A comprehensive index of artificial intelligence and machine-learning research with AI-generated summaries, citation metrics, and direct links to papers and code.
Baohao Liao, Hanze Dong, Christof Monz, et al.
ReOPD enables efficient multi-turn on-policy distillation for LLM agents by reusing teacher trajectories as replayed prefixes, avoiding costly environment interactions.
Xiao Yu, Baolin Peng, Ruize Xu, et al.
OpenForgeRL is an open-source framework that enables end-to-end training of harness-based AI agents in any environment by decoupling training and inference via a lightweight proxy and Kubernetes orchestrator.
Linjun Li
LLMs can appear safer under direct dangerous objectives than when mediated by other agents, revealing a compositional safety gap.
Zhongyuan Peng, Dan Huang, Chuyu Zhang, et al.
ICAE-Bench evaluates coding agents on interactive project-building from fuzzy requirements using automated user simulation and multi-dimensional diagnostics.
Jiazhen Jiang, Boxi Cao, Lingyong Yan, et al.
DocOps introduces a verifiable benchmark for evaluating autonomous agents on complex document operations, revealing critical failure modes in state tracking and editing.
Unknown
This paper introduces world models, enabling agents to learn and plan in latent spaces, leading to the Dreamer series for long-horizon reinforcement learning.
Unknown
Proposes Agent Process Reward Models (AgentPRM) for fine-grained step-level supervision as a tractable alternative to large-scale RL for LLM agents.
Unknown
Proposes AgentPRM, a process reward model that evaluates each decision step of LLM agents via step-wise promise and progress metrics.
Unknown
Evaluates LLM tool-use agents on long-horizon planning in large-scale tool ecosystems, finding most models below two-thirds accuracy.
Unknown
PLAN-AND-ACT improves long-horizon planning for agents by integrating plan generation with execution, achieving 57.58% success on WebArena-Lite.
Frank F. Xu, Yufan Song, Boxuan Li, et al.
Introduces TheAgentCompany, an extensible benchmark for evaluating LLM agents on real-world professional tasks.
Unknown
MultiagentBench is a benchmark for evaluating LLM agents in multi-agent collaboration and competition scenarios.