这篇论文提出智能体和基准一起进化，15代后证明正确率从12.7%飙升到45.1%，思路很新。
一篇新论文主张自改进智能体在优化自身的同时也应演化其评估基准。研究者实现了一个自进化的Lean证明智能体，其工作流、提示和工具均可被重写。该智能体采用共同进化课程，仅在当前难度被掌握后才引入更难的证明任务。经过15代进化，共同进化智能体在miniF2F上的留出解决率达到45.1%，而种子版本仅为12.7%，固定基准最优版本为32.0%。论文强调将每步奖励锚定在形式验证器上，可避免奖励黑客攻击。
I agree with what this AI paper suggests. Self-improving agents should evolve their benchmarks too....
I agree with what this AI paper suggests. Self-improving agents should evolve their benchmarks too. (bookmark it) Self-improving agents are one of the most important directions in AI right now, and most of them optimize against a fixed benchmark. This paper argues the benchmark should evolve alongside the agent. The setup is a self-evolving Lean proof agent. A small trusted runtime wraps a fully mutable workspace of workflow, prompts, and tools, so the agent can rewrite how it decomposes proofs, uses compiler feedback, and repairs failures. What makes it different is coevolution. Between generations the champion revises the active task distribution through a mastery-throttled curriculum that adds harder proof obligations only after the current level is mastered. All of it stays inside a Lean-grounded loop, so a success counts only when the behavior yields verified proofs under a trusted snapshot. Over 15 generations the best coevolving agent reaches 45.1% held-out solve rate on miniF2F, versus 12.7% for the seed and 32.0% for the best fixed-benchmark agent. Grounding every reward in a formal verifier is how you get self-modification without reward hacking, and coevolving the curriculum pushes the ceiling higher. Paper: arxiv.org/abs/2607.17352 Learn to build effective AI agents in our academy: academy.dair.ai 💬 4 🔄 5 ❤️ 5 👀 1769 📊 6 ⚡