Survey: most self-improving agents only tune prompts and tools, rarely weights
maier_ak · x · 2026-09-22
A survey from Jürgen Schmidhuber's group finds most self-improving agents only tweak the fast scaffold — prompts, memory, tools — while keeping model weights frozen. Genuine weight updates are rare and risk collapse. Future work targets test-time adaptation and budget-aware improvement.
It also proposes logging self-improvement as checkpointed performance curves under a fixed update budget.
More from AGI Musings
- Business strategy debate: hide the AI behind humans, not humans behind AI — annetgriffin · 2026-11-12
- If superhuman AI researchers are imminent, will capability problems be easier than safety ones? — ronbodkin · 2026-11-12
- "AI is a terrible teacher": viral pushback against personalized AI tutoring hype — RachelVT42 · 2026-11-12
- TissueMicroscopyBench: 49 evals to test AI agents on biological microscopy — kenbwork · 2026-11-12
- OpenAI, Anthropic and DeepMind job posts signal PMs must write evals and code — aakashgupta · 2026-11-12
- Architecture study sparks monitoring concerns over a multi-hop planning spike — ChowdhuryNeil · 2026-11-12