ADaPT: Token-Level Decoupling for Efficient Large Reasoning Models

2026-06-18

AI Summary

Existing efficient reasoning methods often sacrifice model performance due to sequence-level coupling, making it difficult to balance computational efficiency and accuracy. This work proposes the Adaptive Dual-process Thinking (ADaPT) framework, which explicitly decouples efficiency and correctness objectives at the token level for the first time. ADaPT introduces mode-selection tokens, designs a token-level reward allocation strategy, and incorporates a controllable generation probability modulation mechanism. This approach avoids penalizing correct yet verbose reasoning paths during training and enables continuous adjustment of the efficiency–accuracy trade-off at inference time. Experiments demonstrate that a single ADaPT model spans the Pareto frontier, significantly reducing computational overhead while maintaining strong reasoning capabilities across multiple benchmarks.

Abstract

Large reasoning models rely on long chain-of-thought to achieve strong performance, but applying such reasoning uniformly incurs high computational cost. Existing efficiency-oriented methods attempt to shorten or mix reasoning strategies, yet often degrade reasoning capability. We identify the root cause as sequence-level coupling between efficiency incentives and correctness optimization, which implicitly penalizes long but correct reasoning trajectories. To address this issue, we propose Adaptive Dual-Process Thinking (ADaPT), a token-level dual-process framework that explicitly decouples efficiency and correctness signals during training. ADaPT introduces a mode-selection token to control fast and slow reasoning, applying efficiency-related rewards exclusively to this token to avoid penalizing correct long reasoning while encouraging efficiency when appropriate. Moreover, ADaPT enables precise and continuous control over the efficiency-performance trade-off at inference time: by adjusting the generation probability of the mode-selection token, a single trained model can smoothly move along the efficiency-performance Pareto frontier. Extensive experiments demonstrate that ADaPT significantly reduces inference cost while maintaining strong reasoning performance across multiple benchmarks.

Problem

• large reasoning models
• computational efficiency
• reasoning capability
• sequence-level coupling
• efficiency-performance trade-off

Innovation

• token-level decoupling
• dual-process reasoning
• efficiency-performance trade-off
• mode-selection token
• Pareto frontier