and 3B Ladder Transformer from scratch and observe comparable performance to a standard dense transformer baseline. We also show that it is possible to convert parts of the Llama-3.1 8B model to our Ladder Residual architecture with minimal accuracy degradation by only retraining for 3B tokens. We release our code for training and inference for easier replication of experiments.

Costin-Andrei Oncescu, Qingyang Wu, Wai Tong Chung et al.

An increasing number of LLMs employ Mixture-of-Experts (MoE) architectures where the feed-forward layer is replaced by a pool of experts and each token only activates a small subset of them. During autoregressive generation, these models often enter a memory-bound regime even for moderate batch sizes because the average expert load grows more slowly than in an equivalent dense feedforward layer. Consequently, MoE latency is governed by the number of activated experts. We introduce a framework for dynamically re-routing token-to-expert mapping to lower this number (and thus, the decode latency) while preserving a comparable quality. Our best results use a batch-aware routing that works by having tokens piggyback experts that have already been loaded into memory due to being crucial to other tokens within the same batch. Empirically, we evaluate our method on the Qwen3-30B and Qwen3-235B models with a batch size of $16$. Without any statistically significant loss in accuracy, our approach achieves latency reductions of $39\%$ and $15\%$ in the MoE layer decode latency, respectively.

Linda He, Jue Wang, Maurice Weber et al.

Large Language Models (LLMs) struggle with long-context reasoning, not only due to the quadratic scaling of computational complexity with sequence length but also because of the scarcity and expense of annotating long-context data. There has been barely any open-source work that systematically ablates long-context data, nor is there any openly available instruction tuning dataset with contexts surpassing 100K tokens. To bridge this gap, we introduce a novel post-training synthetic data generation strategy designed to efficiently extend the context window of LLMs while preserving their general task performance. Our approach scalably extends to arbitrarily long context lengths, unconstrained by the length of available real-world data, which effectively addresses the scarcity of raw long-context data. Through a step-by-step rotary position embedding (RoPE) scaling training strategy, we demonstrate that our model, with a context length of up to 1M tokens, performs well on the RULER benchmark and InfiniteBench and maintains robust performance on general language tasks.

Tanmaey Gupta, Hayden Prairie, Xiaoxia Wu et al.

Quantization has emerged as a standard technique for accelerating inference for generative models by enabling faster low-precision computations and reduced memory transfers. Recently, GPU accelerators have added first-class support for microscaling Block Floating Point (BFP) formats. Standard BFP algorithms use a fixed scale based on the maximum magnitude of the block. We observe that this scale choice can be suboptimal with respect to quantization errors. In this work, we propose ScaleSearch, an alternative strategy for selecting these scale factors: using a fine-grained search leveraging the mantissa bits in microscaling formats to minimize the quantization error for the given distribution. ScaleSearch can be integrated with existing quantization methods such as Post Training Quantization and low-precision attention, and is shown to improve their performance. Additionally, we introduce ScaleSearchAttention, an accelerated NVFP4-based attention algorithm, which uses ScaleSearch and adapted prior techniques to ensure near-0 performance loss for causal language modeling. Experiments show that ScaleSearch reduces quantization error by 27% for NVFP4 and improves language model PTQ by up to 15 points for MATH500 (Qwen3-8B), while ScaleSearchAttention improves Wikitext-2 PPL by upto 0.77 points for Llama 3.1 70B. The proposed methods closely match baseline performance while providing quantization accuracy improvements.

Minseo Kim, Chenfeng Xu, Coleman Hooper et al.

Diffusion Language Models (DLMs) offer a promising parallel generation paradigm but suffer from slow inference due to numerous refinement steps and the inability to use standard KV caching. We introduce CDLM (Consistency Diffusion Language Models), a training-based acceleration method that simultaneously tackles both bottlenecks. CDLM integrates consistency modeling to drastically reduce the number of required sampling steps by enabling multi-token finalization. Furthermore, we enforce a block-wise causal attention mask during fine-tuning, making the model fully compatible with KV caching. Experiments show CDLM achieves 3.6x-14.5x lower latency while maintaining competitive accuracy on math and coding tasks. The full training and evaluation code is available at https://github.com/SqueezeAILab/CDLM.

