SuCo: Sufficiency-guided Continuous Adaptive Reasoning

A research team has proposed SuCo, a two-stage training framework designed to give large reasoning models autonomous control over how much they “think,” aiming to cut computational waste while preserving or improving accuracy. The work was posted to the arXiv preprint server on 16 June 2026 [1]. Large Reasoning Models (LRMs) routinely produce excessively long Chain-of-Thought (CoT) sequences, driving up costs even for straightforward queries [2]. The authors argue that existing remedies — discrete reasoning modes or fixed token budgets — lack a principled way to decide when reasoning is enough [2]. To address this, they introduce Minimal Sufficient CoT (MSC), defined as the shortest prefix of a CoT trajectory that is adequate to produce the correct answer [2]. The SuCo framework is built in two stages. The first, MSC-Aligned Fine-Tuning (MFT), constructs training data using problem-adaptive sufficiency thresholds that scale naturally with question difficulty, then fine-tunes the model to internalize concise but sufficient reasoning patterns [2]. The second stage, Sufficiency-Aware Policy Optimization (SAPO), applies reinforcement learning with dynamic complexity tracking and sufficiency-aware rewards that penalize both over-thinking and under-thinking [2]. Across mathematics, code, and science benchmarks, SuCo consistently improved accuracy and reasoning efficiency, the paper states [2]. The preprint appeared on arXiv, an open-access repository that has hosted more than two million e-prints since its launch in 1991 and currently receives about 24,000 submissions per month [10]. The platform is not peer-reviewed; papers are approved after moderation [10]. The SuCo manuscript was submitted under the Computation and Language category [1]. The concept of minimizing reasoning steps to what is sufficient echoes broader debates in cognitive science about efficiency. Embodied cognition research, for example, investigates how cognition is shaped by bodily states and environmental interactions, challenging purely computational models of the mind [3]. While SuCo operates inside a machine-learning paradigm, its core question — when is a cognitive process “enough”? — touches on long-standing inquiries into the nature of reasoning and judgment [3].