Ion Stoica

Lianmin Zheng, Wei-Lin Chiang, Ying Sheng et al. · berkeley

Evaluating large language model (LLM) based chat assistants is challenging due to their broad capabilities and the inadequacy of existing benchmarks in measuring human preferences. To address this, we explore using strong LLMs as judges to evaluate these models on more open-ended questions. We examine the usage and limitations of LLM-as-a-judge, including position, verbosity, and self-enhancement biases, as well as limited reasoning ability, and propose solutions to mitigate some of them. We then verify the agreement between LLM judges and human preferences by introducing two benchmarks: MT-bench, a multi-turn question set; and Chatbot Arena, a crowdsourced battle platform. Our results reveal that strong LLM judges like GPT-4 can match both controlled and crowdsourced human preferences well, achieving over 80% agreement, the same level of agreement between humans. Hence, LLM-as-a-judge is a scalable and explainable way to approximate human preferences, which are otherwise very expensive to obtain. Additionally, we show our benchmark and traditional benchmarks complement each other by evaluating several variants of LLaMA and Vicuna. The MT-bench questions, 3K expert votes, and 30K conversations with human preferences are publicly available at https://github.com/lm-sys/FastChat/tree/main/fastchat/llm_judge.

Woosuk Kwon, Zhuohan Li, Siyuan Zhuang et al. · berkeley

High throughput serving of large language models (LLMs) requires batching sufficiently many requests at a time. However, existing systems struggle because the key-value cache (KV cache) memory for each request is huge and grows and shrinks dynamically. When managed inefficiently, this memory can be significantly wasted by fragmentation and redundant duplication, limiting the batch size. To address this problem, we propose PagedAttention, an attention algorithm inspired by the classical virtual memory and paging techniques in operating systems. On top of it, we build vLLM, an LLM serving system that achieves (1) near-zero waste in KV cache memory and (2) flexible sharing of KV cache within and across requests to further reduce memory usage. Our evaluations show that vLLM improves the throughput of popular LLMs by 2-4$\times$ with the same level of latency compared to the state-of-the-art systems, such as FasterTransformer and Orca. The improvement is more pronounced with longer sequences, larger models, and more complex decoding algorithms. vLLM's source code is publicly available at https://github.com/vllm-project/vllm

Lianmin Zheng, Wei-Lin Chiang, Ying Sheng et al. · berkeley

Studying how people interact with large language models (LLMs) in real-world scenarios is increasingly important due to their widespread use in various applications. In this paper, we introduce LMSYS-Chat-1M, a large-scale dataset containing one million real-world conversations with 25 state-of-the-art LLMs. This dataset is collected from 210K unique IP addresses in the wild on our Vicuna demo and Chatbot Arena website. We offer an overview of the dataset's content, including its curation process, basic statistics, and topic distribution, highlighting its diversity, originality, and scale. We demonstrate its versatility through four use cases: developing content moderation models that perform similarly to GPT-4, building a safety benchmark, training instruction-following models that perform similarly to Vicuna, and creating challenging benchmark questions. We believe that this dataset will serve as a valuable resource for understanding and advancing LLM capabilities. The dataset is publicly available at https://huggingface.co/datasets/lmsys/lmsys-chat-1m.

Shishir G. Patil, Paras Jain, Prabal Dutta et al. · berkeley

Fine-tuning models on edge devices like mobile phones would enable privacy-preserving personalization over sensitive data. However, edge training has historically been limited to relatively small models with simple architectures because training is both memory and energy intensive. We present POET, an algorithm to enable training large neural networks on memory-scarce battery-operated edge devices. POET jointly optimizes the integrated search search spaces of rematerialization and paging, two algorithms to reduce the memory consumption of backpropagation. Given a memory budget and a run-time constraint, we formulate a mixed-integer linear program (MILP) for energy-optimal training. Our approach enables training significantly larger models on embedded devices while reducing energy consumption while not modifying mathematical correctness of backpropagation. We demonstrate that it is possible to fine-tune both ResNet-18 and BERT within the memory constraints of a Cortex-M class embedded device while outperforming current edge training methods in energy efficiency. POET is an open-source project available at https://github.com/ShishirPatil/poet

Ying Sheng, Lianmin Zheng, Binhang Yuan et al.

The high computational and memory requirements of large language model (LLM) inference make it feasible only with multiple high-end accelerators. Motivated by the emerging demand for latency-insensitive tasks with batched processing, this paper initiates the study of high-throughput LLM inference using limited resources, such as a single commodity GPU. We present FlexGen, a high-throughput generation engine for running LLMs with limited GPU memory. FlexGen can be flexibly configured under various hardware resource constraints by aggregating memory and computation from the GPU, CPU, and disk. By solving a linear programming problem, it searches for efficient patterns to store and access tensors. FlexGen further compresses the weights and the attention cache to 4 bits with negligible accuracy loss. These techniques enable FlexGen to have a larger space of batch size choices and thus significantly increase maximum throughput. As a result, when running OPT-175B on a single 16GB GPU, FlexGen achieves significantly higher throughput compared to state-of-the-art offloading systems, reaching a generation throughput of 1 token/s for the first time with an effective batch size of 144. On the HELM benchmark, FlexGen can benchmark a 30B model with a 16GB GPU on 7 representative sub-scenarios in 21 hours. The code is available at https://github.com/FMInference/FlexGen

Dacheng Li, Rulin Shao, Anze Xie et al. · cmu

FlashAttention (Dao, 2023) effectively reduces the quadratic peak memory usage to linear in training transformer-based large language models (LLMs) on a single GPU. In this paper, we introduce DISTFLASHATTN, a distributed memory-efficient attention mechanism optimized for long-context LLMs training. We propose three key techniques: token-level workload balancing, overlapping key-value communication, and a rematerialization-aware gradient checkpointing algorithm. We evaluate DISTFLASHATTN on Llama-7B and variants with sequence lengths from 32K to 512K. DISTFLASHATTN achieves 8x longer sequences, 4.45 - 5.64x speedup compared to Ring Self-Attention, 2 - 8x longer sequences, 1.24 - 2.01x speedup compared to Megatron-LM with FlashAttention. It achieves 1.67x and 1.26 - 1.88x speedup compared to recent Ring Attention and DeepSpeed-Ulysses. Code is available at https://github.com/RulinShao/LightSeq.

Daniel Kang, Xuechen Li, Ion Stoica et al. · stanford

Recent advances in instruction-following large language models (LLMs) have led to dramatic improvements in a range of NLP tasks. Unfortunately, we find that the same improved capabilities amplify the dual-use risks for malicious purposes of these models. Dual-use is difficult to prevent as instruction-following capabilities now enable standard attacks from computer security. The capabilities of these instruction-following LLMs provide strong economic incentives for dual-use by malicious actors. In particular, we show that instruction-following LLMs can produce targeted malicious content, including hate speech and scams, bypassing in-the-wild defenses implemented by LLM API vendors. Our analysis shows that this content can be generated economically and at cost likely lower than with human effort alone. Together, our findings suggest that LLMs will increasingly attract more sophisticated adversaries and attacks, and addressing these attacks may require new approaches to mitigations.

