Megadose The AI curve, compressed.

Re-Search: Dissecting LLM Ideation Through a Two-Player Questioning Game

· HF Daily Papers ·
The models could recognize useful research ideas and questions they struggled to generate themselves.

Re-Search tests whether models can rediscover a published paper’s core idea by asking yes-or-no questions, using 200 ICML 2026 papers. Across six LLMs, nearly all progress came within the first 30 of 100 questions, followed by plateaus for most models. Controlled interventions identified generation as the bottleneck, with models either pursuing one direction too narrowly or exploring too many. The authors suggest externally supplied ideas and questions could help research systems advance.

HF Daily Papers' note

score 5

Categories: Research