AutoHarness：通过自动合成代码运行框架改进 LLM Agent（论文摘要）

完整译文与原文逐段对应。图片、图注、表格和代码保留原文。A complete reading edition. Figures, captions, tables and code are preserved from the source.

中文译文ENGLISH ORIGINAL

阅读论文全文（PDF）

Read the full paper (PDF)

阅读论文全文（HTML）

Read the full paper (HTML)

摘要

Abstract

尽管语言模型在过去几年取得了显著进步，但在作为 Agent 使用时，这类模型仍经常尝试执行某些行动：这些行动不仅在给定状态下并非最优，甚至是外部环境明确禁止的。例如，在近期的 Kaggle GameArena 国际象棋比赛中，Gemini-2.5-Flash 的败局有 78% 归因于非法走法。人们经常围绕 LLM 手动编写“运行框架”（harness），以防止此类失败。在本文中，我们表明，Gemini-2.5-Flash 能够根据（游戏）环境的反馈，经过少量轮次的迭代代码改进，自动合成这样的代码运行框架。生成的运行框架在 145 种不同的 TextArena 游戏中（包括单人和双人游戏）阻止了所有非法行动，使规模较小的 Gemini-2.5-Flash 模型能够超越 Gemini-2.5-Pro 等更大模型。将这项技术推向极限时，我们可以让 Gemini-2.5-Flash 用代码生成完整策略，从而无需在决策时使用 LLM。在 16 款 TextArena 单人游戏中，生成的代码策略获得的平均奖励高于 Gemini-2.5-Pro 和 GPT-5.2-High。我们的结果表明，使用较小的模型合成定制代码运行框架（或完整策略），可以超越大得多的模型，同时也更具成本效益。

Despite significant strides in language models in the last few years, when used as agents, such models often try to perform actions that are not just suboptimal for a given state, but are strictly prohibited by the external environment. For example, in the recent Kaggle GameArena chess competition, 78% of Gemini-2.5-Flash losses were attributed to illegal moves. Often people manually write "harnesses" around LLMs to prevent such failures. In this paper, we demonstrate that Gemini-2.5-Flash can automatically synthesize such a code harness, using a small number of rounds of iterative code refinement given feedback from the (game) environment. The resulting harness prevents all illegal moves in 145 different TextArena games (both 1-player and 2-player), enabling the smaller Gemini-2.5-Flash model to outperform larger models, such as Gemini-2.5-Pro. Pushing our technique to the limit, we can get Gemini-2.5-Flash to generate the entire policy in code, thus eliminating the need to use the LLM at decision making time. The resulting code-policy receives a higher average reward than Gemini-2.5-Pro and GPT-5.2-High on 16 TextArena 1-player games. Our results show that using a smaller model to synthesize a custom code harness (or entire policy) can outperform a much larger model, while also being more cost effective.

文献信息

Bibliographic information

Comments:	agent harness, code synthesis, self-improvement, code-as-policy, text games
Subjects:	Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
Cite as:	arXiv:2603.03329 [cs.CL]
	(or arXiv:2603.03329v1 [cs.CL] for this version)
	https://doi.org/10.48550/arXiv.2603.03329 arXiv-issued DOI via DataCite

提交历史　提交者：Xinghua Lou［查看电子邮件］　[v1] 2026 年 2 月 10 日，星期二，14:12:54 UTC（321 KB）

Submission history From: Xinghua Lou [view email] [v1] Tue, 10 Feb 2026 14:12:54 UTC (321 KB)

— 全文完 —

原文来自 arXiv，中文为非官方学习译文。
查看原始出处 ↗