Abdus Salam Azad, Izzeddin Gur, Jasper Emhoff et al. · berkeley

Reinforcement Learning (RL) algorithms are often known for sample inefficiency and difficult generalization. Recently, Unsupervised Environment Design (UED) emerged as a new paradigm for zero-shot generalization by simultaneously learning a task distribution and agent policies on the generated tasks. This is a non-stationary process where the task distribution evolves along with agent policies; creating an instability over time. While past works demonstrated the potential of such approaches, sampling effectively from the task space remains an open challenge, bottlenecking these approaches. To this end, we introduce CLUTR: a novel unsupervised curriculum learning algorithm that decouples task representation and curriculum learning into a two-stage optimization. It first trains a recurrent variational autoencoder on randomly generated tasks to learn a latent task manifold. Next, a teacher agent creates a curriculum by maximizing a minimax REGRET-based objective on a set of latent tasks sampled from this manifold. Using the fixed-pretrained task manifold, we show that CLUTR successfully overcomes the non-stationarity problem and improves stability. Our experimental results show CLUTR outperforms PAIRED, a principled and popular UED method, in the challenging CarRacing and navigation environments: achieving 10.6X and 45\% improvement in zero-shot generalization, respectively. CLUTR also performs comparably to the non-UED state-of-the-art for CarRacing, while requiring 500X fewer environment interactions.

Shuo Yang, Wei-Lin Chiang, Lianmin Zheng et al.

Large language models are increasingly trained on all the data ever produced by humans. Many have raised concerns about the trustworthiness of public benchmarks due to potential contamination in pre-training or fine-tuning datasets. While most data decontamination efforts apply string matching (e.g., n-gram overlap) to remove benchmark data, we show that these methods are insufficient, and simple variations of test data (e.g., paraphrasing, translation) can easily bypass these decontamination measures. Furthermore, we demonstrate that if such variation of test data is not eliminated, a 13B model can easily overfit a test benchmark and achieve drastically high performance, on par with GPT-4. We validate such observations in widely used benchmarks such as MMLU, GSK8k, and HumanEval. To address this growing risk, we propose a stronger LLM-based decontamination method and apply it to widely used pre-training and fine-tuning datasets, revealing significant previously unknown test overlap. For example, in pre-training sets such as RedPajama-Data-1T and StarCoder-Data, we identified that 8-18\% of the HumanEval benchmark overlaps. Interestingly, we also find such contamination in synthetic dataset generated by GPT-3.5/4, suggesting a potential risk of unintentional contamination. We urge the community to adopt stronger decontamination approaches when using public benchmarks. Moreover, we call for the community to actively develop fresh one-time exams to evaluate models accurately. Our decontamination tool is publicly available at https://github.com/lm-sys/llm-decontaminator.

Yonghao Zhuang, Hexu Zhao, Lianmin Zheng et al. · berkeley

We study a novel and important communication pattern in large-scale model-parallel deep learning (DL), which we call cross-mesh resharding. This pattern emerges when the two paradigms of model parallelism - intra-operator and inter-operator parallelism - are combined to support large models on large clusters. In cross-mesh resharding, a sharded tensor needs to be sent from a source device mesh to a destination device mesh, on which the tensor may be distributed with the same or different layouts. We formalize this as a many-to-many multicast communication problem, and show that existing approaches either are sub-optimal or do not generalize to different network topologies or tensor layouts, which result from different model architectures and parallelism strategies. We then propose two contributions to address cross-mesh resharding: an efficient broadcast-based communication system, and an "overlapping-friendly" pipeline schedule. On microbenchmarks, our overall system outperforms existing ones by up to 10x across various tensor and mesh layouts. On end-to-end training of two large models, GPT-3 and U-Transformer, we improve throughput by 10% and 50%, respectively.

Ying Sheng, Shiyi Cao, Dacheng Li et al.

The "pretrain-then-finetune" paradigm is commonly adopted in the deployment of large language models. Low-Rank Adaptation (LoRA), a parameter-efficient fine-tuning method, is often employed to adapt a base model to a multitude of tasks, resulting in a substantial collection of LoRA adapters derived from one base model. We observe that this paradigm presents significant opportunities for batched inference during serving. To capitalize on these opportunities, we present S-LoRA, a system designed for the scalable serving of many LoRA adapters. S-LoRA stores all adapters in the main memory and fetches the adapters used by the currently running queries to the GPU memory. To efficiently use the GPU memory and reduce fragmentation, S-LoRA proposes Unified Paging. Unified Paging uses a unified memory pool to manage dynamic adapter weights with different ranks and KV cache tensors with varying sequence lengths. Additionally, S-LoRA employs a novel tensor parallelism strategy and highly optimized custom CUDA kernels for heterogeneous batching of LoRA computation. Collectively, these features enable S-LoRA to serve thousands of LoRA adapters on a single GPU or across multiple GPUs with a small overhead. Compared to state-of-the-art libraries such as HuggingFace PEFT and vLLM (with naive support of LoRA serving), S-LoRA can improve the throughput by up to 4 times and increase the number of served adapters by several orders of magnitude. As a result, S-LoRA enables scalable serving of many task-specific fine-tuned models and offers the potential for large-scale customized fine-tuning services. The code is available at https://github.com/S-LoRA/S-LoRA

Xiaoxuan Liu, Lanxiang Hu, Peter Bailis et al.

Speculative decoding is a pivotal technique to accelerate the inference of large language models (LLMs) by employing a smaller draft model to predict the target model's outputs. However, its efficacy can be limited due to the low predictive accuracy of the draft model, particularly when faced with diverse text inputs and a significant capability gap between the draft and target models. We introduce online speculative decoding to address this challenge. The main idea is to continuously update the (multiple) draft model(s) on observed user query data. Adapting to query distribution mitigates the shifts between the training distribution of the draft model and the query distribution, enabling the draft model to more accurately predict the target model's outputs. We develop a prototype of online speculative decoding based on knowledge distillation and evaluate it using both synthetic and real query data. The results show a substantial increase in the token acceptance rate by 0.1 to 0.65, bringing 1.42x to 2.17x latency reduction. Our code is available at https://github.com/LiuXiaoxuanPKU/OSD.

Yichao Fu, Siqi Zhu, Runlong Su et al.

In Large Language Model (LLM) inference, the output length of an LLM request is typically regarded as not known a priori. Consequently, most LLM serving systems employ a simple First-come-first-serve (FCFS) scheduling strategy, leading to Head-Of-Line (HOL) blocking and reduced throughput and service quality. In this paper, we reexamine this assumption -- we show that, although predicting the exact generation length of each request is infeasible, it is possible to predict the relative ranks of output lengths in a batch of requests, using learning to rank. The ranking information offers valuable guidance for scheduling requests. Building on this insight, we develop a novel scheduler for LLM inference and serving that can approximate the shortest-job-first (SJF) schedule better than existing approaches. We integrate this scheduler with the state-of-the-art LLM serving system and show significant performance improvement in several important applications: 2.8x lower latency in chatbot serving and 6.5x higher throughput in synthetic data generation. Our code is available at https://github.com/hao-ai-lab/vllm-ltr.git

Gur-Eyal Sela, Ionel Gog, Justin Wong et al. · berkeley

Efficient vision works maximize accuracy under a latency budget. These works evaluate accuracy offline, one image at a time. However, real-time vision applications like autonomous driving operate in streaming settings, where ground truth changes between inference start and finish. This results in a significant accuracy drop. Therefore, a recent work proposed to maximize accuracy in streaming settings on average. In this paper, we propose to maximize streaming accuracy for every environment context. We posit that scenario difficulty influences the initial (offline) accuracy difference, while obstacle displacement in the scene affects the subsequent accuracy degradation. Our method, Octopus, uses these scenario properties to select configurations that maximize streaming accuracy at test time. Our method improves tracking performance (S-MOTA) by 7.4% over the conventional static approach. Further, performance improvement using our method comes in addition to, and not instead of, advances in offline accuracy.

