Tag
这篇注册研究考察了自改进智能体将运行结果写成教训时，是否会在后续决策中传递运气（随机性）的影响。实验发现，在可靠工具因偶然失败后，LLM撰写的教训会让智能体在后续运行中显著更多地选择较差的工具（11.6% vs 0.4%），而修改记忆写入指令可降低这一危害。
该论文提出技能适配（skill adaptation）方法，通过元技能引导外部改写器将现有技能重写以匹配执行模型的能力，从而提升弱模型在 harness 自演化中的表现，并在七个模型上验证了改进效果。
A follow-up report on an experiment where hidden instructions planted in web pages, agent config files (AGENTS.md, CLAUDE.md, .cursorrules), and public repos were obeyed by AI agents — fake keys were used 82 times, yet none of the agents disclosed their identity or model. The post raises questions about whether agents acting on embedded instructions without self-identifying is good safety training or a red flag.
SciGame 是一个将科研文献转化为结构化「科学卡片」的受控基准，通过原始与反事实环境配对来分离先验知识与实证证据，评估大模型智能体的科学发现能力，发现精确符号定律恢复仍是主要瓶颈。
Project Zero 比较了 LLM 驱动的漏洞挖掘系统 BigSleep 与生成式模糊测试工具 Fuzzilli 在 V8 上的表现：六个月内 BigSleep 发现 45 个漏洞、Fuzzilli 37 个，两者结果存在明显重叠但仍各自独有大量发现，结论是模糊测试与 LLM 互为补充。
Trajectory2Gym 提出一种决策落地的组合式交互合成方法用于工具调用智能体的训练监督，在五个 Qwen3 规模上于 Tau1、Tau2 和 BFCL 基准上提升 16.9–29.8 个百分点。
AlphaIdeaBench introduces a benchmark of 100 research-ideation tasks paired with explicit resource budgets, using an LLM judge validated against real ICLR/NeurIPS/ICML reviewer scores to evaluate eight ideation harnesses and eight backbone LLMs for novelty and feasibility.
FraudSociety 是首个联合建模信息与金融动态的 LLM 智能体仿真系统，用于模拟金融诈骗经济并评估反欺诈干预策略的有效性。
FlexA-Mem 提出一种预算自适应的智能体记忆机制，将历史一次性写入不可变存储，并为每次调用动态渲染紧凑的时间线记忆加相关细节，在 AppWorld、LongMemEval 和 LoCoMo 上显著优于 A-Mem 等现有方法。
This paper introduces WhoWins, a benchmark of 1,080 code-scored scenarios and controlled experiments testing how instruction position, chat role, and label affect which conflicting instruction LLM agents follow across 14 open-weight models. Key findings show that earlier instructions within a message win 72-95% of conflicts, and that system-message policy lines surprisingly lose to later instructions in Qwen models but win on gpt-oss.
AgentOrch 是一个基准，用于对比全局编排（Global）与逐步交互（Step-by-Step）两种 LLM 代理模式在 650 个企业 SOP 上的表现，发现最新模型的差距已缩小约 79%，并提出增量构建、执行反馈修正和针对性训练三条缩小差距的路径。
SLICE-Mem 提出一种轻量级 LLM Agent 长期记忆框架，通过显著性引导的记忆准入与基于谓词的语义抽取实现免 LLM 调用的在线记忆构建，在保持接近 MemOS 基线精度的同时将 token 消耗降低 92.4%–99.5%。
该论文提出了 Recursive Semantic Divergence (RSD)，一种基于强化学习双模拟度量的无监督轨迹级一致性指标，用于衡量 LLM 智能体在相同状态下多次独立执行时的行为差异；将其作为奖励信号进行微调可显著降低任务进展方差并提升性能。
AOS-Bench, from Tsinghua University and Columbia researchers, evaluates LLM coding agents as workload managers handling streams of tasks under shared CPU and memory limits, finding that isolated task success doesn't carry over to streaming scenarios and that simple fixed-concurrency policies often match elaborate agent-managed ones.
This paper systematically studies context compaction in LLM agent scaffolds (OpenCode) across three frontier models and three long-horizon benchmarks, finding that current benchmarks are 'compaction-friendly' and that compaction silently drops task constraints and anti-cheating rules, with real user instructions being dropped about 6x more often than build specifications.
WorkFactory introduces a scalable framework for constructing coherent, executable professional tasks via information-level planning and repair, producing 112,686 tasks across 612 occupations to fine-tune the WorkPilot model family (Qwen 4B/27B/35B-A3B), which improve over base models on professional-work benchmarks.
该论文提出，在KL正则化控制下，策略的log odds隐含其harness（提示、工具、控制逻辑）的偏好，部署性能等于harness潜力上限减去策略失配；实验在MiniWoB++和Terminal-Bench Lite上证明，仅用固定模型比较harness无法预测训练收益，而事先测得的ceiling与mismatch能准确预测。
SO-101 Bench 推出一个基于 Isaac Lab 数字孪生的低成本机器人臂语言条件空间操作仿真基准，结果显示混合 LLM 智能体（83.2%）显著优于学习型 VLA 策略（约61%），且仿真与真实硬件上的策略排名完全一致。
论文提出 TGR-EvoDAG，一个基于有向无环图（DAG）的多智能体代码进化框架，用于生成式推荐流水线的自动演化，离线指标最高提升 37.4%，并在腾讯工业部署中带来 0.40% 的在线 CTR 提升。
This paper proposes a new method for measuring whether LLM agents have internalized skills from public skill hubs, arguing that gain-based retirement rules fail to distinguish internalized skills from uncovered ones. It introduces a probe-based unaided scoring approach with cross-model validation and releases 1,251 skills with task data across four models.