Haojun Xia, Xiaoxia Wu, Jisen Li et al.

The KV cache is a dominant memory bottleneck for LLM inference. While 4-bit KV quantization preserves accuracy, 2-bit often degrades it, especially on long-context reasoning. We close this gap via an algorithm-system co-design for mixed-precision KV caching: Kitty. On the algorithm side, extensive experiments show that Dynamic Channel-wise Precision Boost -- which ranks Key-cache channels by sensitivity and keeps only a small fraction at higher precision -- maintains near-zero loss in accuracy drop while approaching 2-bit memory. The main challenge is handling dynamic 4-bit channel boosts while keeping the page layout coalesced and the dequantization uniform, with no scattered reads or hard-coded masks. Kitty addresses these issues by decompose each mixed-precision Key page into two tensors with unified 2-bit precision. Based on this, Kitty provides a page-centric KV layout, Triton-compatible page dequantization kernels, and a lightweight runtime pipeline that preserves coalescing and avoids divergence. Across seven tasks and two model families (Qwen3, LLaMA3), Kitty cuts KV memory by nearly 8x with negligible accuracy loss, enabling up to 8x larger batches and 2.1x-4.1x higher throughput under the same memory budget. We release the full implementation of Kitty at https://github.com/Summer-Summer/Kitty.

Junlin Wang, Jue Wang, Ben Athiwaratkun et al.

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.

Junlin Wang, Shang Zhu, Jon Saad-Falcon et al.

There is intense interest in investigating how inference time compute (ITC) (e.g. repeated sampling, refinements, etc) can improve large language model (LLM) capabilities. At the same time, recent breakthroughs in reasoning models, such as Deepseek-R1, unlock the opportunity for reinforcement learning to improve LLM reasoning skills. An in-depth understanding of how ITC interacts with reasoning across different models could provide important guidance on how to further advance the LLM frontier. This work conducts a comprehensive analysis of inference-time scaling methods for both reasoning and non-reasoning models on challenging reasoning tasks. Specifically, we focus our research on verifier-free inference time-scaling methods due to its generalizability without needing a reward model. We construct the Pareto frontier of quality and efficiency. We find that non-reasoning models, even with an extremely high inference budget, still fall substantially behind reasoning models. For reasoning models, majority voting proves to be a robust inference strategy, generally competitive or outperforming other more sophisticated ITC methods like best-of-N and sequential revisions, while the additional inference compute offers minimal improvements. We further perform in-depth analyses of the association of key response features (length and linguistic markers) with response quality, with which we can improve the existing ITC methods. We find that correct responses from reasoning models are typically shorter and have fewer hedging and thinking markers (but more discourse markers) than the incorrect responses.

Ben Athiwaratkun, Shiqi Wang, Mingyue Shang et al. · amazon-science

Generative models, widely utilized in various applications, can often struggle with prompts corresponding to partial tokens. This struggle stems from tokenization, where partial tokens fall out of distribution during inference, leading to incorrect or nonsensical outputs. This paper examines a technique to alleviate the tokenization artifact on text completion in generative models, maintaining performance even in regular non-subword cases. The method, termed token alignment, involves backtracking to the last complete tokens and ensuring the model's generation aligns with the prompt. This approach showcases marked improvement across many partial token scenarios, including nuanced cases like space-prefix and partial indentation, with only a minor time increase. The technique and analysis detailed in this paper contribute to the continuous advancement of generative models in handling partial inputs, bearing relevance for applications like code completion and text autocompletion.

Jon Saad-Falcon, E. Kelly Buchanan, Mayee F. Chen et al.

