ThinkDeception Framework Boosts Interpretable Multimodal Deception Detection

Researchers have developed ThinkDeception, an innovative framework for detecting deception that combines multimodal large language models (MLLMs) with a progressive reinforcement learning approach. This system moves beyond traditional black-box classification by framing deception detection as an explicit cognitive reasoning task, providing transparent explanations for its conclusions. A key component is the first meticulously annotated step-by-step multimodal Chain of Thought (CoT) dataset, which helps train the foundational model, ThinkDeception Base, to identify crucial modal inconsistencies indicative of deception. The core innovation, Visual-Audio Consistency Group Relative Policy Optimization (VAC-GRPO), employs a progressive training strategy across four difficulty tiers, guided by a multi-dimensional reward mechanism and reflective learning. This method significantly enhances reasoning quality and achieves state-of-the-art performance on benchmarks.