Tag
微软的研究论文提出在用户轮次边界进行上下文感知的模型路由，发现将会话历史提供给路由器可在相同卸载率下提升多轮电子表格智能体的任务通过率，并分析了提示缓存切换的不对称成本与廉价模型的成本优势衰减。
This paper proposes Critique-Augmented Agentic Fine-Tuning (CAFT), a framework where expert LLMs provide both corrected actions and textual critiques explaining student model mistakes, achieving an 8% average improvement over action-only distillation on agentic benchmarks like τ²-Bench and WebShop.
Google 与高校合作提出 DTDD（Divergence-Triggered Dynamic Distillation），通过学生-教师分歧触发教师接管，为长程推理、从基座初始化及多轮智能体场景提供更可靠的 on-policy 监督，显著降低梯度范数并加速收敛。
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。
DeltaTTG 提出一种可复用的进度计量器（progress meter），以 Token-to-Go（剩余生成 token 数）作为完成进度的可观测代理，在不改变奖励的前提下重新分配多轮智能体训练中的优势值，在 ALFWorld、PaperSearchQA 和 BFCL v3 上显著优于 GRPO 等 RL 基线。
提出STI-OPD框架，通过基于教师-学生策略差异的随机教师干预和重要性加权反向KL目标，解决多轮agent任务中on-policy蒸馏的轨迹漂移问题，在工具推理与长程交互基准上超越先前最强OPD基线。
FC-SWE introduces failure-conditioned reinforcement learning for software engineering agents, using failed patch verifier feedback as context for recovery trajectories. On SWE-bench Verified with Qwen3.5-4B and SWE-agent, it achieves 41.7% Resolved@1 versus 38.9% for GRPO, and reaches 70.7% Resolved@11 at test time.
The paper introduces Calibrated User Embeddings (CUE), a training-free framework that encodes observed sessions and samples continuous persona embeddings decoded into prompts for LLM user simulators, achieving better outcome calibration and failure-mode reproduction than existing persona-based simulation methods on tau^2-Bench.
提出了名为StepLearn的非参数预序测试时学习框架，通过将个别交互转移转化为可即时使用、经前瞻性验证后才持久信任的假设，使LLM智能体在部署期间仅更新外部知识即可持续提升表现，效果显著优于EvoTest等基线。
PivotOPD proposes an on-policy distillation framework that jointly trains language agents to prevent pivotal mistakes and recover from the states they create, outperforming 13 baselines on ALFWorld, WebShop, SearchQA, and SWE-Bench Verified for Qwen3 and Nemotron students.
This paper introduces ActFirst-OPD, a framework that accelerates on-policy distillation for multi-turn language agents by decoupling action execution from full reasoning, achieving significant training speedups while maintaining performance across benchmarks.
The paper proposes Privileged Self-Practice (PSP), a method that injects privileged information into prompts for multi-turn LLM agents, improving task completion rates on AppWorld and SWE-bench benchmarks compared to existing distillation techniques.
This paper empirically analyzes cost savings in context-compression gateways for multi-turn coding agents, revealing that tool-schema filtering provides fixed token savings, while content compression saves quadratically but can be offset by recalls, offering actionable insights for cost optimization.
This paper proposes a dependency-aware trajectory refinement method for efficient multi-turn agent fine-tuning, improving accuracy and reducing inference costs.
The paper proposes RetireOPD, a method for training multi-turn agents using reinforcement learning with self-retiring on-policy distillation, improving performance on ALFWorld and WebShop benchmarks.
This paper introduces methods to calibrate confidence in agentic systems using internal representations, demonstrating improved performance over baselines in multi-turn benchmarks.
The paper introduces A-CEGIS, a framework that uses counterexample feedback to evaluate and improve multi-turn agent self-correction in natural-language-to-regex synthesis, achieving high success rates.
The paper introduces CREST, a method to improve multi-turn agent training by assigning separate credits to each turn and using a self-teacher to focus learning without overriding rewards, showing performance gains on benchmarks.
The paper introduces Retrieval-Grounded Voting (RGV) to address the limitations of confidence-based voting in multi-turn search agents by using lexical overlap with retrieved documents, achieving up to 5.4% accuracy gains.
CommitKV proposes a lifecycle-aware KV cache compression method for multi-turn ReAct agents, distinguishing dormant tokens from completed ones via commit transitions to reduce memory use and speed up inference.