Verifiers can improve language model capabilities by scoring and ranking responses from generated candidates. Currently, high-quality verifiers are either unscalable (e.g., humans) or limited in utility (e.g., tools like Lean). While LM judges and reward models have become broadly useful as general-purpose verifiers, a significant performance gap remains between them and oracle verifiers (verifiers with perfect accuracy). To help close this gap, we introduce Weaver, a framework for designing a strong verifier by combining multiple weak, imperfect verifiers. We find weighted ensembles of verifiers, which typically require learning from labeled data, significantly outperform unweighted combinations due to differences in verifier accuracies. To reduce dependency on labeled data, Weaver leverages weak supervision to estimate each verifier's accuracy and combines outputs into a unified score that better reflects true response quality. However, directly applying weak supervision algorithms poses challenges, including inconsistent verifier output formats and handling low-quality verifiers. Weaver addresses these using dataset statistics to normalize outputs and filter specific verifiers. We study Weaver's effectiveness in test-time repeated sampling, where a model generates multiple candidate responses and selects one. Our evaluations show Weaver significantly improves over Pass@1-performance when selecting the first candidate-across reasoning and math tasks, achieving o3-mini-level accuracy with Llama 3.3 70B Instruct as generator, and an ensemble of 70B or smaller judge and reward models as verifiers (87.7% average). This gain mirrors the jump between GPT-4o and o3-mini (69.0% vs. 86.7%), which required extensive finetuning and post-training. To reduce computational costs of verifier ensembles, we train a 400M cross-encoder using Weaver's combined output scores.

Ben Athiwaratkun, Sujan Kumar Gonugondla, Sanjay Krishna Gouda et al. · amazon-science

This study introduces bifurcated attention, a method designed to enhance language model inference in shared-context batch decoding scenarios. Our approach addresses the challenge of redundant memory IO costs, a critical factor contributing to latency in high batch sizes and extended context lengths. Bifurcated attention achieves this by strategically dividing the attention mechanism during incremental decoding into two separate GEMM operations: one focusing on the KV cache from prefill, and another on the decoding process itself. While maintaining the computational load (FLOPs) of standard attention mechanisms, bifurcated attention ensures precise computation with significantly reduced memory IO. Our empirical results show over 2.1$\times$ speedup when sampling 16 output sequences and more than 6.2$\times$ speedup when sampling 32 sequences at context lengths exceeding 8k tokens on a 7B model that uses multi-head attention. The efficiency gains from bifurcated attention translate into lower latency, making it particularly suitable for real-time applications. For instance, it enables massively parallel answer generation without substantially increasing latency, thus enhancing performance when integrated with post-processing techniques such as re-ranking.

Zhen Xu, Shang Zhu, Jue Wang et al.

We investigate the challenge of applying Large Language Models (LLMs) to long texts. We propose a theoretical framework that distinguishes the failure modes of long context tasks into three categories: cross-chunk dependence (task noise), confusion that grows with context size (model noise), and the imperfect integration of partial results (aggregator noise). Under this view, we analyze when it is effective to use multi-agent chunking, i.e., dividing a length sequence into smaller chunks and aggregating the processed results of each chunk. Our experiments on tasks such as retrieval, question answering, and summarization confirm both the theoretical analysis and the conditions that favor multi-agent chunking. By exploring superlinear model noise growth with input length, we also explain why, for large inputs, a weaker model configured with chunk-based processing can surpass a more advanced model like GPT4o applied in a single shot. Overall, we present a principled understanding framework and our results highlight a direct pathway to handling long contexts in LLMs with carefully managed chunking and aggregator strategies.

Rahul Thapa, Andrew Li, Qingyang Wu et al.

