MIT CSAIL paper: frozen LLM agents design their own self-improvement without being told how
omarsar0 · x · 2026-11-19
A new paper from MIT CSAIL and the University of Tübingen directly tests whether LLM agents still need to be told how to self-improve. Answer: no.
Setup: a frozen model gets a compute budget and simple tools — it can write practice games, watch another copy of itself play them, and edit that copy's prompt and skill files. It never sees real tasks or validation scores and is given no learning instructions.
Findings:
- Sufficiently capable models built a working learning procedure on their own and improved on held-out games with hidden mechanics;
- Models from different families converged on similar strategies: writing games at calibrated difficulty, running controlled A/B comparisons, building their own validation suites, and spending more effort on harder tasks;
- The text artifacts transferred: they improved play on external benchmarks with different horizons, mechanics, and 3D navigation.
The authors note this matters for safety too — when self-improvement no longer requires a human-specified loop, alignment assumptions need revisiting.
More from AGI Musings
- The site-builder era is repeating: people will hire humans to use AI app builders, says dev — sull · 2026-11-19
- Blogger goes contrarian: AI isn't currently moving toward higher general intelligence — LucaAmb · 2026-11-19
- Daily links: AI successionism debate, pluralistic AGI, and AI automation in physics research — gleech · 2026-11-19
- Developers push back against the thin-client future: 'ram altman doesn't want you to have RAM' — sull · 2026-11-19
- Anthropic Says Claude Solved Two Open Problems in Black Hole Physics — ghumare64 · 2026-11-19
- Reuters: China distills US model outputs to close AI gap despite chip export controls — r0ck3t23 · 2026-11-19