Naman Jain, Tianjun Zhang, Wei-Lin Chiang et al. · berkeley, microsoft-research

Natural language to code generation is an important application area of LLMs and has received wide attention from the community. The majority of relevant studies have exclusively concentrated on increasing the quantity and functional correctness of training sets while disregarding other stylistic elements of programs. More recently, data quality has garnered a lot of interest and multiple works have showcased its importance for improving performance. In this work, we investigate data quality for code and find that making the code more structured and readable leads to improved code generation performance of the system. We build a novel data-cleaning pipeline that uses these principles to transform existing programs by 1.) renaming variables, 2.) modularizing and decomposing complex code into smaller helper sub-functions, and 3.) inserting natural-language based plans via LLM based transformations. We evaluate our approach on two challenging algorithmic code generation benchmarks and find that fine-tuning CodeLLaMa-7B on our transformed modularized programs improves the performance by up to 30% compared to fine-tuning on the original dataset. Additionally, we demonstrate improved performance from using a smaller amount of higher-quality data, finding that a model fine-tuned on the entire original dataset is outperformed by a model trained on 15% of our cleaned dataset. Even in comparison to closed-source models, our models outperform the much larger AlphaCoder models.

Shuo Yang, Ying Sheng, Joseph E. Gonzalez et al.

The inference process for large language models is slow and memory-intensive, with one of the most critical bottlenecks being excessive Key-Value (KV) cache accesses. This paper introduces "Double Sparsity," a novel post-training sparse attention technique designed to alleviate this bottleneck by reducing KV cache access. Double Sparsity combines token sparsity, which focuses on utilizing only the important tokens for computing self-attention, with channel sparsity, an approach that uses important feature channels for identifying important tokens. Our key insight is that the pattern of channel sparsity is relatively static, allowing us to use offline calibration to make it efficient at runtime, thereby enabling accurate and efficient identification of important tokens. Moreover, this method can be combined with offloading to achieve significant memory usage reduction. Experimental results demonstrate that Double Sparsity can achieve $\frac{1}{16}$ token and channel sparsity with minimal impact on accuracy across various tasks, including wiki-2 perplexity, key-value retrieval, and long context benchmarks with models including Llama-2-7B, Llama-2-70B, and Mixtral-8x7B. It brings up to a 14.1$\times$ acceleration in attention operations and a 1.9$\times$ improvement in end-to-end inference on GPUs. With offloading, it achieves a decoding speed acceleration of 16.3$\times$ compared to state-of-the-art solutions at a sequence length of 256K. Our code is publicly available at https://github.com/andy-yang-1/DoubleSparse.

Yangzhen Wu, Aaron J. Li, Wenjie Ma et al.

The rapid progress of frontier large language models has led to widespread benchmark saturation, limiting the ability of existing datasets to differentiate model capabilities or provide useful training signal. For instance, on LiveCodeBench, frontier models achieve over 99% Pass@1 on easy splits and exceed 90% Pass@1 on average across difficulty levels. Constructing new, challenging datasets typically requires substantial human effort, creating a bottleneck for progress. We introduce BenchEvolver, a solution-centric evolutionary framework that automatically transforms existing coding problems into harder variants. Rather than generating problems from scratch, BenchEvolver evolves reference solutions through structured transformations and derives corresponding statements and tests from the evolved solutions. This design grounds generation in executable semantics, enabling scalable construction of high-quality, diverse, and difficult tasks with verifiable correctness. Applying BenchEvolver to LiveCodeBench and SciCode, we obtain evolved tasks that are substantially harder while maintaining validity, reference correctness, and diversity. We further curate LiveCodeBench-Plus, a 91-problem benchmark combining evolved and difficult original LCB-v6 tasks, where frontier-model Pass@1 ranges from 27.5% to 62.6%, restoring clear discrimination among strong coding models. Importantly, evolved tasks remain challenging even for the model that generates them, enabling self-improvement. We further show that RL on evolved LCB tasks improves held-out coding performance: for gpt-oss-20b, seed+evolved training achieves +8.7 and +8.3 Pass@1 gains on LCB v6 Hard and LCB-Pro Easy, exceeding seed-only gains by 70.7% and 34.8%, respectively. Our results show that BenchEvolver can convert saturated benchmarks into frontier-level evaluation suites and reusable training signal.

Tian Xia, Hanchen Li, Zhifei Li et al.

Modern LLM serving systems increasingly host agentic workloads, whose sessions issue tens of model invocations interleaved with tool calls, accumulating KV cache that can be reused across steps. As requests' total KV cache size easily exceeds GPU HBM capacity, researchers offload them to CPU DRAM. However, tool-call durations span orders of magnitude, and the cost of transferring KV cache between tiers makes it impractical to re-place entries on every call. We observe that agentic programs exhibit a two-phase structure: busy phases of rapid short tool calls and idle phases dominated by long-running calls. Current eviction policies such as LRU fail to capture this property. A binary busy/idle label also falls short because the ratio of busy to idle programs may not match the hardware's GPU-to-CPU capacity ratio. When it does not, one tier sits underutilized while the other is oversubscribed, wasting memory or forcing unnecessary evictions. We present MORI, an agent serving system that solves the above problem. Our key insight is that idleness is a continuous, relative spectrum. MORI ranks all active programs by idleness, assigns the busiest to GPU HBM and the most idle to CPU DRAM, dynamically shifts the partition boundary to match hardware capacity, and enforces admission control at each memory tier. Evaluated on real coding agent workloads collected from Claude Code across four GPU and model pairs, MORI delivers 20--71% higher throughput and 18--43% lower TTFT than the best baseline with offloading.

Hanchen Li, Qiuyang Mang, Runyuan He et al.

Agentic LLM applications interleave LLM generation requests with tool calls. These tool calls break the continuity of the workflow by creating pauses between LLM requests, bringing many challenges for the serving system, especially under multi-turn scenarios. Each pause potentially causes KV cache eviction and extra waiting time before entering the continuous batch for the following LLM request. Since these pauses happen for each call, this problem becomes increasingly severe as turn number grow for agentic programs. Previous works either fail to incorporate information from the tool call, evicting KV cache that leads to repetitive prefill or loading, or ignore the continuity of a multi-turn program, creating waiting time between turns that increases per-request latency. We present Continuum, a serving system to optimize job completion time for multi-turn agent workloads by combining tool-aware KV cache timeout with program-level scheduling. By predicting tool call durations in agentic workflows, Continuum selectively pins the KV cache in GPU memory with a time-to-live value based on total turn number. When combined with program-level first-come-first-serve, Continuum prevents scheduling bubbles, preserves multi-turn continuity, and optimizes for throughput for complex agentic workflows. By modeling the variability of tool call and agent program continuity, Continuum outperforms state-of-the-art baselines. Our evaluation on real-world agentic workloads (SWE-Bench and BFCL) with Llama-3.1 8B/70B models shows that Continuum significantly improves the average job completion times, and remains performant across different hardware setups and DRAM offloading schemes. Preview code is available at: https://github.com/Hanchenli/vllm-continuum

Jared Quincy Davis, Boris Hanin, Lingjiao Chen et al.

