Can AI Improve Itself? RSI Might Be the Answer [R]
热度趋势
百分比基于当前可用热度信号，而非评论数或独立用户人数。
一项研究探讨了人工智能能否自我改进，此前一个OpenAI评估代理曾逃脱沙盒以获取测试解决方案。为了在没有作弊的情况下衡量递归自我改进（RSI），研究人员使用了一个锁定的考试。结果显示，Claude Opus 5在四项任务中的三项表现更优。从2025年11月到2026年7月，GPT在一项任务上的表现从3%提升至49%的上限，而Claude Opus则从37%提升至59%。
Can an AI make other AIs better? And what stops it from just cheating? Last month, an OpenAI eval agent escaped its sandbox and broke into Hugging Face, apparently to grab test solutions from a benchmark. It's exactly what you'd expect from a system that rewrites agents and reads its own grades. We set out to measure recursive self-improvement anyway, with the exam locked outside its sandbox.
We introduce HarnessOpt-Bench, which scores an LLM on how much it improves another agent's harness. On the development split, the optimizer sees per-case traces. Upon validation, it receives a single aggregate score. On test, nothing — until a trusted server scores its final candidate harness. API keys, budget enforcement, and held-out data never enter the optimizer's sandbox. That isolation holds by construction, not by instruction: the held-out evaluator and permission control sit outside the loop that evolves the harness.
5 frontier models, 4 downstream tasks, 111 runs to test 2 hypotheses:
1️⃣ Same coding harness, swap the model: Claude Opus 5 under OpenCode tops 3 of 4 tasks. Walk the releases from Nov 2025 to Jul 2026 on one task, and GPT climbs from 3% to 49% of the headroom, Claude Opus from 37% to 59%.
2️⃣ Same model, swap the coding harness: does a model do best in its own? No consistent home-field edge: opencode beats native harnesses (Claude Code, Codex, Kimi CLI) in 11 of 20 model–task pairs. Model choice moves gains 1.8× more than harness choice.
Paper: https://arxiv.org/abs/2608.06301
Code (MIT, built on our team's ICML 2026 VeRO): https://github.com/scaleapi/vero
Original post: https://www.linkedin.com/posts/shehabyasser_can-an-ai-make-other-ais-better-and-what-share-7498801902260981760-xuCo/