1
美团上线 LongCat-2.5-Preview 模型：1.6T 参数，主打 AI 长程任务与多模态能力
IT之家 9 月 26 日消息，美团旗下 LongCat API 开放平台于 9 月 25 日上线新一代大模型 LongCat-2.5-Preview， 主打“长程任务”与多模态能力，并同步开放 API 与网页端体验入口。 IT之家附上官方更新日志内容如下： LongCat-2.5-Preview …
2
MemArbiter: Decision-Time Memory Arbitration for Long-Horizon LLM Agents
长程任务中LLM智能体记忆混乱？这项决策时记忆仲裁机制让智能体精准调用关键信息。
arXiv:2608.02113v1 Announce Type: new Abstract: Large language model (LLM) agents must retain and use cross-step information to act coherently in long…
3
4
QVal: Cheaply Evaluating Dense Supervision Signals for Long-Horizon LLM Agents
提出QVal框架，低成本评估长程LLM代理的密集监督信号，提升任务效率。
arXiv:2606.32034v1 Announce Type: cross Abstract: LLM agents increasingly act over long horizons, where a single trajectory can contain hundreds or th…
5
Semantic Consistency Policy Optimization for Reinforcement Learning of LLM Agents
突破组强化学习信用分配瓶颈，语义一致步骤不再因轨迹结局被错判奖励，显著提升LLM智能体长程稀疏奖励任务表现。
arXiv:2606.25852v1 Announce Type: cross Abstract: Group-based reinforcement learning effectively post-trains LLM agents for long-horizon, sparse-rewar…
6
BiPACE: Bisimulation-Guided Policy Optimization with Action Counterfactual Estimation for LLM Agents
用双模拟和动作反事实估计突破组相对策略优化的瓶颈，为LLM智能体长程决策提供新思路。
arXiv:2606.25556v1 Announce Type: cross Abstract: Stepwise group-based RL is an attractive way to train long-horizon LLM agents without a learned crit…
7
The Optimal Token Baseline: Variance Reduction for Long-Horizon LLM-RL
提出最优token基线方法，有效降低长程LLM强化学习的方差，提升训练稳定性。
arXiv:2602.07078v2 Announce Type: replace-cross Abstract: Reinforcement Learning (RL) for Large Language Models (LLMs) often suffers from training col…
8
When Agents Commit Too Soon: Diagnosing Premature Commitment in LLM Agents
揭秘LLM代理在长程任务中过早承诺的失败模式，提供系统诊断方法，提升AI决策可靠性。
arXiv:2606.22936v1 Announce Type: new Abstract: Long-horizon LLM agents can fail quietly: they settle on one reading of the evidence early, then spend…
9
Retrospective Progress-Aware Self-Refinement for LLM Agent Training
LLM智能体训练新突破：用回顾性进度感知自我改进，弥补元认知缺口，提升长程任务效果
arXiv:2606.14302v1 Announce Type: new Abstract: LLM-based agents trained with reinforcement learning optimize step-wise action prediction but lack met…
10