Frontier language-model agents excel at complex tasks, but high inference costs make transferring their capabilities to smaller models essential. Traditional knowledge distillation achieves this through parameter updates, requiring access to the student's weights and additional training compute while encoding the transferred knowledge implicitly in the parameters. We introduce D-Skill, a framework that treats the teacher as a black box and distills its behavioral advantage into an interpretable natural-language skill by analyzing paired rollouts and identifying behavioral divergences and agreements. The resulting lessons are consolidated offline, with any compatible skill-update backend, into a skill document that is provided to the student at inference time. D-Skill constructs the skill entirely offline from fixed trajectories, without labels, further environment interaction, or weight updates after rollout collection. Across SpreadsheetBench-Verified, ALFWorld, and chess puzzles spanning multiple teacher-student pairs, D-Skill closes 75%-95% of the teacher-student gap without touching model weights. After a single offline pass over the paired traces, D-Skill outperforms one-epoch SFT baselines and remains competitive with multi-epoch SFT. Under the same model family, the resulting student also incurs 5.7x-8.3x lower inference cost per task than the teacher. Our results demonstrate that natural-language skill distillation offers a performant, interpretable, and efficient alternative to weight-based capability transfer.
Huanzhi Mao, Yichuan Wang, Michael Koo, Wenyuan Huang, Huy Doan, Jinghan Xia, Ziyao Wang, Joseph E. Gonzalez
Vision-language model (VLM) agents that act in the open world need two abilities: world knowledge to interpret what they see, and exploration to find what they cannot yet see. We study how agents combine the two through interactive visual geolocation. GeoGuesser Bench contains 282 tasks in an offline Street View environment where agents rotate, zoom, and move between real panoramas. Its three task types shift the balance from knowledge to exploration: agents (1) report the coordinates of their starting location, (2) name the enclosing street, city, state, or country, or (3) find, read, and count nearby objects. The coordinate tasks in (1) also include 42 real GeoGuessr World Championship rounds, whose expert guesses provide a human reference. We evaluate 23 VLMs: the best accuracy is 62.5% on place naming and 35.0% on object finding, and the best agent averages 4352.4 GeoGuessr points on the championship rounds, against 4770.3 for the better expert in each round. Beyond these headline numbers, world knowledge and exploration are only weakly correlated across models: agents that localize well do not necessarily explore well. To understand where agents fail, we categorize their errors into vision, navigation, and tool-use failures. Further ablations show that the strongest agents, such as Claude Opus 4.8, still perform well with limited views, while weaker agents degrade; customizing the environment thus adjusts task difficulty. The environment's verifiable rewards and restorable state also support training, and our Planner-Worker-Compressor harness cuts planner tokens by 57-61% by compacting the interaction history. We hope GeoGuesser Bench and its environment serve as a high-quality foundation for simulating real-world exploration at scale.
Huanzhi Mao, Chengkun Cao, Shuo Yuan, Joseph E. Gonzalez
Large language models are increasingly capable of writing and executing code, suggesting a natural paradigm for tool-augmented agents: expressing tool use as programs rather than isolated calls. Programmatic tool calling (PTC) captures this idea through loops, batching, abstraction, and structured computation, but existing PTC approaches are largely stateless, limiting their ability to reuse intermediate results and adapt over time. This is problematic for real-world tasks, where tool environments and problem-solving processes are inherently stateful. We therefore present I-PTC, an interactive framework that executes stateful PTC: models issue code snippets that call tools, maintain reusable state, and evolve the tool interface through API corrections. To evaluate the effectiveness of I-PTC and the general PTC setting, we introduce PTC-Bench, a benchmark of long-tail, production-style user requests whose difficulty comes from large initial environments, many dependent API interactions, extended planning, and recovery across follow-ups and transient failures. Experiments on PTC-Bench show that I-PTC improves performance on complex multi-step tasks with large tool sets while substantially reducing token consumption over conventional tool-calling baselines and the original PTC baseline, highlighting stateful programmatic tool calling as a promising direction for scalable tool-augmented agents.
Shiyi Cao, Ziming Mao, Dacheng Li, Joseph E. Gonzalez, Ion Stoica
The rapid progress of software engineering agents drives interest in applying them to open-ended problems such as GPU kernel optimization. However, we find that existing coding agents converge to local optima on real-world, expert-written kernels (e.g., FlashInfer). We introduce Evolving Agent Teams, a multi-agent framework that escapes such local optima. It features (1) a central supervisor that diagnoses each local optimum and restructures a team of agents; and (2) a persistent file system that comprises a learning document, an evolution history, and per-team artifacts, grounding restructuring with accumulated experience. Under a matched token budget, Evolving Agent Teams consistently outperforms strong coding-agent baselines. With an extended token budget, Evolving Agent Teams generates a CUDA kernel that reaches 1.2x mean speedup over the production FlashInfer MLA paged decode kernel from scratch and wins on 36 of 47 production workloads from FlashInfer-Bench (0.86-0.99x on the remaining 11). It also demonstrates superior performance on tasks such as KDA prefill on AMD MI350X (3.01x over vLLM) and NVIDIA B200 (3.44x over FlashKDA), MoE expert-offload serving on Apple silicon (30.3 tokens/s decode for Qwen3.8-Flash-Next in MLX 4-bit on a 48 GB M4 Pro), and RoboDojo robot manipulation (62.1 on held-out seeds vs. 55.3 for the best leaderboard entry). Code is available at https://github.com/caoshiyi/EAT.
