ACL 2026: Huawei Taylor Lab Proposes SHAPE, Adding a Reasoning Tax to LLM Inference

ACL 2026｜答得更准还写得更短？华为泰勒实验室提出SHAPE，给LLM推理装了个「推理税」

Huawei Taylor Lab, Peking University, and Shanghai University of Finance and Economics proposed SHAPE, accepted by ACL 2026, with about 3% average accuracy gain. It uses entropy segmentation, short rollouts for potential estimation, dynamic length discounts, and token-level credit assignment, cutting token use by about 30%. The key mechanism is a reasoning tax: long high-potential late-stage segments are penalized to reduce verbose confirmation loops.

Why it matters: HKR-H/K/R all pass: the paper gives testable gains of about +3% math accuracy and -30% tokens, with concrete mechanisms. It is a strong research item, not a same-day model-launch story.