BiMind: Dual-Head Reasoning Model Revolutionizes Disinformation Detection, Attention Geometry Adapter Solves Attention Collapse Problem

BiMind separates in-content reasoning and knowledge-enhanced reasoning through an innovative dual-head reasoning framework, introduces an attention geometry adapter, a self-retrieval knowledge mechanism, and an uncertainty-aware fusion strategy, effectively solving the attention collapse problem. It also proposes the VoX metric to quantify knowledge contribution, achieves breakthrough progress in disinformation detection tasks, and provides a new direction for AI content moderation.

Disinformation detection needs to handle both in-content reasoning (text logic, linguistic features) and knowledge-enhanced reasoning (external fact verification) simultaneously. Traditional methods struggle to balance the two: either they lack fact-checking capabilities or ignore textual clues. What's more challenging is that attention collapse tends to occur when handling both, leading to a decline in model performance.

BiMind decouples the reasoning task into two independent heads:

• Content Reasoning Head: Focuses on the intrinsic features of text (logic, style, coherence) without external knowledge;
• Knowledge Reasoning Head: Retrieves external knowledge and verifies facts by comparing with the text. This design avoids attention conflicts and allows each head to focus on its specialized area.

1. Attention Geometry Adapter: Reshapes attention logits via token-conditional offsets to alleviate attention collapse;
2. Self-Retrieval Knowledge Mechanism: Builds a domain semantic memory bank, retrieves relevant knowledge using kNN, and smoothly injects it into the model via FiLM;
3. Uncertainty-Aware Fusion: Gated fusion based on entropy (weighted by confidence) + trainable consensus head, combined with symmetric KL divergence regularization to stabilize training.

BiMind significantly outperforms existing methods on public datasets, and proposes the VoX metric: by measuring the logit gain before and after introducing external knowledge, it quantifies the contribution of knowledge to sample judgment. A high VoX value indicates that knowledge is critical, while a low VoX value means text analysis is sufficient, enhancing the model's interpretability.

The success of BiMind implies:

• Decoupling complex tasks can improve performance;
• Interpretability (e.g., VoX) is crucial in sensitive applications;
• Fine-tuning the attention mechanism can solve multi-source information allocation problems. In the future, such AI systems that deeply understand text and effectively utilize knowledge will play a key role in maintaining the information ecosystem.