MSA-Thinker: Discrimination-Calibration Reasoning with Hint-Guided Reinforcement Learning for Multimodal Sentiment Analysis

A research team has introduced MSA-Thinker, a novel framework designed to make Multimodal Large Language Models (MLLMs) more interpretable and robust for sentiment analysis. The core innovation is a two-stage training process that integrates structured Discrimination-Calibration (DC) reasoning with a new reinforcement learning technique called Hint-GRPO. The process begins with a 'cold-start' supervised fine-tuning (SFT) phase, where a powerful teacher model (Qwen3Omni-30B) synthesizes high-quality Chain-of-Thought (CoT) data. This data inherently contains the DC structure, teaching a smaller student model (like Qwen2.5Omni-7B) a reasoning paradigm that first performs a macro discrimination (e.g., positive/negative) followed by fine-grained calibration (e.g., intensity).

Building on this foundation, the team's Hint-GRPO method tackles a major RL challenge: sparse rewards on difficult samples. It leverages the initial discrimination phase as a verifiable anchor to provide directional hints during policy optimization. This guidance helps the model learn more efficiently by mitigating reward sparsity. Experiments on the Qwen2.5Omni-7B model show the framework not only achieves higher accuracy in fine-grained sentiment regression tasks but also generates high-quality, structured reasoning chains. Crucially, it demonstrates superior generalization in cross-domain evaluations, validating that explicit reasoning steps contribute positively to model robustness. This offers a new paradigm for building more trustworthy and efficient AI systems for understanding human emotion across text, audio, and visual data.