Reproduce “ThinkDeception: A Progressive Reinforcement Learning Framework for Interpretable Multimodal Deception Detection” on OpenResearch

About this paper

Jinhao Song, Shan Liang, Yiqun Yue, Zhuhuayang Zhang, Tianqi Gao

Researchers at Xi'an Jiaotong-Liverpool University developed ThinkDeception, an interpretable multimodal deception detection framework leveraging Multimodal Large Language Models (MLLMs) and progressive reinforcement learning. This framework achieved an average accuracy of 73.76%, an 8.52% absolute improvement over previous state-of-the-art models, while providing transparent, step-by-step reasoning.

Key insights

• General MLLMs, despite their capabilities in other visual tasks, perform poorly (near 50% accuracy) on deception detection without domain-specific adaptation and fine-grained reasoning guidance.
• Explicitly modeling cross-modal inconsistencies and providing process-aware rewards (visual-audio consistency, logic alignment) within an RL framework is crucial for both factual grounding and deep reflective reasoning in deception detection.
• A progressive training strategy, moving from easy to difficult samples, significantly enhances model stability and performance, enabling the framework to internalize a generalized reasoning paradigm.

Summary by alphaXiv.