Paper: strategy coopetition explains why in-context learning emerges and then disappears
scychan_brains · x · 2026-09-01
arXiv 2503.05631, from authors at DeepMind and Oxford (Aaditya K. Singh, Felix Hill, Stephanie C.Y. Chan, et al.), studies why in-context learning (ICL) emerges during training and later disappears.
Key findings:
- After ICL vanishes, the model's asymptotic strategy is a hybrid of in-weights and in-context learning the authors call CIWL (context-constrained in-weights learning);
- CIWL and ICL both compete and share sub-circuits — "strategy coopetition": ICL cannot emerge quickly on its own and is enabled by the slow development of CIWL;
- They propose a minimal mathematical model reproducing these dynamics and use it to identify a setup where ICL is truly emergent and persistent.
More from Research
- Boaz Barak: abandoning chain-of-thought before validated alternatives is irresponsible — inductionheads · 2026-09-03
- Developer once tried building AI benchmark from Puzzlescript, similar to ARC-AGI-3 — Darpinian · 2026-09-03
- He quarantined pre-1996 sources to build a 'clone' of Prof. Milhaupt as a sounding board — KarlMuth · 2026-09-03
- Do induction heads already explain LLMs' 'unprecedented' abilities? Researchers debate — aryaman2020 · 2026-09-03
- Do induction heads and attention sinks count? Debate over interpretability's missed milestone — aryaman2020 · 2026-09-03
- Counterfactual debugging scales sim2real failure diagnosis to 1M steps in world models — sarahcat21 · 2026-09-03