When2Think replaces uniform length penalties with instance-level difficulty control, ending the efficiency tax

arXiv / HuggingFace Daily Papers·medium signalXBlueskyLinkedInCopy link

arXiv 2609.19671 (2026-09-17) frames reasoning models' systematic inefficiency (overthinking easy problems, underthinking hard ones) as an instance-adaptive compute allocation problem, and argues uniform length penalties and rigid routing both pay an efficiency tax by buying cheap easy instances with accuracy loss on hard ones. Its Instance-level Difficulty-Aware Control shapes rewards using pre-computed reference statistics of accuracy and token usage per instance, combined with verifier rewards and batch-wise standardized advantages, giving stable critic-free optimization with no learned reward model and no online reference-model queries.