SuCo: Sufficiency-guided Continuous Adaptive Reasoning

2026-06-16

AI Summary

This work addresses the inefficiency of large reasoning models that often generate unnecessarily verbose chains of thought for simple tasks, leading to computational waste and a lack of principled criteria for assessing reasoning sufficiency. To remedy this, the paper introduces the notion of “Minimal Sufficient Chain-of-Thought” (MSC) as a new standard for reasoning adequacy and proposes a two-stage training framework. First, MSC-aligned fine-tuning internalizes concise yet sufficient reasoning patterns; second, reinforcement learning combined with dynamic complexity tracking enables adaptive control of reasoning length based on problem difficulty. Evaluated across mathematical, coding, and scientific reasoning benchmarks, the approach significantly improves accuracy while substantially shortening reasoning chains, achieving an optimal balance between efficiency and performance.

Abstract

Despite remarkable performance on complex tasks, Large Reasoning Models (LRMs) often generate excessively long Chain-of-Thoughts (CoT), inflating computational costs even for simple queries. Existing efforts to mitigate this inefficiency typically rely on discrete reasoning modes or fixed budget tiers, lacking a principled criterion of when reasoning is sufficient. In this work, we introduce Minimal Sufficient CoT (MSC), defined as the shortest prefix of a CoT trajectory which is adequate for producing the correct answer. We empirically show that MSC not only reduces reasoning tokens, but also improves accuracy across difficulty levels. Building on MSC, we propose Sufficiency-guided Continuous Adaptive Reasoning (SuCo), a two-stage training framework for autonomous reasoning control along a continuous spectrum. In stage 1, MSC-Aligned Fine-Tuning (MFT) constructs MSC data using problem-adaptive sufficiency thresholds that naturally scale with question difficulty, then fine-tunes the model to internalize concise yet sufficient reasoning patterns. In stage 2, Sufficiency-Aware Policy Optimization (SAPO) further optimizes the model through reinforcement learning with dynamic complexity tracking and sufficiency-aware rewards that penalize both over- and under-thinking. Extensive experiments across mathematics, code, and science benchmarks show that SuCo consistently achieves improvements in both accuracy and reasoning efficiency.

Problem

• Large Reasoning Models
• Chain-of-Thought
• reasoning efficiency
• sufficiency criterion
• computational cost

Innovation

• Minimal Sufficient CoT
• Continuous Adaptive Reasoning
• Sufficiency-guided Control
• Reasoning Efficiency
• Reinforcement Learning