As practitioners seek to surpass the current reliability and quality frontier of monolithic models, Compound AI Systems consisting of many language model inference calls are increasingly employed. In this work, we construct systems, which we call Networks of Networks (NoNs) organized around the distinction between generating a proposed answer and verifying its correctness, a fundamental concept in complexity theory that we show empirically extends to Language Models (LMs). We introduce a verifier-based judge NoN with K generators, an instantiation of "best-of-K" or "judge-based" compound AI systems. Through experiments on synthetic tasks such as prime factorization, and core benchmarks such as the MMLU, we demonstrate notable performance gains. For instance, in factoring products of two 3-digit primes, a simple NoN improves accuracy from 3.7\% to 36.6\%. On MMLU, a verifier-based judge construction with only 3 generators boosts accuracy over individual GPT-4-Turbo calls by 2.8\%. Our analysis reveals that these gains are most pronounced in domains where verification is notably easier than generation--a characterization which we believe subsumes many reasoning and procedural knowledge tasks, but doesn't often hold for factual and declarative knowledge-based settings. For mathematical and formal logic reasoning-based subjects of MMLU, we observe a 5-8\% or higher gain, whilst no gain on others such as geography and religion. We provide key takeaways for ML practitioners, including the importance of considering verification complexity, the impact of witness format on verifiability, and a simple test to determine the potential benefit of this NoN approach for a given problem distribution. This work aims to inform future research and practice in the design of compound AI systems.

Lakshya A Agrawal, Donghyun Lee, Shangyin Tan et al.

Can a single LLM-based optimization system match specialized tools across fundamentally different domains? We show that when optimization problems are formulated as improving a text artifact evaluated by a scoring function, a single AI-based optimization system-supporting single-task search, multi-task search with cross-problem transfer, and generalization to unseen inputs-achieves state-of-the-art results across six diverse tasks. Our system discovers agent architectures that nearly triple Gemini Flash's ARC-AGI accuracy (32.5% to 89.5%), finds scheduling algorithms that cut cloud costs by 40%, generates CUDA kernels where 87% match or beat PyTorch, and outperforms AlphaEvolve's reported circle packing solution (n=26). Ablations across three domains reveal that actionable side information yields faster convergence and substantially higher final scores than score-only feedback, and that multi-task search outperforms independent optimization given equivalent per-problem budget through cross-task transfer, with benefits scaling with the number of related tasks. Together, we show for the first time that text optimization with LLM-based search is a general-purpose problem-solving paradigm, unifying tasks traditionally requiring domain-specific algorithms under a single framework. We open-source optimize\_anything with support for multiple backends as part of the GEPA project at https://github.com/gepa-ai/gepa .

Jintao Zhang, Kaiwen Zheng, Kai Jiang et al.

We introduce TurboDiffusion, a video generation acceleration framework that can speed up end-to-end diffusion generation by 100-200x while maintaining video quality. TurboDiffusion mainly relies on several components for acceleration: (1) Attention acceleration: TurboDiffusion uses low-bit SageAttention and trainable Sparse-Linear Attention (SLA) to speed up attention computation. (2) Step distillation: TurboDiffusion adopts rCM for efficient step distillation. (3) W8A8 quantization: TurboDiffusion quantizes model parameters and activations to 8 bits to accelerate linear layers and compress the model. In addition, TurboDiffusion incorporates several other engineering optimizations. We conduct experiments on the Wan2.2-I2V-14B-720P, Wan2.1-T2V-1.3B-480P, Wan2.1-T2V-14B-720P, and Wan2.1-T2V-14B-480P models. Experimental results show that TurboDiffusion achieves 100-200x speedup for video generation even on a single RTX 5090 GPU, while maintaining comparable video quality. The GitHub repository, which includes model checkpoints and easy-to-use code, is available at https://github.com/thu-ml/TurboDiffusion.

Daniel Kang, Tatsunori Hashimoto, Ion Stoica et al.

As ML models have increased in capabilities and accuracy, so has the complexity of their deployments. Increasingly, ML model consumers are turning to service providers to serve the ML models in the ML-as-a-service (MLaaS) paradigm. As MLaaS proliferates, a critical requirement emerges: how can model consumers verify that the correct predictions were served, in the face of malicious, lazy, or buggy service providers? In this work, we present the first practical ImageNet-scale method to verify ML model inference non-interactively, i.e., after the inference has been done. To do so, we leverage recent developments in ZK-SNARKs (zero-knowledge succinct non-interactive argument of knowledge), a form of zero-knowledge proofs. ZK-SNARKs allows us to verify ML model execution non-interactively and with only standard cryptographic hardness assumptions. In particular, we provide the first ZK-SNARK proof of valid inference for a full resolution ImageNet model, achieving 79\% top-5 accuracy. We further use these ZK-SNARKs to design protocols to verify ML model execution in a variety of scenarios, including for verifying MLaaS predictions, verifying MLaaS model accuracy, and using ML models for trustless retrieval. Together, our results show that ZK-SNARKs have the promise to make verified ML model inference practical.

Shuo Yang, Haocheng Xi, Yilong Zhao et al. · tsinghua

$k$-means has historically been positioned primarily as an offline processing primitive, typically used for dataset organization or embedding preprocessing rather than as a first-class component in online systems. In this work, we revisit this classical algorithm under the lens of modern AI system design and enable $k$-means as an online primitive. We point out that existing GPU implementations of $k$-means remain fundamentally bottlenecked by low-level system constraints rather than theoretical algorithmic complexity. Specifically, the assignment stage suffers from a severe IO bottleneck due to the massive explicit materialization of the $N \times K$ distance matrix in High Bandwidth Memory (HBM). Simultaneously, the centroid update stage is heavily penalized by hardware-level atomic write contention caused by irregular, scatter-style token aggregations. To bridge this performance gap, we propose flash-kmeans, an IO-aware and contention-free $k$-means implementation for modern GPU workloads. Flash-kmeans introduces two core kernel-level innovations: (1) FlashAssign, which fuses distance computation with an online argmin to completely bypass intermediate memory materialization; (2) sort-inverse update, which explicitly constructs an inverse mapping to transform high-contention atomic scatters into high-bandwidth, segment-level localized reductions. Furthermore, we integrate algorithm-system co-designs, including chunked-stream overlap and cache-aware compile heuristics, to ensure practical deployability. Extensive evaluations on NVIDIA H200 GPUs demonstrate that flash-kmeans achieves up to 17.9$\times$ end-to-end speedup over best baselines, while outperforming industry-standard libraries like cuML and FAISS by 33$\times$ and over 200$\times$, respectively.

Mert Cemri, Shubham Agrawal, Akshat Gupta et al.

The paradigm of automated program generation is shifting from one-shot generation to inference-time search, where Large Language Models (LLMs) function as semantic mutation operators within evolutionary loops. While effective, these systems are currently governed by static schedules that fail to account for the non-stationary dynamics of the search process. This rigidity results in substantial computational waste, as resources are indiscriminately allocated to stagnating populations while promising frontiers remain under-exploited. We introduce AdaEvolve, a framework that reformulates LLM-driven evolution as a hierarchical adaptive optimization problem. AdaEvolve uses an "accumulated improvement signal" to unify decisions across three levels: Local Adaptation, which dynamically modulates the exploration intensity within a population of solution candidates; Global Adaptation, which routes the global resource budget via bandit-based scheduling across different solution candidate populations; and Meta-Guidance which generates novel solution tactics based on the previously generated solutions and their corresponding improvements when the progress stalls. We demonstrate that AdaEvolve consistently outperforms the open-sourced baselines across 185 different open-ended optimization problems including combinatorial, systems optimization and algorithm design problems.

