Your Policy Is Secretly a Harness: The Decision Geometry of Agent Interfaces

Hugging Face Daily Papers Papers
llm-agents reinforcement-learning kl-regularization agent-harness post-training benchmark

Summary

该论文提出，在KL正则化控制下，策略的log odds隐含其harness（提示、工具、控制逻辑）的偏好，部署性能等于harness潜力上限减去策略失配；实验在MiniWoB++和Terminal-Bench Lite上证明，仅用固定模型比较harness无法预测训练收益，而事先测得的ceiling与mismatch能准确预测。

Language model agents are post-trained and deployed through a harness, the prompts, tools, and control logic that present a task to the model and execute its outputs. Harnesses are often revised and judged by how an already trained model performs under them, which assumes that what the model learned under one harness carries over to the next. We ask what a policy learns from its harness and what survives when the harness changes. Under KL-regularized control, we show that the log odds of an optimal policy recover how strongly its harness favors each action over the others, so a trained policy carries the preferences of its harness. Deployed return then equals the ceiling of the harness, the best return it offers, minus the mismatch of the policy, the return lost to its own suboptimal decisions. Training changes only the mismatch, while changing the harness moves both terms. Comparing harnesses with a fixed model therefore mixes what each harness offers with how well that model uses it. In MiniWoB++, post-training makes a model's preferences specific to its training harness. Reversing the order of the observation fields leaves every optimal decision unchanged yet lowers a trained model's success from 85.0% to 19.2%. Conversely, every new harness that a fixed-model comparison rejects but that raises the ceiling outperforms the original agent after further training under it, and ceiling and mismatch measured beforehand predict these gains. For coding-agent harnesses on Terminal-Bench Lite, ceiling minus mismatch estimated with a stronger model as a stand-in for the optimum reproduces the accuracy ranking, while either term alone does not. Code is available at https://github.com/2020-qqtcg/HRP.
Original Article
View Cached Full Text

Cached at: 10/28/26, 05:24 AM

Paper page - Your Policy Is Secretly a Harness: The Decision Geometry of Agent Interfaces

Source: https://huggingface.co/papers/2610.37255 Published on Oct 27

·

Submitted byhttps://huggingface.co/mujianijan

mjon Oct 28

Abstract

Under KL-regularized control, a policy encodes its harness preferences, so deployed performance equals harness potential minus policy mismatch, and harness comparisons must separate these terms to predict training gains.

Language model agents are post-trained and deployed through aharness, the prompts, tools, and control logic that present a task to the model and execute its outputs.Harnesses are often revised and judged by how an already trained model performs under them, which assumes that what the model learned under oneharnesscarries over to the next. We ask what a policy learns from itsharnessand what survives when theharnesschanges. UnderKL-regularized control, we show that thelog oddsof anoptimal policyrecover how strongly itsharnessfavors each action over the others, so a trained policy carries the preferences of itsharness. Deployed return then equals theceilingof theharness, the best return it offers, minus themismatchof the policy, the return lost to its own suboptimal decisions. Training changes only themismatch, while changing theharnessmoves both terms. Comparingharnesses with a fixed model therefore mixes what eachharnessoffers with how well that model uses it. InMiniWoB++, post-training makes a model’s preferences specific to its trainingharness. Reversing the order of the observation fields leaves every optimal decision unchanged yet lowers a trained model’s success from 85.0% to 19.2%. Conversely, every newharnessthat a fixed-model comparison rejects but that raises theceilingoutperforms the original agent after further training under it, andceilingandmismatchmeasured beforehand predict these gains. For coding-agentharnesses onTerminal-Bench Lite,ceilingminusmismatchestimated with a stronger model as a stand-in for the optimum reproduces the accuracy ranking, while either term alone does not. Code is available at https://github.com/2020-qqtcg/HRP.

View arXiv pageView PDFGitHubAdd to collection

Get this paper in your agent:

hf papers read 2610\.37255

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/2610.37255 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2610.37255 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2610.37255 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles

The Harness as the Only Mutable Surface: Compliance-Bounded Self-Evolution of LLM Agents in Credit Pipelines, with a Measured Admission Gate

arXiv cs.AI

该论文提出一种将LLM代理的自进化限制在运行时harness（指令文本、工具调用逻辑与原语组合）的框架，模型权重保持固定，使每次适应都成为可审查的diff，并通过哈希链记录的准入门进行管控。模拟实验显示受控双循环仅放行144/7449个候选变更且未恶化误差，而无约束系统放行了309个有害变更。

Stop Comparing LLM Agents Without Disclosing the Harness

arXiv cs.AI

This position paper argues that in long-horizon LLM agent tasks, the execution harness often determines performance more than the model itself, and current benchmarks misattribute harness-level gains to model improvements. It proposes a harness-aware evaluation framework with disclosure standards and variance decomposition protocols.