K-Search: LLM Kernel Generation via Co-Evolving Intrinsic World Model
Shiyi Cao, Ziming Mao, Joseph E. Gonzalez, Ion Stoica
AI 速览：用协同进化的世界模型替代静态启发式搜索，让LLM显式规划GPU内核的多步结构优化。
Optimizing GPU kernels is critical for efficient modern machine learning systems yet remains challenging due to the complex interplay of design factors and rapid hardware evolution. Existing automated approaches typically treat Large Language Models (LLMs) merely as stochastic code generators within heuristic-guided evolutionary loops. These methods often struggle with complex kernels requiring coordinated, multi-step structural transformations, as they lack explicit planning capabilities and frequently discard promising strategies due to inefficient or incorrect intermediate implementations. To address this, we propose Search via Co-Evolving World Model and build K-Search based on this method. By replacing static search heuristics with a co-evolving world model, our framework leverages LLMs' prior domain knowledge to guide the search, actively exploring the optimization space. This approach explicitly decouples high-level algorithmic planning from low-level program instantiation, enabling the system to navigate non-monotonic optimization paths while remaining resilient to temporary implementation defects. We evaluate K-Search on diverse, complex kernels from FlashInfer, including GQA, MLA, and MoE kernels. Our results show that K-Search significantly outperforms state-of-the-art evolutionary search methods, achieving an average 2.10x improvement and up to a 14.3x gain on complex MoE kernels. On the GPUMode TriMul task, K-Search achieves state-of-the-art performance on H100, reaching 1030us and surpassing both prior evolution and human-designed solutions.
Supporting Our AI Overlords: Redesigning Data Systems to be Agent-First
Shu Liu, Soujanya Ponnapalli, Shreya Shankar, Sepanta Zeighami, Alan Zhu, Shubham Agarwal, Ruiqi Chen, Samion Suwito 等 15 位作者
AI 速览：作者主张未来数据系统应以LLM智能体为主要工作负载，并基于智能体推测过程的特征提出智能体优先的数据系统架构研究方向。
Large Language Model (LLM) agents, acting on their users' behalf to manipulate and analyze data, are likely to become the dominant workload for data systems in the future. When working with data, agents employ a high-throughput process of exploration and solution formulation for the given task, one we call agentic speculation. The sheer volume and inefficiencies of agentic speculation can pose challenges for present-day data systems. We argue that data systems need to adapt to more natively support agentic workloads. We take advantage of the characteristics of agentic speculation that we identify, i.e., scale, heterogeneity, redundancy, and steerability - to outline a number of new research opportunities for a new agent-first data systems architecture, ranging from new query interfaces, to new query processing techniques, to new agentic memory stores.
Shu Liu, Asim Biswal, Amog Kamsetty, Audrey Cheng, Luis Gaspar Schroeder, Liana Patel, Shiyi Cao, Xiangxi Mo 等 11 位作者
Batch data analytics is a growing application for Large Language Models (LLMs). LLMs enable users to perform a wide range of natural language tasks, such as classification, entity extraction, and translation, over large datasets. However, LLM inference is highly costly and slow: for example, an NVIDIA L4 GPU running Llama3-8B can only process 6 KB of text per second, taking about a day to handle 15 GB of data; processing a similar amount of data costs around $10K on OpenAI's GPT-4o. In this paper, we propose novel techniques that can significantly reduce the cost of LLM calls for relational data analytics workloads. Our key contribution is developing efficient algorithms for reordering the rows and the fields within each row of an input table to maximize key-value (KV) cache reuse when performing LLM serving. As such, our approach can be easily applied to existing analytics systems and serving platforms. Our evaluation shows that our solution can yield up to 3.4x improvement in job completion time on a benchmark of diverse LLM-based queries using Llama 3 models. Our solution also achieves a 32% cost savings under OpenAI and Anthropic pricing models.
Dacheng Li, Shiyi Cao, Chengkun Cao, Xiuyu Li, Shangyin Tan, Kurt Keutzer, Jiarong Xing, Joseph E. Gonzalez 等 9 位作者
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.
Charles Packer, Sarah Wooders, Kevin Lin, Vivian Fang, Shishir G. Patil, Ion Stoica, Joseph E. Gonzalez
AI 速览：受操作系统分层内存启发，提出虚拟上下文管理，让LLM在有限上下文窗口内处理超长文档与多轮长期对话。
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.
Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li 等 13 位作者
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.