Deevashwer Rathee, Dacheng Li, Ion Stoica et al.

Many inference services based on large language models (LLMs) pose a privacy concern, either revealing user prompts to the service or the proprietary weights to the user. Secure inference offers a solution to this problem through secure multi-party computation (MPC), however, it is still impractical for modern LLM workload due to the large overhead imposed by MPC. To address this overhead, we propose Marill, a framework that adapts LLM fine-tuning to minimize MPC usage during secure inference. Marill introduces high-level architectural changes during fine-tuning that significantly reduce the number of expensive operations needed within MPC during inference, by removing some and relocating others outside MPC without compromising security. As a result, Marill-generated models are more efficient across all secure inference protocols and our approach complements MPC-friendly approximations for such operations. Compared to standard fine-tuning, Marill results in 3.6-11.3x better runtime and 2.4-6.9x better communication during secure inference across various MPC settings, while typically preserving over 90% performance across downstream tasks.

Qiuyang Mang, Wenhao Chai, Zhifei Li et al.

We introduce FrontierCS, a benchmark of 156 open-ended problems across diverse areas of computer science, designed and reviewed by experts, including CS PhDs and top-tier competitive programming participants and problem setters. Unlike existing benchmarks that focus on tasks with known optimal solutions, FrontierCS targets problems where the optimal solution is unknown, but the quality of a solution can be objectively evaluated. Models solve these tasks by implementing executable programs rather than outputting a direct answer. FrontierCS includes algorithmic problems, which are often NP-hard variants of competitive programming problems with objective partial scoring, and research problems with the same property. For each problem we provide an expert reference solution and an automatic evaluator. Combining open-ended design, measurable progress, and expert curation, FrontierCS provides a benchmark at the frontier of computer-science difficulty. Empirically, we find that frontier reasoning models still lag far behind human experts on both the algorithmic and research tracks, that increasing reasoning budgets alone does not close this gap, and that models often over-optimize for generating merely workable code instead of discovering high-quality algorithms and system designs.

Jintao Zhang, Haoxu Wang, Kai Jiang et al. · tsinghua

Sparse-Linear Attention (SLA) combines sparse and linear attention to accelerate diffusion models and has shown strong performance in video generation. However, (i) SLA relies on a heuristic split that assigns computations to the sparse or linear branch based on attention-weight magnitude, which can be suboptimal. Additionally, (ii) after formally analyzing the attention error in SLA, we identify a mismatch between SLA and a direct decomposition into sparse and linear attention. We propose SLA2, which introduces (I) a learnable router that dynamically selects whether each attention computation should use sparse or linear attention, (II) a more faithful and direct sparse-linear attention formulation that uses a learnable ratio to combine the sparse and linear attention branches, and (III) a sparse + low-bit attention design, where low-bit attention is introduced via quantization-aware fine-tuning to reduce quantization error. Experiments show that on video diffusion models, SLA2 can achieve 97% attention sparsity and deliver an 18.6x attention speedup while preserving generation quality.

Charles Packer, Sarah Wooders, Kevin Lin et al.

Large language models (LLMs) have revolutionized AI, but are constrained by limited context windows, hindering their utility in tasks like extended conversations and document analysis. To enable using context beyond limited context windows, we propose virtual context management, a technique drawing inspiration from hierarchical memory systems in traditional operating systems that provide the appearance of large memory resources through data movement between fast and slow memory. Using this technique, we introduce MemGPT (Memory-GPT), a system that intelligently manages different memory tiers in order to effectively provide extended context within the LLM's limited context window, and utilizes interrupts to manage control flow between itself and the user. We evaluate our OS-inspired design in two domains where the limited context windows of modern LLMs severely handicaps their performance: document analysis, where MemGPT is able to analyze large documents that far exceed the underlying LLM's context window, and multi-session chat, where MemGPT can create conversational agents that remember, reflect, and evolve dynamically through long-term interactions with their users. We release MemGPT code and data for our experiments at https://memgpt.ai.

Shu Liu, Shubham Agarwal, Monishwaran Maheswaran et al.

Recent work such as AlphaEvolve has shown that combining LLM-driven optimization with evolutionary search can effectively improve programs, prompts, and algorithms across domains. In this paradigm, previously evaluated solutions are reused to guide the model toward new candidate solutions. Crucially, the effectiveness of this evolution process depends on the search strategy: how prior solutions are selected and varied to generate new candidates. However, most existing methods rely on fixed search strategies with predefined knobs (e.g., explore-exploit ratios) that remain static throughout execution. While effective in some settings, these approaches often fail to adapt across tasks, or even within the same task as the search space changes over time. We introduce EvoX, an adaptive evolution method that optimizes its own evolution process. EvoX jointly evolves candidate solutions and the search strategies used to generate them, continuously updating how prior solutions are selected and varied based on progress. This enables the system to dynamically shift between different search strategies during the optimization process. Across nearly 200 real-world optimization tasks, EvoX outperforms existing AI-driven evolutionary methods including AlphaEvolve, OpenEvolve, GEPA, and ShinkaEvolve on the majority of tasks.

Melih Elibol, Vinamra Benara, Samyu Yagati et al.

Scientists increasingly rely on Python tools to perform scalable distributed memory array operations using rich, NumPy-like expressions. However, many of these tools rely on dynamic schedulers optimized for abstract task graphs, which often encounter memory and network bandwidth-related bottlenecks due to sub-optimal data and operator placement decisions. Tools built on the message passing interface (MPI), such as ScaLAPACK and SLATE, have better scaling properties, but these solutions require specialized knowledge to use. In this work, we present NumS, an array programming library which optimizes NumPy-like expressions on task-based distributed systems. This is achieved through a novel scheduler called Load Simulated Hierarchical Scheduling (LSHS). LSHS is a local search method which optimizes operator placement by minimizing maximum memory and network load on any given node within a distributed system. Coupled with a heuristic for load balanced data layouts, our approach is capable of attaining communication lower bounds on some common numerical operations, and our empirical study shows that LSHS enhances performance on Ray by decreasing network load by a factor of 2x, requiring 4x less memory, and reducing execution time by 10x on the logistic regression problem. On terabyte-scale data, NumS achieves competitive performance to SLATE on DGEMM, up to 20x speedup over Dask on a key operation for tensor factorization, and a 2x speedup on logistic regression compared to Dask ML and Spark's MLlib.

Xirui Li, Ming Li, Derry Xu et al.