Publicly available biomedical videos, such as those on YouTube, serve as valuable educational resources for medical students. Unlike standard machine learning datasets, these videos are designed for human learners, often mixing medical imagery with narration, explanatory diagrams, and contextual framing. In this work, we investigate whether such pedagogically rich, yet non-standardized and heterogeneous videos can effectively teach general-domain vision-language models biomedical knowledge. To this end, we introduce OpenBiomedVi, a biomedical video instruction tuning dataset comprising 1031 hours of video-caption and Q/A pairs, curated through a multi-step human-in-the-loop pipeline. Diverse biomedical video datasets are rare, and OpenBiomedVid fills an important gap by providing instruction-style supervision grounded in real-world educational content. Surprisingly, despite the informal and heterogeneous nature of these videos, the fine-tuned Qwen-2-VL models exhibit substantial performance improvements across most benchmarks. The 2B model achieves gains of 98.7% on video tasks, 71.2% on image tasks, and 0.2% on text tasks. The 7B model shows improvements of 37.09% on video and 11.2% on image tasks, with a slight degradation of 2.7% on text tasks compared to their respective base models. To address the lack of standardized biomedical video evaluation datasets, we also introduce two new expert curated benchmarks, MIMICEchoQA and SurgeryVideoQA. On these benchmarks, the 2B model achieves gains of 99.1% and 98.1%, while the 7B model shows gains of 22.5% and 52.1%, respectively, demonstrating the models' ability to generalize and perform biomedical video understanding on cleaner and more standardized datasets than those seen during training. These results suggest that educational videos created for human learning offer a surprisingly effective training signal for biomedical VLMs.

Zelei Shao, Vikranth Srivatsa, Sanjana Srivastava et al.

Reinforcement learning(RL) post-training has become essential for aligning large language models (LLMs), yet its efficiency is increasingly constrained by the rollout phase, where long trajectories are generated token by token. We identify a major bottleneck:the long-tail distribution of rollout lengths, where a small fraction of long generations dominates wall clock time and a complementary opportunity; the availability of historical rollouts that reveal stable prompt level patterns across training epochs. Motivated by these observations, we propose DAS, a Distribution Aware Speculative decoding framework that accelerates RL rollouts without altering model outputs. DAS integrates two key ideas: an adaptive, nonparametric drafter built from recent rollouts using an incrementally maintained suffix tree, and a length aware speculation policy that allocates more aggressive draft budgets to long trajectories that dominate makespan. This design exploits rollout history to sustain acceptance while balancing base and token level costs during decoding. Experiments on math and code reasoning tasks show that DAS reduces rollout time up to 50% while preserving identical training curves, demonstrating that distribution-aware speculative decoding can significantly accelerate RL post training without compromising learning quality.

Junlin Wang, Jue Wang, Zhen et al.

Recent advances in large language models (LLMs) opened up new directions for leveraging the collective expertise of multiple LLMs. These methods, such as Mixture-of-Agents, typically employ additional inference steps to generate intermediate outputs, which are then used to produce the final response. While multi-agent inference can enhance response quality, it can significantly increase the time to first token (TTFT), posing a challenge for latency-sensitive applications and hurting user experience. To address this issue, we propose staircase streaming for low-latency multi-agent inference. Instead of waiting for the complete intermediate outputs from previous steps, we begin generating the final response as soon as we receive partial outputs from these steps. Experimental results demonstrate that staircase streaming reduces TTFT by up to 93% while maintaining response quality.

Zhongzhu Zhou, Yibo Yang, Ziyan Chen et al.

Policy gradient (PG) methods in reinforcement learning frequently utilize deep neural networks (DNNs) to learn a shared backbone of feature representations used to compute likelihoods in an action selection layer. Numerous studies have been conducted on the convergence and global optima of policy networks, but few have analyzed representational structures of those underlying networks. While training an optimal policy DNN, we observed that under certain constraints, a gentle structure resembling neural collapse, which we refer to as Action Collapse (AC), emerges. This suggests that 1) the state-action activations (i.e. last-layer features) sharing the same optimal actions collapse towards those optimal actions respective mean activations; 2) the variability of activations sharing the same optimal actions converges to zero; 3) the weights of action selection layer and the mean activations collapse to a simplex equiangular tight frame (ETF). Our early work showed those aforementioned constraints to be necessary for these observations. Since the collapsed ETF of optimal policy DNNs maximally separates the pair-wise angles of all actions in the state-action space, we naturally raise a question: can we learn an optimal policy using an ETF structure as a (fixed) target configuration in the action selection layer? Our analytical proof shows that learning activations with a fixed ETF as action selection layer naturally leads to the AC. We thus propose the Action Collapse Policy Gradient (ACPG) method, which accordingly affixes a synthetic ETF as our action selection layer. ACPG induces the policy DNN to produce such an ideal configuration in the action selection layer while remaining optimal. Our experiments across various OpenAI Gym environments demonstrate that our technique can be integrated into any discrete PG methods and lead to favorable reward improvements more quickly and robustly.

