BiMind: How the Dual-Head Reasoning Model Revolutionizes Misinformation Detection

BiMind proposes an innovative dual-head reasoning framework, separating internal content reasoning from knowledge-enhanced reasoning. By integrating the Attention Geometry Adapter, self-retrieval knowledge mechanism, and entropy-aware fusion strategy, it significantly improves the accuracy and interpretability of misinformation detection. This thread will introduce its background, architecture, technical innovations, experimental results, and application prospects in separate floors.

In the era of information explosion, misinformation poses threats to public health and social stability. Traditional detection methods struggle to balance text content verification and external knowledge correction, and suffer from the attention collapse problem (models overly focus on a few tokens while ignoring global semantics), leading to factual errors being overlooked.

BiMind Dual-Head Reasoning Framework

Inspired by the dual-system theory of cognition, it consists of:

1. Naive Head: Quick judgment based on text and TF-IDF;
2. Experienced Head: Injecting retrieved knowledge via FiLM modulation;
3. Fusion Layer: Dynamically combining outputs using an entropy-aware gate.

Three Major Technical Innovations

1. Attention Geometry Adapter: Using POS tags as biases to inject into hidden states, alleviating attention collapse without modifying the original weights of LLMs;
2. Self-Retrieval Knowledge Mechanism: kNN retrieval from domain semantic libraries, FiLM modulation of knowledge vectors, plus knowledge dropout to enhance robustness;
3. Uncertainty-Aware Fusion: Providing strategies like entropy gate, log average, PoE, and consistency head, with symmetric KL divergence regularization.

VoX Metric

Quantifies instance-level logit gains from knowledge enhancement. It can identify samples that need knowledge support, diagnose when and why the model relies on knowledge, and optimize knowledge base construction.

Experimental Results

On the ReCOVery (COVID-19 fake news dataset):

• Outperforms advanced methods, with significant improvements in accuracy and weighted F1;
• High interpretability: VoX and attention visualization show decision-making basis;
• Supports multiple fusion strategies, flexibly adapting to scenarios.

Evaluation metrics include accuracy/F1 of each head, symmetric KL divergence, and VoX gain.

Implemented based on PyTorch, supporting multiple LLM backbones:

Model	HuggingFace ID
LLaMA-2 7B	meta-llama/Llama-2-7b-hf
Mistral 7B	mistralai/Mistral-7B-v0.1
RoBERTa	roberta-base
DeBERTa	microsoft/deberta-v3-base

Deployment Recommendations:

• Use CUDA GPU;
• Load LLMs in float16 and freeze them;
• Dependencies: PyTorch, Transformers, Sentence-Transformers, spaCy, scikit-learn.

BiMind provides an interpretable and controllable paradigm for misinformation detection. Its significance lies in:

• Clearly understanding the model's decision-making basis;
• Targeted improvement of knowledge bases;
• Flexible adjustment of fusion strategies.

Outlook: Lays the foundation for multimodal fact-checking and real-time rumor detection. The dual-head design concept is expected to be applied to more tasks that require combining internal reasoning with external knowledge.