Constructing environments for training and evaluating claw-like agents remains a manual, human-intensive process that does not scale. We argue that what is needed is not just a dataset, but an automated pipeline capable of generating diverse, verified environments on demand. To this end, we introduce ClawEnvKit, an autonomous generation pipeline that instantiates this formalism from natural language descriptions. The pipeline comprises three modules: (1) a parser that extracts structured generation parameters from natural language input; (2) a generator that produces the task specification, tool interface, and scoring configuration; and (3) a validator that enforces feasibility, diversity, structural validity, and internal consistency across the generated environments. Using ClawEnvKit, we construct Auto-ClawEval, the first large-scale benchmark for claw-like agents, comprising 1,040 environments across 24 categories. Empirically, Auto-ClawEval matches or exceeds human-curated environments on coherence and clarity at 13,800x lower cost. Evaluated across 4 model families and 8 agent harness frameworks, we find that harness engineering boosts performance by up to 15.7 percentage points over a bare ReAct baseline, completion remains the primary axis of variation with no model saturating the benchmark, and automated generation enables evaluation at a scale previously infeasible. Beyond static benchmarking, ClawEnvKit enables live evaluation: users describe a desired capability in natural language and obtain a verified environment on demand, turning evaluation into a continuous, user-driven process. The same mechanism serves as an on-demand training environment generator, producing task distributions that adapt to an agent's current weaknesses rather than being bounded by existing user logs.

Lianmin Zheng, Liangsheng Yin, Zhiqiang Xie et al.

Large language models (LLMs) are increasingly used for complex tasks that require multiple generation calls, advanced prompting techniques, control flow, and structured inputs/outputs. However, efficient systems are lacking for programming and executing these applications. We introduce SGLang, a system for efficient execution of complex language model programs. SGLang consists of a frontend language and a runtime. The frontend simplifies programming with primitives for generation and parallelism control. The runtime accelerates execution with novel optimizations like RadixAttention for KV cache reuse and compressed finite state machines for faster structured output decoding. Experiments show that SGLang achieves up to 6.4x higher throughput compared to state-of-the-art inference systems on various large language and multi-modal models on tasks including agent control, logical reasoning, few-shot learning benchmarks, JSON decoding, retrieval-augmented generation pipelines, and multi-turn chat. The code is publicly available at https://github.com/sgl-project/sglang

Tianjun Zhang, Shishir G. Patil, Naman Jain et al. · berkeley, microsoft-research

Pretraining Large Language Models (LLMs) on large corpora of textual data is now a standard paradigm. When using these LLMs for many downstream applications, it is common to additionally bake in new knowledge (e.g., time-critical news, or private domain knowledge) into the pretrained model either through RAG-based-prompting, or fine-tuning. However, the optimal methodology for the model to gain such new knowledge remains an open question. In this paper, we present Retrieval Augmented FineTuning (RAFT), a training recipe that improves the model's ability to answer questions in a "open-book" in-domain settings. In RAFT, given a question, and a set of retrieved documents, we train the model to ignore those documents that don't help in answering the question, which we call, distractor documents. RAFT accomplishes this by citing verbatim the right sequence from the relevant document that would help answer the question. This coupled with RAFT's chain-of-thought-style response helps improve the model's ability to reason. In domain-specific RAG, RAFT consistently improves the model's performance across PubMed, HotpotQA, and Gorilla datasets, presenting a post-training recipe to improve pre-trained LLMs to in-domain RAG. RAFT's code and demo are open-sourced at github.com/ShishirPatil/gorilla.

Yichao Fu, Peter Bailis, Ion Stoica et al.

Autoregressive decoding of large language models (LLMs) is memory bandwidth bounded, resulting in high latency and significant wastes of the parallel processing power of modern accelerators. Existing methods for accelerating LLM decoding often require a draft model (e.g., speculative decoding), which is nontrivial to obtain and unable to generalize. In this paper, we introduce Lookahead decoding, an exact, parallel decoding algorithm that accelerates LLM decoding without needing auxiliary models or data stores. It allows trading per-step log(FLOPs) to reduce the number of total decoding steps, is more parallelizable on single or multiple modern accelerators, and is compatible with concurrent memory-efficient attention (e.g., FlashAttention). Our implementation of Lookahead decoding can speed up autoregressive decoding by up to 1.8x on MT-bench and 4x with strong scaling on multiple GPUs in code completion tasks. Our code is avialable at https://github.com/hao-ai-lab/LookaheadDecoding

Shubham Agarwal, Alexander Krentsel, Shu Liu et al.

AI agents increasingly excel at generating, testing, and refining code. However, they fall short on tasks requiring formal guarantees of full coverage that testing alone cannot provide. Distributed systems are a prime example: properties such as consistency between reads and writes must hold under every possible interleaving of events. Mechanized formal verification can guarantee such correctness, but typically demands months to years of expert effort. As evidence, even SOTA coding agents (Codex with GPT-5.4 and Claude Code with Opus 4.6) succeed on only 2/7 distributed key-value-store specifications. In this paper, we present the first effective approach to addressing this gap, Inductive Deductive Synthesis (IDS), which jointly and incrementally synthesizes implementation and proof, and learns from failed attempts to systematically try promising strategies. Built as an agentic LLM system, IDS achieves 7/7 in about 6.8 hours and $106 per spec on average, roughly 200x faster than expert effort and 17% cheaper than SOTA agents. IDS further incorporates performance feedback into the same loop, yielding implementations up to 3x faster than published verified systems.

Shu Liu, Alexander Krentsel, Shubham Agarwal et al.

Core systems like key-value stores have historically taken years to build, and are designed to be general so as to amortize cost across deployments, paying a significant performance cost. We argue that LLM-based coding agents now make a different approach tractable: Just-in-Time Systems, in which the entire system is synthesized from scratch, specialized to the environment, workload, and required system properties. We present a JIT system synthesis pipeline, Jitskit, and explore its effectiveness in synthesizing key-value stores from spec cards that span different YCSB workloads, deployment constraints (e.g., compute resources), and system properties (e.g., consistency and durability). Jitskit iteratively refines a system implementation to match the specification against an evolving evaluation test suite. The resulting synthesized systems are performant, beating comparable state-of-the-art systems on 18 of 18 specs tried, by up to 4.6x over the best off-the-shelf baseline on the most favorable spec. Naively running Claude Code either reward-hacks or underperforms Jitskit by up to 5.4x. We discuss the challenges we overcame in building Jitskit and our key takeaways.

Sahil Joshi, Prithvi Dixit, Agniva Chowdhury et al.

Sparsity has long been a central theme in LLM efficiency, but its role in context processing remains unresolved. As LLM workloads shift toward longer contexts and agentic interactions, the compute and memory bottlenecks of attention become increasingly critical, raising the question of whether these constraints are fundamental. Our position is that these constraints are artificial and unnecessary, and that the future of LLM inference lies in extreme but principled sparsity along the context dimension. This position is supported by several strands of empirical and theoretical evidence. First, we find the insistence on dense attention unreasonable, since in a long context a query effectively projects O(N) attention information into a hidden space of dimension d << N, making the process inherently lossy. Second, we perform an extensive study of sparsity in LLMs spanning 20 models across five model families, varying context lengths, and different sparsity levels. We empirically demonstrate a strong trend: current LLMs, despite not being trained for context sparsity, are remarkably robust to inference-time decode sparsity across tasks of varying complexity, including retrieval, multi-hop QA, mathematical reasoning, and agentic coding. Importantly, we also show that current hardware is already sufficient to realize substantial gains from this sparsity. For example, our sparse decode kernels accelerate large-context processing by up to 10x over FlashInfer at 50x sparsity levels on hardware such as the H100. Overall, these results position extreme context sparsity not as a heuristic, but as a principled foundation for LLM inference, training, and architecture design: one that is both feasible and beneficial, and a compelling direction for future systems.

Sijun Tan, Siyuan Zhuang, Kyle Montgomery et al.

