Why Do Multi-Agent LLM Systems Fail?

Despite growing enthusiasm for Multi-Agent Systems (MAS), where multiple LLM agents collaborate to accomplish tasks, their performance gains across popular benchmarks remain minimal compared to single-agent frameworks. This gap highlights the need to analyze the challenges hindering MAS effectiveness. In this paper, we present the first comprehensive study of MAS challenges. We analyze five popular MAS frameworks across over 150 tasks, involving six expert human annotators. We identify 14 unique failure modes and propose a comprehensive taxonomy applicable to various MAS frameworks. This taxonomy emerges iteratively from agreements among three expert annotators per study, achieving a Cohen's Kappa score of 0.88. These fine-grained failure modes are organized into 3 categories, (i) specification and system design failures, (ii) inter-agent misalignment, and (iii) task verification and termination. To support scalable evaluation, we integrate MASFT with LLM-as-a-Judge. We also explore if identified failures could be easily prevented by proposing two interventions: improved specification of agent roles and enhanced orchestration strategies. Our findings reveal that identified failures require more complex solutions, highlighting a clear roadmap for future research. We open-source our dataset and LLM annotator.

WorldModelBench: Judging Video Generation Models As World Models

Video generation models have rapidly progressed, positioning themselves as video world models capable of supporting decision-making applications like robotics and autonomous driving. However, current benchmarks fail to rigorously evaluate these claims, focusing only on general video quality, ignoring important factors to world models such as physics adherence. To bridge this gap, we propose WorldModelBench, a benchmark designed to evaluate the world modeling capabilities of video generation models in application-driven domains. WorldModelBench offers two key advantages: (1) Against to nuanced world modeling violations: By incorporating instruction-following and physics-adherence dimensions, WorldModelBench detects subtle violations, such as irregular changes in object size that breach the mass conservation law - issues overlooked by prior benchmarks. (2) Aligned with large-scale human preferences: We crowd-source 67K human labels to accurately measure 14 frontier models. Using our high-quality human labels, we further fine-tune an accurate judger to automate the evaluation procedure, achieving 8.6% higher average accuracy in predicting world modeling violations than GPT-4o with 2B parameters. In addition, we demonstrate that training to align human annotations by maximizing the rewards from the judger noticeably improve the world modeling capability. The website is available at https://worldmodelbench-team.github.io.

S*: Test Time Scaling for Code Generation

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.

Autellix: An Efficient Serving Engine for LLM Agents as General Programs

Large language model (LLM) applications are evolving beyond simple chatbots into dynamic, general-purpose agentic programs, which scale LLM calls and output tokens to help AI agents reason, explore, and solve complex tasks. However, existing LLM serving systems ignore dependencies between programs and calls, missing significant opportunities for optimization. Our analysis reveals that programs submitted to LLM serving engines experience long cumulative wait times, primarily due to head-of-line blocking at both the individual LLM request and the program. To address this, we introduce Autellix, an LLM serving system that treats programs as first-class citizens to minimize their end-to-end latencies. Autellix intercepts LLM calls submitted by programs, enriching schedulers with program-level context. We propose two scheduling algorithms-for single-threaded and distributed programs-that preempt and prioritize LLM calls based on their programs' previously completed calls. Our evaluation demonstrates that across diverse LLMs and agentic workloads, Autellix improves throughput of programs by 4-15x at the same latency compared to state-of-the-art systems, such as vLLM.

The Danger of Overthinking: Examining the Reasoning-Action Dilemma in Agentic Tasks

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.

LLMs Can Easily Learn to Reason from Demonstrations Structure, not content, is what matters!

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.

