cs.AINov 12, 2026

The Value of Thinking: Learning When to Reason in Hybrid Decision Models

Authors: Zhengxu Yu, Ke Xu, Meng Fang, Haitham Ammar, Weilin Luo, Jun Wang

Organizations: Huawei Noah’s Ark Lab · University of Liverpool · University College London

Abstract

Existing adaptive-reasoning methods shorten the explicit reasoning a model writes, or skip it by switching modes; either way the answer is still written out token by token. Rather than deciding how much to think, we explore a different paradigm, \emph{zero-or-TT thinking}: a single model either answers from the implicit thinking inside one forward pass, generating no token, or performs full autoregressive reasoning at a cost of TT tokens when that is worth paying. We realise it in a hybrid decision model (HDM), where one frozen backbone serves both modes and a lightweight self-routing block chooses between them. The block estimates the value of thinking, the utility the slow mode is expected to add over the fast one, from what that implicit thinking has already produced, and the slow mode runs only where that value exceeds a price λλ set per request. Where no per-decision label exists, as in episodic agentic tasks, we introduce Calibrated Decision Policy Optimisation (CDPO), a reinforcement-learning recipe that trains the block from episode-level returns. We evaluate HDM on MMLU-Pro, AGIEval, GPQA-Diamond, JevBench, BANKING77 and MiniWoB. Specifically, on JevBench HDM-4B scores 0.8570.857 against Jev's 0.8510.851 (+0.7%+0.7\%) at a median 0.1790.179,s against 0.6520.652 (−72.5%-72.5\%). On the agentic venue MiniWoB HDM-27B cuts token cost by 80.1%80.1\%.

Figures & tables

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

Explore similar work

CardsList
  1. DART: Draft-Agreement Routing for Training-Free Adaptive Thinking Budgets in Hybrid Reasoning Models

    Jun 22, 2026Jungseob Lee, Seongtae Hong, Seungjun Lee +7LLM RoutingEfficient Language Model Inference

  2. ADaPT: Token-Level Decoupling for Efficient Large Reasoning Models

    Jun 18, 2026Tingyun Li, Zishang Jiang, Jinyi Han +8Inference-Time OptimizationRL for Language Model Reasoning