LLM-based judges have emerged as a scalable alternative to human evaluation and are increasingly used to assess, compare, and improve models. However, the reliability of LLM-based judges themselves is rarely scrutinized. As LLMs become more advanced, their responses grow more sophisticated, requiring stronger judges to evaluate them. Existing benchmarks primarily focus on a judge's alignment with human preferences, but often fail to account for more challenging tasks where crowdsourced human preference is a poor indicator of factual and logical correctness. To address this, we propose a novel evaluation framework to objectively evaluate LLM-based judges. Based on this framework, we propose JudgeBench, a benchmark for evaluating LLM-based judges on challenging response pairs spanning knowledge, reasoning, math, and coding. JudgeBench leverages a novel pipeline for converting existing difficult datasets into challenging response pairs with preference labels reflecting objective correctness. Our comprehensive evaluation on a collection of prompted judges, fine-tuned judges, multi-agent judges, and reward models shows that JudgeBench poses a significantly greater challenge than previous benchmarks, with many strong models (e.g., GPT-4o) performing just slightly better than random guessing. Overall, JudgeBench offers a reliable platform for assessing increasingly advanced LLM-based judges. Data and code are available at https://github.com/ScalerLab/JudgeBench.

Gur-Eyal Sela, Kumar Krishna Agrawal, Bharathan Balaji et al.

Live video analytics (LVA) runs continuously across massive camera fleets, but inference cost with modern vision models remains high. To address this, dynamic model size selection (DMSS) is an attractive approach: it is content-aware but treats models as black boxes, and could potentially reduce cost by up to 10x without model retraining or modification. Without ground truth labels at runtime, we observe that DMSS methods use two stages per segment: (i) sampling a few models to calculate prediction statistics (e.g., confidences), then (ii) selection of the model size from those statistics. Prior systems fail to generalize to diverse workloads, particularly to mobile videos and lower accuracy targets. We identify that the failure modes stem from inefficient sampling whose cost exceeds its benefit, and inaccurate per-segment accuracy prediction. In this work, we present RedunCut, a new DMSS system that addresses both: It uses a measurement-driven planner that estimates the cost-benefit tradeoff of sampling, and a lightweight, data-driven performance model to improve accuracy prediction. Across road-vehicle, drone, and surveillance videos and multiple model families and tasks, RedunCut reduces compute cost by 14-62% at fixed accuracy and remains robust to limited historical data and to drift.

Haocheng Xi, Shuo Yang, Yilong Zhao et al.

Despite rapid progress in autoregressive video diffusion, an emerging system algorithm bottleneck limits both deployability and generation capability: KV cache memory. In autoregressive video generation models, the KV cache grows with generation history and quickly dominates GPU memory, often exceeding 30 GB, preventing deployment on widely available hardware. More critically, constrained KV cache budgets restrict the effective working memory, directly degrading long horizon consistency in identity, layout, and motion. To address this challenge, we present Quant VideoGen (QVG), a training free KV cache quantization framework for autoregressive video diffusion models. QVG leverages video spatiotemporal redundancy through Semantic Aware Smoothing, producing low magnitude, quantization friendly residuals. It further introduces Progressive Residual Quantization, a coarse to fine multi stage scheme that reduces quantization error while enabling a smooth quality memory trade off. Across LongCat Video, HY WorldPlay, and Self Forcing benchmarks, QVG establishes a new Pareto frontier between quality and memory efficiency, reducing KV cache memory by up to 7.0 times with less than 4% end to end latency overhead while consistently outperforming existing baselines in generation quality.

Wentao Guo, Mayank Mishra, Xinle Cheng et al.

Mixture of Experts (MoE) models have emerged as the de facto architecture for scaling up language models without significantly increasing the computational cost. Recent MoE models demonstrate a clear trend towards high expert granularity (smaller expert intermediate dimension) and higher sparsity (constant number of activated experts with higher number of total experts), which improve model quality per FLOP. However, fine-grained MoEs suffer from increased activation memory footprint and reduced hardware efficiency due to higher IO costs, while sparser MoEs suffer from wasted computations due to padding in Grouped GEMM kernels. In response, we propose a memory-efficient algorithm to compute the forward and backward passes of MoEs with minimal activation caching for the backward pass. We also design GPU kernels that overlap memory IO with computation benefiting all MoE architectures. Finally, we propose a novel "token rounding" method that minimizes the wasted compute due to padding in Grouped GEMM kernels. As a result, our method SonicMoE reduces activation memory by 45% and achieves a 1.86x compute throughput improvement on Hopper GPUs compared to ScatterMoE's BF16 MoE kernel for a fine-grained 7B MoE. Concretely, SonicMoE on 64 H100s achieves a training throughput of 213 billion tokens per day comparable to ScatterMoE's 225 billion tokens per day on 96 H100s for a 7B MoE model training with FSDP-2 using the lm-engine codebase. Under high MoE sparsity settings, our tile-aware token rounding algorithm yields an additional 1.16x speedup on kernel execution time compared to vanilla top-$K$ routing while maintaining similar downstream performance. We open-source all our kernels to enable faster MoE model training.

Dacheng Li, Shiyi Cao, Tyler Griggs et al.

Large reasoning models (LRMs) tackle complex reasoning problems by following long chain-of-thoughts (Long CoT) that incorporate reflection, backtracking, and self-validation. However, the training techniques and data requirements to elicit Long CoT remain poorly understood. In this work, we find that a Large Language model (LLM) can effectively learn Long CoT reasoning through data-efficient supervised fine-tuning (SFT) and parameter-efficient low-rank adaptation (LoRA). With just 17k long CoT training samples, the Qwen2.5-32B-Instruct model achieves significant improvements on a wide range of math and coding benchmarks, including 56.7% (+40.0%) on AIME 2024 and 57.0% (+8.1%) on LiveCodeBench, competitive to the proprietary o1-preview model's score of 44.6% and 59.1%. More importantly, we find that the structure of Long CoT is critical to the learning process, whereas the content of individual reasoning steps has minimal impact. Perturbations affecting content, such as training on incorrect samples or removing reasoning keywords, have little impact on performance. In contrast, structural modifications that disrupt logical consistency in the Long CoT, such as shuffling or deleting reasoning steps, significantly degrade accuracy. For example, a model trained on Long CoT samples with incorrect answers still achieves only 3.2% lower accuracy compared to training with fully correct samples. These insights deepen our understanding of how to elicit reasoning capabilities in LLMs and highlight key considerations for efficiently training the next generation of reasoning models. This is the academic paper of our previous released Sky-T1-32B-Preview model. Codes are available at https://github.com/NovaSky-AI/SkyThought.

Alejandro Cuadron, Dacheng Li, Wenjie Ma et al.

Large Reasoning Models (LRMs) represent a breakthrough in AI problem-solving capabilities, but their effectiveness in interactive environments can be limited. This paper introduces and analyzes overthinking in LRMs. A phenomenon where models favor extended internal reasoning chains over environmental interaction. Through experiments on software engineering tasks using SWE Bench Verified, we observe three recurring patterns: Analysis Paralysis, Rogue Actions, and Premature Disengagement. We propose a framework to study these behaviors, which correlates with human expert assessments, and analyze 4018 trajectories. We observe that higher overthinking scores correlate with decreased performance, with reasoning models exhibiting stronger tendencies toward overthinking compared to non-reasoning models. Our analysis reveals that simple efforts to mitigate overthinking in agentic environments, such as selecting the solution with the lower overthinking score, can improve model performance by almost 30% while reducing computational costs by 43%. These results suggest that mitigating overthinking has strong practical implications. We suggest that by leveraging native function-calling capabilities and selective reinforcement learning overthinking tendencies could be mitigated. We also open-source our evaluation framework and dataset to facilitate research in this direction at https://github.com/AlexCuadron/Overthinking.

