V-STAR: Visual Anchoring Training Solves the Hallucination Problem of Multimodal Reasoning Models

The V-STAR framework solves the reasoning-visual disconnection problem of multimodal reasoning models at cognitive branching points through hierarchical visual attention rewards (HVAR) and forced reflection mechanisms (FRM). It transforms external debiasing interventions into the model's intrinsic hallucination suppression capability, achieving more reliable visual reasoning.

Multimodal reasoning models (MLRM) exhibit the reasoning-visual disconnection (RVTD) phenomenon. Hallucinations arise at cognitive branching points in long-chain reasoning—these are key decision moments in high-entropy states where the model tends to fall back on language priors instead of anchoring to visual evidence, and anchoring failure often occurs in intermediate layers with dense visual-language interactions.

1. Hierarchical Visual Attention Reward (HVAR)：Integrated into the GRPO framework, it dynamically incentivizes the model to focus on visual input at cognitive branching points. Basic rewards are given for bottom-level visual attention, and additional bonuses for visual attention at key nodes of high-level reasoning; 2. Forced Reflection Mechanism (FRM)：When a high-entropy branching point is detected, it forces the insertion of reflection steps, which are transformed into the model's autonomous behavior through training, actively verifying reasoning against visual input.

V-STAR is a lightweight training paradigm that can fine-tune existing MLRMs with low computational cost, flexible deployment, and fast iteration; it also has versatility, not targeting specific tasks or domains, and the trained model can be transferred to various downstream tasks.

Theoretical Significance: Challenges the traditional multimodal fusion assumption (intermediate layer visual anchoring is crucial), proves that attention patterns can be used as training targets, and provides an interpretability perspective for the reasoning process; Application Prospects: Medical image analysis (anchoring image features to assist diagnosis), autonomous driving perception (avoiding hallucinations or ignoring dangers), scientific image analysis (reducing subjective assumptions).

Limitations: Hyperparameter adjustment requires domain knowledge, and hallucinations may still exist in extremely complex scenarios; Future Directions: Develop intelligent entropy detection mechanisms, explore new multimodal attention architectures, and extend to more multimodal reasoning tasks such as audio/video.