Training the Whole Agent, Not Just the Model
Summary
HarnessAdapt-RL jointly trains a coding policy and adapts its evaluation harness via mechanism-level skills and evolving tools, improving pass@1 over GRPO by up to 12.6 points on SWE-bench variants with a 9B base model.
View Cached Full Text
Cached at: 10/21/26, 06:23 PM
Paper page - Training the Whole Agent, Not Just the Model
Source: https://huggingface.co/papers/2610.28069
Abstract
HarnessAdapt-RL jointly trains a coding policy and adapts its evaluation harness via mechanism-level skills and evolving tools, improving benchmark pass rates over fixed-harness reinforcement learning.
Training coding agents typically updates the policy while keeping the harness fixed. During training, trajectories reveal flawed repair strategies and interaction problems that vary across tasks and change as the policy learns. We introduceHarnessAdapt-RL, which trains the policy and adapts its harness duringreinforcement learning. For flawed repair strategies, a frozen teacher generates task-specificmechanism-level skills. For interaction problems recurring across tasks, a separate teacher model evolves tools and hooks, keeping each change only if it improves the observedpass rateor the observedtrainable ratewithout lowering either. The policy trains with the adapted harness, and its new trajectories guide further adaptation. With a 9B base model,HarnessAdapt-RLimproves pass@1 overGRPOby 9.8, 12.6, and 6.4 percentage points onSWE-benchVerified, Multilingual, and Pro, and outperformsGRPOwith the human-designed OpenHands harness. Gains frommechanism-level skillspersist after their removal at evaluation. Re-evolving tools and hooksas the policy changes outperforms freezing early versions. Removing them at deployment, in contrast, costs up to 7.7 percentage points. Training therefore produces a model-harness pair: the trained policy and the evolved tools and hooks, with skills removed at deployment.
View arXiv pageView PDFGitHub1Add to collection
Get this paper in your agent:
hf papers read 2610\.28069
Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash
Models citing this paper0
No model linking this paper
Cite arxiv.org/abs/2610.28069 in a model README.md to link it from this page.
Datasets citing this paper0
No dataset linking this paper
Cite arxiv.org/abs/2610.28069 in a dataset README.md to link it from this page.
Spaces citing this paper0
No Space linking this paper
Cite arxiv.org/abs/2610.28069 in a Space README.md to link it from this page.
Collections including this paper1
Similar Articles
HarnessRL: Co-Evolving Policies with a Just-in-Time Training Harness for Open-Ended RL
HarnessRL 提出一个与策略共同进化的训练框架，通过自适应任务采样、rollout 引导和 rubric 标准更新来解决开放式强化学习中的奖励失效问题，并发布了五领域评测集 Rubric-Mix-10k，在 4B 和 9B 模型规模上均优于最强基线。
LEGO-RL: Harness-Native Reinforcement Learning for Coding Agents
LEGO-RL presents a framework to bridge native coding-agent harnesses with scalable policy-gradient reinforcement learning, improving performance on benchmarks like SWE-bench.
EvoHarness-RL: Learning Self-Evolving Runtime Harness for Long-Horizon LLM Agents
Introduces EvoHarness-RL, a framework that learns runtime harness policies for long-horizon LLM agents, enabling them to construct and update external state (belief, progress, experience) during task execution. Using Qwen3-8B on ALFWorld, it achieves 96.9% success and reveals harness annealing and evolution dynamics.
ActiveSaddler: Automated Curriculum Learning for Agent Harness Optimization
ActiveSaddler 通过将动态课程学习建模为非平稳多臂老虎机问题，实现了 LLM Agent 优化中训练场景课程与 harness 的协同进化，在 GAIA2 和 Terminal-Bench2.0 上分别提升了 4.4 和 7.5 个百分点的 Pass@1。
Prime Agent: A Self-Improving RLM Harness
Prime Agent is an open-source harness that uses recursive subagents and persistent computation to extend language models' long-horizon capabilities across coding and reasoning tasks, significantly improving performance on benchmarks like ARC-AGI-3.