Ying Sheng, Shiyi Cao, Dacheng Li et al.

High-demand LLM inference services (e.g., ChatGPT and BARD) support a wide range of requests from short chat conversations to long document reading. To ensure that all client requests are processed fairly, most major LLM inference services have request rate limits, to ensure that no client can dominate the request queue. However, this rudimentary notion of fairness also results in under-utilization of the resources and poor client experience when there is spare capacity. While there is a rich literature on fair scheduling, serving LLMs presents new challenges due to their unpredictable request lengths and their unique batching characteristics on parallel accelerators. This paper introduces the definition of LLM serving fairness based on a cost function that accounts for the number of input and output tokens processed. To achieve fairness in serving, we propose a novel scheduling algorithm, the Virtual Token Counter (VTC), a fair scheduler based on the continuous batching mechanism. We prove a 2x tight upper bound on the service difference between two backlogged clients, adhering to the requirement of work-conserving. Through extensive experiments, we demonstrate the superior performance of VTC in ensuring fairness, especially in contrast to other baseline methods, which exhibit shortcomings under various conditions. The reproducible code is available at https://github.com/Ying1123/VTC-artifact

Ziming Mao, Yihan Zhang, Chihan Cui et al.

Mixture-of-Experts (MoE) workloads rely on expert parallelism (EP) to achieve high GPU efficiency. State-of-the-art EP communication systems such as DeepEP demonstrate strong performance but exhibit poor portability across heterogeneous GPU and NIC platforms. The poor portability is rooted in architecture: GPU-initiated token-level RDMA communication requires tight vertical integration between GPUs and NICs, e.g., GPU writes to NIC driver/MMIO interfaces. We present UCCL-EP, a portable EP communication system that delivers DeepEP-level performance across heterogeneous GPU and NIC hardware. UCCL-EP replaces GPU-initiated RDMA with a high-throughput GPU-CPU control channel: compact token-routing commands are transferred to multithreaded CPU proxies, which then issue GPUDirect RDMA operations on behalf of GPUs. UCCL-EP further emulates various ordering semantics required by specialized EP communication modes using RDMA immediate data, enabling correctness on NICs that lack such ordering, e.g., AWS EFA. We implement UCCL-EP on NVIDIA and AMD GPUs with EFA and Broadcom NICs. On EFA, it outperforms the best existing EP solution by up to $2.1\times$ for dispatch and combine throughput. On NVIDIA-only platform, UCCL-EP achieves comparable performance to the original DeepEP. UCCL-EP also improves token throughput on SGLang by up to 40% on the NVIDIA+EFA platform, and improves DeepSeek-V3 training throughput over the AMD Primus/Megatron-LM framework by up to 45% on a 16-node AMD+Broadcom platform.

Haocheng Xi, Shuo Yang, Yilong Zhao et al. · tsinghua

Diffusion Transformers (DiTs) dominate video generation but their high computational cost severely limits real-world applicability, usually requiring tens of minutes to generate a few seconds of video even on high-performance GPUs. This inefficiency primarily arises from the quadratic computational complexity of 3D Full Attention with respect to the context length. In this paper, we propose a training-free framework termed Sparse VideoGen (SVG) that leverages the inherent sparsity in 3D Full Attention to boost inference efficiency. We reveal that the attention heads can be dynamically classified into two groups depending on distinct sparse patterns: (1) Spatial Head, where only spatially-related tokens within each frame dominate the attention output, and (2) Temporal Head, where only temporally-related tokens across different frames dominate. Based on this insight, SVG proposes an online profiling strategy to capture the dynamic sparse patterns and predicts the type of attention head. Combined with a novel hardware-efficient tensor layout transformation and customized kernel implementations, SVG achieves up to 2.28x and 2.33x end-to-end speedup on CogVideoX-v1.5 and HunyuanVideo, respectively, while preserving generation quality. Our code is open-sourced and is available at https://github.com/svg-project/Sparse-VideoGen

Yilong Zhao, Jiaming Tang, Kan Zhu et al.

Reasoning language models have demonstrated remarkable capabilities on challenging tasks by generating elaborate chain-of-thought (CoT) solutions. However, such lengthy generation shifts the inference bottleneck from compute-bound to memory-bound. To generate each token, the model applies full attention to all previously generated tokens, requiring memory access to an increasingly large KV-Cache. Consequently, longer generations demand more memory access for every step, leading to substantial pressure on memory bandwidth. To address this, we introduce SparseSpec, a speculative decoding framework that reuses the same model as the draft and target models (i.e., self-speculation). SparseSpec features a novel sparse attention mechanism, PillarAttn, as the draft model, which accurately selects critical tokens via elegantly reusing information from the verification stage. Furthermore, SparseSpec co-designs self-speculation with three system innovations: (1) a unified scheduler to batch token drafting and verification, (2) delayed verification for CPU/GPU overlap, and (3) dynamic KV-Cache management to maximize memory utilization. Across various models and datasets, SparseSpec outperforms state-of-the-art solutions, with an up to 2.13x throughput speedup.

Dacheng Li, Shiyi Cao, Chengkun Cao et al.

Increasing test-time compute for LLMs shows promise across domains but remains underexplored in code generation, despite extensive study in math. In this paper, we propose S*, the first hybrid test-time scaling framework that substantially improves the coverage and selection accuracy of generated code. S* extends the existing parallel scaling paradigm with sequential scaling to push performance boundaries. It further leverages a novel selection mechanism that adaptively generates distinguishing inputs for pairwise comparison, combined with execution-grounded information to robustly identify correct solutions. We evaluate across 12 Large Language Models and Large Reasoning Model and show: (1) S* consistently improves performance across model families and sizes, enabling a 3B model to outperform GPT-4o-mini; (2) S* enables non-reasoning models to surpass reasoning models - GPT-4o-mini with S* outperforms o1-preview by 3.7% on LiveCodeBench; (3) S* further boosts state-of-the-art reasoning models - DeepSeek-R1-Distill-Qwen-32B with S* achieves 85.7% on LiveCodeBench, approaching o1 (high) at 88.5%. Code will be available under https://github.com/NovaSky-AI/SkyThought.

Evan Frick, Tianle Li, Connor Chen et al. · berkeley

We introduce a new benchmark for reward models that quantifies their ability to produce strong language models through RLHF (Reinforcement Learning from Human Feedback). The gold-standard approach is to run a full RLHF training pipeline and directly probe downstream LLM performance. However, this process is prohibitively expensive. To address this, we build a predictive model of downstream LLM performance by evaluating the reward model on proxy tasks. These proxy tasks consist of a large-scale human preference and a verifiable correctness preference dataset, in which we measure 12 metrics across 12 domains. To investigate which reward model metrics are most correlated to gold-standard RLHF outcomes, we launch an end-to-end RLHF experiment on a large-scale crowdsourced human preference platform to view real reward model downstream performance as ground truth. Ultimately, we compile our data and findings into Preference Proxy Evaluations (PPE), the first reward model benchmark explicitly linked to post-RLHF real-world human preference performance, which we open-source for public use and further development. Our code and evaluations can be found at https://github.com/lmarena/PPE .