Knowledge Injection Exists in MoE? Exploring Expert-Aware Contrast Decoding in MoE for Mitigating LLMs'Hallucinations
The study investigates whether layer-wise differences, previously observed in standard Transformers, exist in Mixture-of-Experts (MoE) architectures for hallucination mitigation. Empirical results indicate that while shared-expert MoEs lack these differences, distinct expert activation patterns emerge in higher layers between factual and non-factual outputs across different MoE models. The authors propose Expert-Aware Adaptive Contrast Decoding (EAACD), a method that splits high-layer experts in
Analysis
TL;DR
- The study investigates whether layer-wise differences, previously observed in standard Transformers, exist in Mixture-of-Experts (MoE) architectures for hallucination mitigation.
- Empirical results indicate that while shared-expert MoEs lack these differences, distinct expert activation patterns emerge in higher layers between factual and non-factual outputs across different MoE models.
- The authors propose Expert-Aware Adaptive Contrast Decoding (EAACD), a method that splits high-layer experts into reliability groups to calibrate predictions through contrastive decoding.
- EAACD amplifies hallucinations from lower-reliability experts via attention masking to create stronger negative references, significantly outperforming baselines on four QA datasets.
Why It Matters
This research bridges a critical gap by extending contrastive decoding techniques, previously limited to dense Transformer models, to the increasingly popular MoE architecture. It provides empirical evidence that expert activation patterns can serve as a reliable signal for detecting and mitigating hallucinations without requiring model retraining. For practitioners, this offers a plug-in inference-time solution to improve the factual accuracy of MoE-based LLMs.
Technical Details
- Empirical Analysis: The authors analyze layer-wise differences in MoE models, finding that shared experts do not exhibit the same patterns as dense layers, but higher layers show distinct activation distributions for factual vs. non-factual outputs.
- EAACD Mechanism: The proposed method divides high-layer experts into a "higher-reliability" group and multiple "lower-reliability" groups based on confidence and consistency metrics.
- Contrastive Calibration: Predictions from the higher-reliability group are contrasted with those from lower-reliability groups to adjust the final output probabilities.
- Hallucination Amplification: To enhance the contrast effect, the method uses attention mechanisms and masking to amplify the errors or hallucinations generated by lower-reliability experts, providing clearer negative signals.
- Evaluation: The approach is validated on four question-answering datasets, demonstrating superior performance compared to existing baseline methods.
Industry Insight
- Inference-Time Optimization: Organizations using MoE models can adopt EAACD to reduce hallucinations without the cost of fine-tuning or retraining, making it a scalable solution for production environments.
- Architectural Specificity: Developers must recognize that techniques effective for dense Transformers may not directly transfer to MoEs; understanding expert activation dynamics is crucial for optimizing next-generation sparse models.
- Reliability Grouping Strategy: The concept of dynamically grouping experts by reliability during inference could inspire new adaptive decoding strategies for other model types beyond MoEs.
Disclaimer: The above content is generated by AI and is for reference only.