Berkan Dokmeci, Qingyang Wu, Ben Athiwaratkun et al.

While recent advances in preference learning have enhanced alignment in human feedback, mathematical reasoning remains a persistent challenge. We investigate how data diversification strategies in preference optimization can improve the mathematical reasoning abilities of large language models (LLMs). We evaluate three common data generation methods: temperature sampling, Chain-of-Thought prompting, and Monte Carlo Tree Search (MCTS), and introduce Diversified-ThinkSolve (DTS), a novel structured approach that systematically decomposes problems into diverse reasoning paths. Our results show that with strategically diversified preference data, models can substantially improve mathematical reasoning performance, with the best approach yielding gains of 7.1% on GSM8K and 4.2% on MATH over the base model. Despite its strong performance, DTS incurs only a marginal computational overhead (1.03x) compared to the baseline, while MCTS is nearly five times more costly with lower returns. These findings demonstrate that structured exploration of diverse problem-solving methods creates more effective preference data for mathematical alignment than traditional approaches.

Junlin Wang, Siddhartha Jain, Dejiao Zhang et al.

A diverse array of reasoning strategies has been proposed to elicit the capabilities of large language models. However, in this paper, we point out that traditional evaluations which focus solely on performance metrics miss a key factor: the increased effectiveness due to additional compute. By overlooking this aspect, a skewed view of strategy efficiency is often presented. This paper introduces a framework that incorporates the compute budget into the evaluation, providing a more informative comparison that takes into account both performance metrics and computational cost. In this budget-aware perspective, we find that complex reasoning strategies often don't surpass simpler baselines purely due to algorithmic ingenuity, but rather due to the larger computational resources allocated. When we provide a simple baseline like chain-of-thought self-consistency with comparable compute resources, it frequently outperforms reasoning strategies proposed in the literature. In this scale-aware perspective, we find that unlike self-consistency, certain strategies such as multi-agent debate or Reflexion can become worse if more compute budget is utilized.

Rahul Thapa, Kezhen Chen, Ian Covert et al.

Recent advances in vision-language models (VLMs) have demonstrated the advantages of processing images at higher resolutions and utilizing multi-crop features to preserve native resolution details. However, despite these improvements, existing vision transformers (ViTs) still struggle to capture fine-grained details from less prominent objects, charts, and embedded text, limiting their effectiveness in certain tasks. In this paper, we extend recent high-resolution and multi-crop techniques by not only preserving the native resolution, but zooming in beyond it and extracting features from a large number of image sub-crops. This enhancement allows our model to better capture fine-grained details, overcoming the limitations of current ViTs. To manage the increased token count and computational complexity, we demonstrate that a simple mean-pooling aggregation over tokens is effective. Our model, Dragonfly, achieves competitive performance on general-domain tasks such as ScienceQA and AI2D, and excels in tasks requiring fine-grained image understanding, including TextVQA and ChartQA. Among models in the 7-8B parameter range, Dragonfly consistently ranks at the top across ten general-domain benchmarks, achieving the highest or second-highest scores in most cases, outperforming models that are significantly larger or trained on larger datasets. Our biomedical model, Dragonfly-Med, sets new benchmarks on several medical tasks, achieving 91.6% accuracy on SLAKE (compared to 84.8% for Med-Gemini), a 67.1% token F1 score on Path-VQA (compared to 62.7% for Med-PaLM M), and state-of-the-art results across the majority of image captioning tasks. Overall, our work highlights the persistent challenge of engineering visual representations with fixed-resolution ViTs, and proposes a simple yet effective solution to address this issue and boost performance in both general and specialized domains.

Yang Li, Ben Athiwaratkun, Cicero Nogueira dos Santos et al.

