Tag
The paper introduces Communicative On-Policy Distillation (COPD), which uses natural-language communication to distill black-box teacher capabilities into smaller student models for long-horizon tasks, achieving 2.46x the gain of the strongest black-box OPD baseline on CAMG-II with generalization to Terminal-Bench 2.1 and SWE-Bench Pro V2.
OfficePrism introduces a benchmark of 420 single-domain enterprise tasks across seven business functions, evaluating language agents with executable reference witnesses; across 124 model-harness-package configurations, pass rates range from 22.6% to 76.4%, with performance decreasing from L1 to L3 task complexity.
Skill∞ 介绍了一种递归自我改进框架,让元智能体通过可复用的元技能同时优化任务技能与优化流程本身,在 83 个 SkillsBench 任务上达到 60.7% 的通过率,优于人工整理技能和 SkillEvolver。
该论文提出 MICA 与 DART 框架,研究在可执行能力相同的情况下,MCP 与 CLI 接口表示差异如何显著改变语言智能体的行为,并通过针对性修复接口实现提升性能、降低交互成本。
论文提出 GRSE(Group-Relative Skill Evolution),通过重复执行并对比同一技能库下的多次运行,将共享团队奖励正确归因到具体的 agent-技能对上,从而改进协作式语言智能体的技能演化,在 HotpotQA 与 LoCoMo 上超越 MASkills 并能从有害技能中恢复。
该论文研究了工具使用型语言智能体在监督微调下的「选择性撤回」行为,提出一种能发现聚合准确率所掩盖的规则级回归的诊断方法,并表明微调本身并不足以保证智能体安全。
PEARL提出了一种面向多轮智能体的强化学习方法,通过分支轨迹的首次分歧点构建局部决策特征,并利用成功/失败记忆聚类识别重复模式,从而在不引入额外critic的情况下改进信用分配。在AppWorld和ALFWorld基准上,PEARL在4B和8B规模模型上均超越了ARPO等基线,ALFWorld 4B达到97.9%成功率。
CoPE 论文提出通过在成功轨迹上训练多步动作目标,使语言智能体的行动与预测能力协同演化,从而在不增加推理开销的情况下提升样本效率,相比 GRPO 与 ECHO 在多项基准上取得更强表现。
该论文提出 agentic SElf-distilLation with environmental Feedback modeling (SELF) 框架,将环境反馈建模与后见自蒸馏联合优化,以在缺乏可验证奖励的交互环境中提升语言智能体的训练效果,在 τ-bench 和 AppWorld 基准上显著优于 SDPO 和 GRPO。
This paper shows that language agents fail to reliably connect past actions with their outcomes when using interaction history, and proposes a simple outcome-labeling annotation plus a learned calibrator that selectively records experience, improving task success and reducing action repetition.
The paper audits the shape of memory use in long-horizon language agents, finding heavy-tailed (core-tail) retrieval traces, and proposes Core-Tail World Model (CTWM), a rank-based memory controller that cuts prompt tokens by up to 24.48% on LongMemEval while preserving accuracy.
NAQD-Env is a synthetic benchmark that evaluates how language agents selectively suspend actions when evidence, permissions, or stop instructions change, finding near-zero withdrawal recall across models despite fine-tuning gains in decision accuracy.
RideWay introduces an efficiency-centered benchmark and the Efficiency Utility metric for evaluating tool-using language agents in ridehailing tasks, measuring success-gapped performance based on tool calls and user turns.
This paper introduces RD-Forget, a training-free framework for persistent language agents that separates stored memory from query-conditioned evidence to handle changing facts while preserving historical information.
This study investigates how procedural memory mismatch in language agents affects web tasks, finding that under controlled conditions, mismatch does not lead to behavioral disruption or errors.
ContractEval is a diagnostic framework that evaluates procedural conformance in LLM agents by making active obligations explicit and matching them against evidence, detecting failures missed by traditional judges.
This paper introduces Outcome Monitors, a deterministic detector that identifies silent failures in tool calls for language agents and provides recovery receipts, significantly improving task completion rates in benchmarks like ToolMaze and τ-bench.
AQuA is a research system with two independent language-model-driven agents that recursively self-improve in quantitative trading research, achieving strong information coefficients on crypto and US equities while using sealed sandboxes to prevent data leakage.
This paper introduces DARC, a diagnosis-guided recovery harness that makes agent self-correction selective by profiling failure modes and pruning mismatched interventions before test-time correction, improving performance on ALFWorld, AppWorld, and XBRL Finance.
Presents EvoGraph-Mem, a failure-aware editable graph memory framework for long-term language agents that tracks positive/negative evidence and activation states for insights, enabling memory maintenance through utility-aware retrieval and graph-level editing.