Can AI Improve Itself? RSI Might Be the Answer [R]

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/