SuCo: Sufficiency-guided Continuous Adaptive Reasoning

Researchers have introduced two new frameworks, SuCo and CARE, aimed at improving reasoning in Large Reasoning Models (LRMs) and multimodal video reasoning.

The SuCo framework, submitted on June 16, 2026^[1], uses a two-stage training approach to enhance accuracy and efficiency in LRMs. It introduces Minimal Sufficient Chain-of-Thought (MSC), defined as the shortest prefix of a CoT trajectory adequate for producing the correct answer. SuCo's first stage, MSC-Aligned Fine-Tuning (MFT), constructs MSC data using problem-adaptive sufficiency thresholds. The second stage, Sufficiency-Aware Policy Optimization (SAPO), optimizes the model through reinforcement learning with dynamic complexity tracking and sufficiency-aware rewards. Meanwhile, a separate framework called CARE was submitted on June 18, 2026^[2], focusing on multimodal video reasoning. CARE is a competence-aware reward shaping framework that adapts reasoning length to the model's competence. It maintains a smoothed competence estimate via an exponential moving average of pass rates and normalizes reasoning effort with batch-level statistics. CARE is integrated into the GRPO training pipeline without incurring additional inference-time overhead.