@Xudong07452910: Nowadays, many Agents are self-evolving, roughly taking two paths: One modifies model weights, the other modifies Harne…

X AI KOLs Timeline Papers
self-evolving-agents harness-evolution reinforcement-learning agentic-systems model-training recursive-improvement

Summary

这篇关于 Evo² 的文章提出将自进化 Agent 的两条路径（修改模型权重与修改 Harness）连接起来，让 Proposer 与 Reflector 双角色进入持续训练循环，在 Polyglot 上达到 82.1%、SWE-bench Verified 上达到 48.3%，且演化出的 Harness 可直接迁移复用。

Nowadays, many Agents are self-evolving, roughly taking two paths: One modifies model weights, the other modifies Harness. This Evo² paper directly connects the two paths. In the system, the same model simultaneously plays two roles: Proposer is responsible for modifying Harness, Reflector is responsible for reviewing execution results, analyzing what was modified correctly, and what still needs fixing. The key is in the next step. These Harness modifications and actual execution results will continue to be used to train Proposer and Reflector. The model becomes better at modifying Harness, and the new Harness generates better training experiences, forming a continuous loop. The result is that Evo² achieves 82.1% on Polyglot, compared to the strongest baseline of 62.6%; on SWE-bench Verified, it reaches 48.3%, an improvement over the baseline's 41.0%. Moreover, the evolved Harness can be directly transferred to other benchmarks without needing to search again. When fixing GPT-5-mini weights, just relying on this Harness evolution process also boosts performance from 24.4% to 48.3%. The truly forward-pushing step in this paper is to bring the "improver" itself into the learning loop. In the future, self-evolving Agents may increasingly resemble two mutually reinforcing systems: One side accumulates better external execution structures, while training models that are better at modifying these structures. When "how to make oneself stronger" also starts to be continuously trained, self-evolution truly gains a second layer of recursion. 📎 arxiv:
Original Article
View Cached Full Text

Cached at: 10/15/26, 02:07 AM

Nowadays, many Agents are self-evolving, roughly taking two paths:

One modifies model weights, the other modifies Harness.

This Evo² paper directly connects the two paths.

In the system, the same model simultaneously plays two roles:

Proposer is responsible for modifying Harness, Reflector is responsible for reviewing execution results, analyzing what was modified correctly, and what still needs fixing.

The key is in the next step.

These Harness modifications and actual execution results will continue to be used to train Proposer and Reflector. The model becomes better at modifying Harness, and the new Harness generates better training experiences, forming a continuous loop.

The result is that Evo² achieves 82.1% on Polyglot, compared to the strongest baseline of 62.6%; on SWE-bench Verified, it reaches 48.3%, an improvement over the baseline’s 41.0%.

Moreover, the evolved Harness can be directly transferred to other benchmarks without needing to search again. When fixing GPT-5-mini weights, just relying on this Harness evolution process also boosts performance from 24.4% to 48.3%.

The truly forward-pushing step in this paper is to bring the “improver” itself into the learning loop.

In the future, self-evolving Agents may increasingly resemble two mutually reinforcing systems:

One side accumulates better external execution structures, while training models that are better at modifying these structures.

When “how to make oneself stronger” also starts to be continuously trained, self-evolution truly gains a second layer of recursion.

📎 arxiv:

Similar Articles

Self-Evolving Agent Harness

Hugging Face Daily Papers

提出一个自我进化代理框架（Harness RSI），在保持模型权重固定的情况下，通过反馈驱动的修订与进化优化来持续改进代理的软件脚手架（提示、记忆、工作流和可执行组件），并借助持续 PAC 学习给出泛化与保留的理论保证。

Evo$^2$: Evolving the Evolver through Weight-Harness Co-Evolution

arXiv cs.CL

Evo²提出权重-支架协同进化框架，通过双角色（Proposer/Reflector）管线和EvoTIDE训练流程（SFT+GRPO），在Polyglot达到82.1%、SWE-bench Verified达到48.3%，显著优于现有自进化方法，并可跨模型与任务迁移。