Agent fixes looping Qwen3.5 0.8B by editing weights — no training, no gradients
VictorSanh · x · 2026-11-12
The Beaker team pointed their autonomous agent-improver at a looping Qwen3.5 0.8B (93% fail rate on MATH-500, stuck generating "Wait"). Without ever running the model or using gradients, Beaker hypothesized, validated and hill-climbed its way to a weight edit toning down that tendency — a case study in viral "weight-space programming." Victor Sanh endorsed it as a real look behind the hype.
More from coding & agent
- Ora's Ranker 3.0 grades whether agents can actually use your site; scores drop 40+ points — yenkel · 2026-11-12
- Gemini API Managed Agents now support self-hosted sandboxes, incl. Vercel — steren · 2026-11-12
- Rayline launches ARC, a per-turn model router claiming 90% cost cuts — testingcatalog · 2026-11-12
- Karpathy's llm-wiki: a pattern for building personal knowledge bases with LLM agents — stochasticchasm · 2026-11-12
- Agent self-improvement with frozen weights: failure traces add just 4pp, blind search gets 60% — burny_tech · 2026-11-12
- Microsoft on agentic schedulers: LLM proposes, deterministic code validates — WirelessLife · 2026-11-12