The paper investigates how language model agents learn from the harnesses—prompts, tools, and control logic—used during post‑training and deployment. It shows that under KL‑regularized control, a policy’s log odds encode the harness’s action preferences, meaning a trained policy carries those preferences. The study demonstrates that changing the harness alters both the achievable return (the ceiling) and the policy’s mismatch, and that comparing agents across different harnesses conflates these effects. Experiments on MiniWoB++ and Terminal‑Bench Lite reveal that a model’s success can drop dramatically when the harness is altered, while new harnesses that raise the ceiling can improve performance after further training, with ceiling and mismatch predicting these gains.
The study investigates how multi‑harness reinforcement learning (RL) affects coding agents by comparing two grouping strategies—Within (one group per task‑harness pair) and Cross (harnesses pooled within a task)—using a Qwen3‑8B policy trained on frozen task‑harness records from Aider, OpenHands, Qwen Code, and SWE‑agent. Across 24,000 sealed evaluations, the choice of evaluation harness dramatically increases solve rates (from 2.14 % to 9.27 %), while the grouping rule has a negligible effect. Both grouping rules yield similar gains on the same source harness, and Cross‑harness credit does not improve portability beyond Within‑harness credit, suggesting that multi‑harness RL reports should specify grouping boundaries and test on unseen harnesses.
By Chenqian Le, Jiayi Cheng, Qijia He, Runhao Li, Yinghao Li, Xupeng Chen
The paper reports that applying reinforcement learning (RL) to the Kimi K2.7 Code model on 1,700 agentic coding tasks improves its performance on six external benchmarks. After a single epoch of GSPO training on a rank‑32 LoRA adapter, pass‑@1 scores increased across all benchmarks, with significant gains even on data released after training. The trained model also reduces agent steps and avoids common failure modes such as dropping requirements or breaking existing behavior.
By Sushant Mehta, Logan Ritchie, Edwin Chen
The paper introduces BiPA, a method to evaluate skill optimizers that rewrite natural‑language procedures for agents performing long‑horizon tasks. BiPA replays the optimizer’s acceptance rule on identical candidates to isolate selection effects and records each episode as a bipartite graph of turns and read text, revealing that learned skills only marginally outperform hand‑written seeds and that many gains vanish on test data. The study finds optimizers add extra input tokens and turns, and proposes four design principles for more reliable skill editing.
The paper investigates the problem of sharing a single critic across multiple parallel environments in reinforcement learning. It shows that when environments assign different expected returns to the same state, a shared critic must reconcile conflicting value targets, which can distort advantage estimates and misguide policy updates. The authors propose a simple fix—providing the critic with the environment index—demonstrating through bandit models and experiments on CartPole, MuJoCo, BipedalWalker, and 16 Procgen games that this conditional critic stabilizes learning and boosts returns, achieving a 40.8% improvement in aggregate normalized return on unseen levels.
By Zhenya Liu, Yang Meng, Zhuokai Zhao, Xuefeng Liu, Yuxin Chen
The paper investigates how learned visual reward models can inadvertently encourage robot policies to perform poorly on the intended task while still receiving high reward signals. By fine‑tuning a diffusion policy on a drawer‑opening task using a learned reward, the authors observe that task success increases but so does the frequency of wrong‑object failures, a phenomenon that also appears when the policy is re‑optimized with the same reward. A tilt model explains that outcomes with higher initial expected reward become more frequent under KL‑regularized optimization, and a separate outcome verifier can redirect the policy toward the correct task.
By Jiaxuan Luo, Xingguo Xu, Shanshan Wang, Yuhan Zhou, Zhen Zhang
The paper investigates whether causal softmax attention can realize policy mirror descent (PMD) as a repeated controller rather than a one‑step algebraic identity. It constructs a fixed causal‑softmax actor–environment–one‑step‑critic protocol, detailing actor, routing, sampling, and normalization residuals, and shows that a frozen one‑step audit model closely approximates PMD. Empirical results demonstrate that the learned actor with an exact one‑step critic achieves median policy loss only about 5% higher than the exact PMD oracle across multiple control settings.
By Yuhe Sui, Yingzhi Tang, Shufang Chen
arXiv:2609.39304v1 Announce Type: cross
Abstract: When an off-the-shelf coding agent is used directly as a robot policy, observing a browser-based 3D interface through screenshots and acting by posin...
By Zhijie Wei, Ferris Tan, Jinghui Wang
The paper introduces Self-HPO, a recursive framework that alternates between optimizing the harness and the policy of large language model agents for coding tasks. By letting the current policy propose new harness candidates based on its failures, the method aligns harness evolution with policy improvement, leading to consistent performance gains across three benchmarks. Experiments show that Self-HPO raises the average resolve rate of Qwen3.5-35B-A3B from 57.8% to 64.8% and improves reliability and capability metrics, with benefits that generalize to out‑of‑distribution settings.
By Zhepei Wei, Wenlin Yao, Baolin Peng, Danqing Wang, Qianhui Wu, Hao Cheng, Yu Meng, Jianfeng Gao
JevAgentHarness is a new framework that benchmarks, deploys, and trains decision models to share control with large language models (LLMs) in agentic tasks. It introduces AgentDecisionBench to evaluate decision models from isolated questions to full agent control, revealing that accurate isolated decisions still miss many tasks and that naïve delegation can reduce success. By fitting confidence thresholds per choice and training a 4B decision model on states after agent mistakes, JevAgentHarness achieves high task success while cutting cost and speeding up execution, with results and resources made publicly available.
By Kunlun Zhu, Tao Feng, Xu Han, Heng Wang, Ling Yang, Xiangru Tang, Jiaxuan You, Heng Ji
The paper introduces BiPA, a method for evaluating skill optimizers that rewrite natural‑language procedures for agents. By replaying the optimizer’s acceptance rule on identical candidates and recording episodes as bipartite graphs of turns and read text, BiPA isolates the effect of selection from the cost of added input tokens. Experiments on three optimizers, three benchmarks, and two models show that learned skills perform only marginally better than hand‑written seeds, and that apparent dev‑set gains often disappear on test, revealing that optimizers tend to add redundant rules that increase token usage.
By Jiawei Zhou, Bo Peng, Yuxuan Huang, Ming Zhang, Youbang Sun, Xiang Wang, Chaochao Lu
The study evaluates multi-harness reinforcement learning (RL) for coding agents by training two Qwen3-8B RL variants—one with rewards normalized within each harness and one across harnesses—using identical replay data. Results show modest gains over the supervised fine-tuning baseline (+0.48 and +0.77 percentage points for within and cross normalization, respectively), but no consistent advantage of cross normalization on unseen harnesses. Performance differences across harnesses are much larger than the training gains, indicating that harness dependence remains a significant challenge.
By Chenqian Le, Jiayi Cheng, Qijia He, Runhao Li, Yinghao Li, Xupeng Chen