Yuanli Wang

Xiangyi Li, Wenbo Chen, Yimin Liu et al. · berkeley

Agent Skills are structured packages of procedural knowledge that augment LLM agents at inference time. Despite rapid adoption, there is no standard way to measure whether they actually help. We present SkillsBench, a benchmark of 86 tasks across 11 domains paired with curated Skills and deterministic verifiers. Each task is evaluated under three conditions: no Skills, curated Skills, and self-generated Skills. We test 7 agent-model configurations over 7,308 trajectories. Curated Skills raise average pass rate by 16.2 percentage points(pp), but effects vary widely by domain (+4.5pp for Software Engineering to +51.9pp for Healthcare) and 16 of 84 tasks show negative deltas. Self-generated Skills provide no benefit on average, showing that models cannot reliably author the procedural knowledge they benefit from consuming. Focused Skills with 2--3 modules outperform comprehensive documentation, and smaller models with Skills can match larger models without them.

Yuanli Wang, Yaoyao Qian, Yue Zhang et al.

LLM agents have made rapid progress on software engineering and ML research tasks, but these advances often assume access to a working runnable environment. For research artifacts released alongside published papers, setting up such an environment from a fresh machine remains a major bottleneck. Existing environment setup benchmarks do not cover the full scope of research artifact deployment, which involves multi-language toolchains, system-level dependencies beyond containers (e.g. GPU/CUDA and kernel configurations), and legacy artifact compatibility. We introduce DeployBench, a multi-domain benchmark of 51 research-artifact deployment tasks spanning AI/ML, computer systems, and scientific computing, covering all these dimensions. Each task is verified by a hidden pipeline that executes the paper's designated experiment and checks its outputs. Evaluating four state-of-the-art LLMs with OpenHands yields pass-rates from 7.8% - 51.0% . Failures are dominated by a completion-judgment problem: 97 of 154 are agent-terminated self-stops, where the agent's pre-finish checks validate a different or weaker target than the paper-specific task requires. DeployBench highlights the gap between current agents and autonomous deployment, and offers a realistic testbed for scientific research agents.

Xiangyi Li, Kyoung Whan Choe, Yimin Liu et al. · apple-ml

Large language model (LLM) agents are increasingly deployed to automate productivity tasks (e.g., email, scheduling, document management), but evaluating them on live services is risky due to potentially irreversible changes. Existing benchmarks rely on simplified environments and fail to capture realistic, stateful, multi-service workflows. We introduce ClawsBench, a benchmark for evaluating and improving LLM agents in realistic productivity settings. It includes five high-fidelity mock services (Gmail, Slack, Google Calendar, Google Docs, Google Drive) with full state management and deterministic snapshot/restore, along with 44 structured tasks covering single-service, cross-service, and safety-critical scenarios. We decompose agent scaffolding into two independent levers (domain skills that inject API knowledge via progressive disclosure, and a meta prompt that coordinates behavior across services) and vary both to measure their separate and combined effects. Experiments across 6 models, 4 agent harnesses, and 33 conditions show that with full scaffolding, agents achieve task success rates of 39-64% but exhibit unsafe action rates of 7-33%. On OpenClaw, the top five models fall within a 10 percentage-point band on task success (53-63%), with unsafe action rates from 7% to 23% and no consistent ordering between the two metrics. We identify eight recurring patterns of unsafe behavior, including multi-step sandbox escalation and silent contract modification.

Long Phan, Alice Gatti, Ziwen Han et al. · amazon-science, apple-ml

Benchmarks are important tools for tracking the rapid advancements in large language model (LLM) capabilities. However, benchmarks are not keeping pace in difficulty: LLMs now achieve over 90\% accuracy on popular benchmarks like MMLU, limiting informed measurement of state-of-the-art LLM capabilities. In response, we introduce Humanity's Last Exam (HLE), a multi-modal benchmark at the frontier of human knowledge, designed to be the final closed-ended academic benchmark of its kind with broad subject coverage. HLE consists of 2,500 questions across dozens of subjects, including mathematics, humanities, and the natural sciences. HLE is developed globally by subject-matter experts and consists of multiple-choice and short-answer questions suitable for automated grading. Each question has a known solution that is unambiguous and easily verifiable, but cannot be quickly answered via internet retrieval. State-of-the-art LLMs demonstrate low accuracy and calibration on HLE, highlighting a significant gap between current LLM capabilities and the expert human frontier on closed-ended academic questions. To inform research and policymaking upon a clear understanding of model capabilities, we publicly release HLE at https://lastexam.ai.