Generalization is a central problem in machine learning, especially when data is limited. Using prior information to enforce constraints is the principled way of encouraging generalization. In this work, we propose to leverage the prior information embedded in pretrained language models (LM) to improve generalization for intent classification and slot labeling tasks with limited training data. Specifically, we extract prior knowledge from pretrained LM in the form of synthetic data, which encode the prior implicitly. We fine-tune the LM to generate an augmented language, which contains not only text but also encodes both intent labels and slot labels. The generated synthetic data can be used to train a classifier later. Since the generated data may contain noise, we rephrase the learning from generated data as learning with noisy labels. We then utilize the mixout regularization for the classifier and prove its effectiveness to resist label noise in generated data. Empirically, our method demonstrates superior performance and outperforms the baseline by a large margin.

Dheeru Dua, Cicero Nogueira dos Santos, Patrick Ng et al.

Compositional reasoning tasks like multi-hop question answering, require making latent decisions to get the final answer, given a question. However, crowdsourced datasets often capture only a slice of the underlying task distribution, which can induce unanticipated biases in models performing compositional reasoning. Furthermore, discriminatively trained models exploit such biases to get a better held-out performance, without learning the right way to reason, as they do not necessitate paying attention to the question representation (conditioning variable) in its entirety, to estimate the answer likelihood. In this work, we propose a generative context selection model for multi-hop question answering that reasons about how the given question could have been generated given a context pair. While being comparable to the state-of-the-art answering performance, our proposed generative passage selection model has a better performance (4.9% higher than baseline) on adversarial held-out set which tests robustness of model's multi-hop reasoning capabilities.

Giovanni Paolini, Ben Athiwaratkun, Jason Krone et al.

We propose a new framework, Translation between Augmented Natural Languages (TANL), to solve many structured prediction language tasks including joint entity and relation extraction, nested named entity recognition, relation classification, semantic role labeling, event extraction, coreference resolution, and dialogue state tracking. Instead of tackling the problem by training task-specific discriminative classifiers, we frame it as a translation task between augmented natural languages, from which the task-relevant information can be easily extracted. Our approach can match or outperform task-specific models on all tasks, and in particular, achieves new state-of-the-art results on joint entity and relation extraction (CoNLL04, ADE, NYT, and ACE2005 datasets), relation classification (FewRel and TACRED), and semantic role labeling (CoNLL-2005 and CoNLL-2012). We accomplish this while using the same architecture and hyperparameters for all tasks and even when training a single model to solve all tasks at the same time (multi-task learning). Finally, we show that our framework can also significantly improve the performance in a low-resource regime, thanks to better use of label semantics.

Ben Athiwaratkun, Cicero Nogueira dos Santos, Jason Krone et al.

We propose a generative framework for joint sequence labeling and sentence-level classification. Our model performs multiple sequence labeling tasks at once using a single, shared natural language output space. Unlike prior discriminative methods, our model naturally incorporates label semantics and shares knowledge across tasks. Our framework is general purpose, performing well on few-shot, low-resource, and high-resource tasks. We demonstrate these advantages on popular named entity recognition, slot labeling, and intent classification benchmarks. We set a new state-of-the-art for few-shot slot labeling, improving substantially upon the previous 5-shot ($75.0\% \rightarrow 90.9\%$) and 1-shot ($70.4\% \rightarrow 81.0\%$) state-of-the-art results. Furthermore, our model generates large improvements ($46.27\% \rightarrow 63.83\%$) in low-resource slot labeling over a BERT baseline by incorporating label semantics. We also maintain competitive results on high-resource tasks, performing within two points of the state-of-the-art on all tasks and setting a new state-of-the-art on the SNIPS dataset.

Ben Athiwaratkun, Marc Finzi, Pavel Izmailov et al.

Presently the most successful approaches to semi-supervised learning are based on consistency regularization, whereby a model is trained to be robust to small perturbations of its inputs and parameters. To understand consistency regularization, we conceptually explore how loss geometry interacts with training procedures. The consistency loss dramatically improves generalization performance over supervised-only training; however, we show that SGD struggles to converge on the consistency loss and continues to make large steps that lead to changes in predictions on the test data. Motivated by these observations, we propose to train consistency-based methods with Stochastic Weight Averaging (SWA), a recent approach which averages weights along the trajectory of SGD with a modified learning rate schedule. We also propose fast-SWA, which further accelerates convergence by averaging multiple points within each cycle of a cyclical learning rate schedule. With weight averaging, we achieve the best known semi-supervised results on CIFAR-10 and CIFAR-100, over many different quantities of labeled training data. For example, we achieve 5.0% error on CIFAR-10 with only 4000 labels, compared to the previous best result in the literature of 6.3%.

