DIFFHEADS: Eliminating Bias in Large Language Models via Differential Analysis and Inference-Time Masking

DIFFHEADS is an open-source research project by the GeniusHTX team. Its core is to identify 'bias heads' that drive LLMs to produce unfair outputs through differential analysis, and mask these heads during inference to achieve lightweight, reversible debiasing without retraining the model. The project includes an automated evaluation tool and a multi-turn dialogue experimental framework, providing a new path for LLM fairness assurance.

With the widespread application of LLMs in various industries, the problem of systematic bias in their outputs regarding sensitive attributes such as race and gender has become increasingly prominent, which may exacerbate social inequality. Although traditional fine-tuning methods can alleviate bias, they require a large amount of labeled data and computing resources, and may impair the model's general capabilities.

DIFFHEADS stands for 'Differential Analysis and Inference-Time Masking of Bias Heads'. Its core idea is to identify 'bias heads' through differential analysis and mask these heads during inference to eliminate bias. This method is an inference-time intervention that does not require retraining, is lightweight and reversible, and can maintain the overall performance of the model.

The project includes four core modules:

1. Inference Pipeline (inference.py)：Supports single-turn/multi-turn dialogue experiments, allows configuration of model, batch size, etc. via command-line parameters, and enables inference mode to observe the thinking process;
2. Automated Evaluation (evaluate_judgellm.py)：Introduces 'judge LLM' for automated fairness evaluation, avoiding the high cost of manual annotation and ensuring consistency and scalability;
3. Unfairness Metric Calculation (evaluate.py)：Implements multiple quantitative metrics and supports fine-grained difference analysis;
4. Model Encapsulation Layer (llm_models/)：Provides a unified interface for easy integration of different pre-trained models for comparative experiments.

The project's usage workflow is intuitive:

1. Generate model outputs: Single-turn dialogue: python -u inference.py --model MODEL_NAME --batch_size 128 --output_path results/one_round --reasoning Multi-turn dialogue: python -u inference.py --model MODEL_NAME --batch_size 128 --output_path results/two_round --reasoning --data_name DATASET_NAME
2. Automated fairness evaluation: python -u evaluate_judgellm.py
3. Calculate unfairness metrics: python evaluate.py

DIFFHEADS has important theoretical and practical value:

• Theoretically: It reveals the causal relationship between specific sub-components of the attention mechanism and model bias, providing a new perspective for understanding the internal mechanisms of LLMs;
• Practically: The inference-time intervention method can be easily integrated into existing model service frameworks, providing a feasible path for ensuring model fairness in production environments. The project team stated that the code for 'guard head' identification and ablation experiments will be released soon, which will further improve the toolchain.

DIFFHEADS represents an important direction in LLM fairness research, shifting from 'black-box fine-tuning' to 'white-box intervention'. By accurately locating and masking bias heads, fairness can be improved without sacrificing model capabilities, and it may also be extended to scenarios such as reducing hallucinations and controlling tone for model behavior correction. For researchers and engineers concerned with AI ethics and interpretability, it is an open-source project worth paying attention to.