Jiahang Li, Taoyu Chen, Yuanli Wang

This study introduces a novel approach for analyzing and modifying entity relationships in GPT models, diverging from ROME's entity-focused methods. We develop a relation tracing technique to understand the influence of language model computations on relationship judgments. Using the FewRel dataset, we identify key roles of MLP modules and attention mechanisms in processing relationship information. Our method, tested against ROME on a new dataset, shows improved balance in specificity and generalization, underscoring the potential of manipulating early-layer modules for enhanced model understanding and accuracy.

Yaoyao Qian, Yuanli Wang, Jinda Zhang et al.

Current evaluation of web agents largely reduces to binary success metrics or conformity to a single reference trajectory, ignoring the structural diversity present in benchmark datasets. We present WebGraphEval, a framework that abstracts trajectories from multiple agents into a unified, weighted action graph. This representation is directly compatible with benchmarks such as WebArena, leveraging leaderboard runs and newly collected trajectories without modifying environments. The framework canonically encodes actions, merges recurring behaviors, and applies structural analyses including reward propagation and success-weighted edge statistics. Evaluations across thousands of trajectories from six web agents show that the graph abstraction captures cross-model regularities, highlights redundancy and inefficiency, and identifies critical decision points overlooked by outcome-based metrics. By framing web interaction as graph-structured data, WebGraphEval establishes a general methodology for multi-path, cross-agent, and efficiency-aware evaluation of web agents.

Yaoyao Qian, Jindan Huang, Yuanli Wang et al.

Dialogue systems often fail when user utterances are semantically complete yet lack the clarity and completeness required for appropriate system action. This mismatch arises because users frequently do not fully understand their own needs, while systems require precise intent definitions. This highlights the critical Intent-Action Alignment Problem: determining when an expression is not just understood, but truly ready for a system to act upon. We present STORM, a framework modeling asymmetric information dynamics through conversations between UserLLM (full internal access) and AgentLLM (observable behavior only). STORM produces annotated corpora capturing trajectories of expression phrasing and latent cognitive transitions, enabling systematic analysis of how collaborative understanding develops. Our contributions include: (1) formalizing asymmetric information processing in dialogue systems; (2) modeling intent formation tracking collaborative understanding evolution; and (3) evaluation metrics measuring internal cognitive improvements alongside task performance. Experiments across four language models reveal that moderate uncertainty (40-60%) can outperform complete transparency in certain scenarios, with model-specific patterns suggesting reconsideration of optimal information completeness in human-AI collaboration. These findings contribute to understanding asymmetric reasoning dynamics and inform uncertainty-calibrated dialogue system design.

Yuanli Wang, Lei Huang

Federated Learning(FL) is a privacy-preserving machine learning paradigm where a global model is trained in-situ across a large number of distributed edge devices. These systems are often comprised of millions of user devices and only a subset of available devices can be used for training in each epoch. Designing a device selection strategy is challenging, given that devices are highly heterogeneous in both their system resources and training data. This heterogeneity makes device selection very crucial for timely model convergence and sufficient model accuracy. To tackle the FL client heterogeneity problem, various client selection algorithms have been developed, showing promising performance improvement in terms of model coverage and accuracy. In this work, we study the overhead of client selection algorithms in a large scale FL environment. Then we propose an efficient data distribution summary calculation algorithm to reduce the overhead in a real-world large scale FL environment. The evaluation shows that our proposed solution could achieve up to 30x reduction in data summary time, and up to 360x reduction in clustering time.