Newest papers

Training the Whole Agent, Not Just the Model

Lingxiao Tang, Wei Tao, Yiwen Guo, Lingfeng Bao · 2026 · arXiv (Cornell University)

DOI: 10.48550/arxiv.2610.28069 · cited by 0 works (OpenAlex) · open-access PDF

Spanish

Los agentes de codificación de entrenamiento suelen actualizar la política mientras mantienen el arnés fijado. Durante el entrenamiento, las trayectorias revelan estrategias de reparación defectuosas y problemas de interacción que varían entre tareas y cambios a medida que aprende la política. Introducimos HarnessAdapt-RL, que capacita la política y adapta su arnés durante el aprendizaje de reforzamiento. Para las estrategias de reparación defectuosas, un maestro congelado genera habilidades específicas de mecanismo a nivel de tareas. Para los problemas de interacción que se repiten entre tareas, un maestro modelo separado evoluciona herramientas y hooks, manteniendo cada cambio solo si mejora la tasa de paso observada o la tasa de entrenamiento observada sin bajar ni siquiera. La política trae el arnés adaptado

Machine translation (Meta M2M100); the English original follows.

Abstract

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 introduce HarnessAdapt-RL, which trains the policy and adapts its harness during reinforcement learning. For flawed repair strategies, a frozen teacher generates task-specific mechanism-level skills. For interaction problems recurring across tasks, a separate teacher model evolves tools and hooks, keeping each change only if it improves the observed pass rate or the observed trainable rate without lowering either. The policy trains with the adapted harness, and its new trajectories guide further adaptation. With a 9B base model, HarnessAdapt-RL improves pass@1 over GRPO by 9.8, 12.6, and 6.4 percentage points on SWE-bench Verified, Multilingual, and Pro, and outperforms GRPO with the human-designed OpenHands harness. Gains from mechanism-level skills persist after their removal at evaluation. Re-evolving tools and hooks as 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.

Read in: English · Spanish · Chinese · Hindi · Arabic · Portuguese · French · German · Japanese · Russian · Korean · Indonesian · Vietnamese

Topics: Train, Training (meteorology), Computer science, Coding (social sciences), Reinforcement learning, Policy learning

Sources: OpenAlex (CC0), arXiv, Crossref. Agent Exchange