Ben Athiwaratkun, Andrew Gordon Wilson, Anima Anandkumar

We introduce Probabilistic FastText, a new model for word embeddings that can capture multiple word senses, sub-word structure, and uncertainty information. In particular, we represent each word with a Gaussian mixture density, where the mean of a mixture component is given by the sum of n-grams. This representation allows the model to share statistical strength across sub-word structures (e.g. Latin roots), producing accurate representations of rare, misspelt, or even unseen words. Moreover, each component of the mixture can capture a different word sense. Probabilistic FastText outperforms both FastText, which has no probabilistic model, and dictionary-level probabilistic embeddings, which do not incorporate subword structures, on several word-similarity benchmarks, including English RareWord and foreign language datasets. We also achieve state-of-art performance on benchmarks that measure ability to discern different meanings. Thus, the proposed model is the first to achieve multi-sense representations while having enriched semantics on rare words.

Ben Athiwaratkun, Andrew Gordon Wilson

By representing words with probability densities rather than point vectors, probabilistic word embeddings can capture rich and interpretable semantic information and uncertainty. The uncertainty information can be particularly meaningful in capturing entailment relationships -- whereby general words such as "entity" correspond to broad distributions that encompass more specific words such as "animal" or "instrument". We introduce density order embeddings, which learn hierarchical representations through encapsulation of probability densities. In particular, we propose simple yet effective loss functions and distance metrics, as well as graph-based schemes to select negative samples to better learn hierarchical density representations. Our approach provides state-of-the-art performance on the WordNet hypernym relationship prediction task and the challenging HyperLex lexical entailment dataset -- while retaining a rich and interpretable density representation.

Ben Athiwaratkun, Andrew Gordon Wilson

Word embeddings provide point representations of words containing useful semantic information. We introduce multimodal word distributions formed from Gaussian mixtures, for multiple word meanings, entailment, and rich uncertainty information. To learn these distributions, we propose an energy-based max-margin objective. We show that the resulting approach captures uniquely expressive semantic information, and outperforms alternatives, such as word2vec skip-grams, and Gaussian embeddings, on benchmark datasets such as word similarity and entailment.

Xilun Chen, Yu Sun, Ben Athiwaratkun et al.

In recent years great success has been achieved in sentiment classification for English, thanks in part to the availability of copious annotated resources. Unfortunately, most languages do not enjoy such an abundance of labeled data. To tackle the sentiment classification problem in low-resource languages without adequate annotated data, we propose an Adversarial Deep Averaging Network (ADAN) to transfer the knowledge learned from labeled data on a resource-rich source language to low-resource languages where only unlabeled data exists. ADAN has two discriminative branches: a sentiment classifier and an adversarial language discriminator. Both branches take input from a shared feature extractor to learn hidden representations that are simultaneously indicative for the classification task and invariant across languages. Experiments on Chinese and Arabic sentiment classification demonstrate that ADAN significantly outperforms state-of-the-art systems.

Ben Athiwaratkun, Keegan Kang

Convolutional Neural Networks (CNNs) are powerful models that achieve impressive results for image classification. In addition, pre-trained CNNs are also useful for other computer vision tasks as generic feature extractors. This paper aims to gain insight into the feature aspect of CNN and demonstrate other uses of CNN features. Our results show that CNN feature maps can be used with Random Forests and SVM to yield classification results that outperforms the original CNN. A CNN that is less than optimal (e.g. not fully trained or overfitting) can also extract features for Random Forest/SVM that yield competitive classification accuracy. In contrast to the literature which uses the top-layer activations as feature representation of images for other tasks, using lower-layer features can yield better results for classification.