Microsoft's When2Think boosts AIME24 Pass@3 by 10% while cutting tokens 27.9%
microsoft · hf · 2026-09-18
Microsoft proposes When2Think, a post-training framework for hybrid reasoning that tackles the systematic inefficiency of large reasoning models: overthinking easy problems and underthinking hard ones. It formulates efficient reasoning as instance-adaptive compute allocation, using Instance-level Difficulty-Aware Control (IDAC) reward shaping with pre-computed accuracy/token reference statistics, enabling stable critic-free optimization without learned reward models. On AIME24, Pass@3 rises 10.0% while token usage drops 27.9%; on AIME25 it reaches 40.0% Pass@3, beating compression- and routing-only baselines.
More from Research
- PixelBench: testing multimodal models pixel-by-pixel instead of with QA — bdsqlsz · 2026-11-04
- François Chollet: modeling brain wiring as an engineering problem constrained by genome size and time — CatAstro_Piyush · 2026-11-04
- OffsecCyber dataset combines HackerOne reports and CVEs to train pentesting LLMs — AccBalanced · 2026-11-04
- Monofact rate in training data governs LLM scientific discovery, preprint finds — CatAstro_Piyush · 2026-11-04
- qwenTPU compiles a TPU into Qwen3.5 weights, achieving exact integer matmul — CatAstro_Piyush · 2026-11-04
- Jev model improves ICL demonstration retrieval via distillation into a lightweight MLP — kalyan_kpl · 2026-11-04