Adaptive Semantic Prompt Caching with VectorQ

Semantic prompt caches reduce the latency and cost of large language model (LLM) inference by reusing cached LLM-generated responses for semantically similar prompts. Vector similarity metrics assign a numerical score to quantify the similarity between an embedded prompt and its nearest neighbor in the cache. Existing systems rely on a static threshold to classify whether the similarity score is sufficiently high to result in a cache hit. We show that this one-size-fits-all threshold is insufficient across different prompts. We propose VectorQ, a framework to learn embedding-specific threshold regions that adapt to the complexity and uncertainty of an embedding. Through evaluations on a combination of four diverse datasets, we show that VectorQ consistently outperforms state-of-the-art systems across all static thresholds, achieving up to 12x increases in cache hit rate and error rate reductions up to 92%.

BARE: Combining Base and Instruction-Tuned Language Models for Better Synthetic Data Generation

As the demand for high-quality data in model training grows, researchers and developers are increasingly generating synthetic data to tune and train LLMs. A common assumption about synthetic data is that sampling from instruct-tuned models is sufficient; however, these models struggle to produce diverse outputs-a key requirement for generalization. Despite various prompting methods, in this work we show that achieving meaningful diversity from instruct-tuned models remains challenging. In contrast, we find base models without post-training exhibit greater diversity, but are less capable at instruction following and hence of lower quality. Leveraging this insight, we propose Base-Refine (BARE), a synthetic data generation method that combines the diversity of base models with the quality of instruct-tuned models through a two-stage process. With minimal few-shot examples and curation, BARE generates diverse and high-quality datasets, improving downstream task performance. We show that fine-tuning with as few as 1,000 BARE-generated samples can reach performance comparable to the best similarly sized models on LiveCodeBench tasks. Furthermore, fine-tuning with BARE-generated data achieves a 101% improvement over instruct-only data on GSM8K and a 18.4% improvement over SOTA methods on RAFT.

HashAttention: Semantic Sparsity for Faster Inference

Utilizing longer contexts is increasingly essential to power better AI systems. However, the cost of attending to long contexts is high due to the involved softmax computation. While the scaled dot-product attention (SDPA) exhibits token sparsity, with only a few pivotal tokens significantly contributing to attention, leveraging this sparsity effectively remains an open challenge. Previous methods either suffer from model degradation or require considerable additional resources. We propose HashAttention --a principled approach casting pivotal token identification as a recommendation problem. Given a query, HashAttention encodes keys and queries in Hamming space capturing the required semantic similarity using learned mapping functions. HashAttention efficiently identifies pivotal tokens for a given query in this Hamming space using bitwise operations, and only these pivotal tokens are used for attention computation, significantly improving overall attention efficiency. HashAttention can reduce the number of tokens used by a factor of $1/32\times$ for the Llama-3.1-8B model with LongBench, keeping average quality loss within 0.6 points, while using only 32 bits per token auxiliary memory. At $32\times$ sparsity, HashAttention is $3{-}6\times$ faster than LightLLM and $2.5{-}4.5\times$ faster than gpt-fast on Nvidia-L4 GPU.

VisionArena: 230K Real World User-VLM Conversations with Preference Labels

With the growing adoption and capabilities of vision-language models (VLMs) comes the need for benchmarks that capture authentic user-VLM interactions. In response, we create VisionArena, a dataset of 230K real-world conversations between users and VLMs. Collected from Chatbot Arena - an open-source platform where users interact with VLMs and submit preference votes - VisionArena spans 73K unique users, 45 VLMs, and 138 languages. Our dataset contains three subsets: VisionArena-Chat, 200k single and multi-turn conversations between a user and a VLM; VisionArena-Battle, 30K conversations comparing two anonymous VLMs with user preference votes; and VisionArena-Bench, an automatic benchmark of 500 diverse user prompts that efficiently approximate the live Chatbot Arena model rankings. Additionally, we highlight the types of question asked by users, the influence of response style on preference, and areas where models often fail. We find open-ended tasks like captioning and humor are highly style-dependent, and current VLMs struggle with spatial reasoning and planning tasks. Lastly, we show finetuning the same base model on VisionArena-Chat outperforms Llava-Instruct-158K, with a 17-point gain on MMMU and a 46-point gain on the WildVision benchmark. Dataset at https://huggingface.co/lmarena-ai