AdaCAD: Adaptively Decoding to Balance Conflicts between Contextual and Parametric Knowledge (2409.07394v1)

Abstract: Knowledge conflict arises from discrepancies between information in the context of a LLM and the knowledge stored in its parameters. This can hurt performance when using standard decoding techniques, which tend to ignore the context. Existing test-time contrastive methods seek to address this by comparing the LLM's output distribution with and without the context and adjust the model according to the contrast between them. However, we find that these methods frequently misjudge the degree of conflict and struggle to handle instances that vary in their amount of conflict, with static methods over-adjusting when conflict is absent. We propose a fine-grained, instance-level approach called AdaCAD, which dynamically infers the weight of adjustment based on the degree of conflict, as measured by the Jensen-Shannon divergence between distributions representing contextual and parametric knowledge. Our experiments across four models on six diverse question-answering (QA) datasets and three summarization tasks demonstrate that our training-free adaptive method consistently outperforms other decoding methods on QA, with average accuracy gains of 14.21% (absolute) over a static contrastive baseline, and improves the factuality of summaries by 5.59 (AlignScore). Furthermore, our analysis shows that while decoding with contrastive baselines hurts performance when conflict is absent, AdaCAD mitigates these losses, making it more applicable to real-world datasets in which some examples have conflict and others do not.

• The paper presents AdaCAD, which dynamically balances contextual and parametric knowledge using Jensen–Shannon divergence.
• It achieves significant performance gains, including a 14.21% improvement in QA accuracy and enhanced factuality in summarization.
• The method minimizes overcorrection in low-conflict scenarios, offering a robust solution for context-sensitive AI applications.

AdaCAD: Adaptively Decoding to Balance Conflicts between Contextual and Parametric Knowledge

The paper "AdaCAD: Adaptively Decoding to Balance Conflicts between Contextual and Parametric Knowledge" addresses a significant issue observed in LLMs—the discrepancies between contextual knowledge provided to the model and the knowledge encapsulated within its parameters, termed as "knowledge conflicts." Such conflicts can degrade the quality and accuracy of generated responses in tasks such as question answering (QA) and summarization. Traditional decoding methods, including greedy decoding, often predominantly rely on parametric knowledge, ignoring crucial context. Although test-time contrastive methods like context-aware decoding (CAD) offer some improvement by adjusting the output distribution based on contextual information, they frequently miscalculate the degree of conflict, leading to overcorrection in low-conflict scenarios.

Method

The authors propose a novel instance-level adaptive method, Adaptive Context-Aware Decoding (AdaCAD), to dynamically balance the influence of context and parametric knowledge. AdaCAD leverages Jensen-Shannon divergence (JSD) to measure the conflict degree between distributions representing contextual and parametric knowledge and infers adjustment weights accordingly, on a per-instance and per-token basis. This dynamic adjustment ensures accurate and reliable performance across varying degrees of conflict.

Key components of the proposed methodology include:

1. Jensen-Shannon Divergence: Used to compute a normalized, symmetric measure of the divergence between two probability distributions, enabling an interpretable metric for the degree of conflict.
2. Dynamic Adjustment: The JSD-derived value is used to reweight the combination of distributions, allowing the method to adaptively decide the influence of contextual knowledge versus parametric knowledge for every token.

Experimental Results

The effectiveness of AdaCAD is demonstrated through extensive experiments on four LLMs across six diverse QA datasets and three summarization tasks. The results are compelling:

• QA Tasks: AdaCAD yields an average accuracy gain of 14.21% over the static contrastive baseline CAD and 4.82% over COIECD across multiple LLMs. This robust performance highlights AdaCAD's ability to handle both high and low-conflict scenarios effectively.
• Summarization: AdaCAD improves the factuality of summaries significantly. For instance, using Llama3-70B, AdaCAD achieves an AlignScore gain of 4.16 over greedy decoding, 2.19 over CAD, and 10.44 over COIECD. These improvements indicate enhanced quality and factual consistency in long-form text generation.

Analysis

The analysis reveals that AdaCAD's dynamic approach effectively mitigates losses typically experienced by static contrastive methods like CAD in low-conflict scenarios. The approach's adaptive nature assigns lower adjustment weights in low-conflict instances and higher adjustments where conflicts are prominent, balancing the trade-off comprehensively. Quantitative assessments underline that AdaCAD perturbs the model’s output less in low-conflict cases compared to CAD, ensuring accuracy by avoiding overcorrection.

Implications and Future Directions

Practically, AdaCAD presents a significant improvement for real-world applications of LLMs, where mixed-context scenarios are prevalent. The method's adaptability could be pivotal in developing more reliable and context-sensitive AI systems, enhancing performance in knowledge-intensive tasks without the need for extensive manual tuning.

Theoretically, this work contributes to the ongoing discourse on dynamic inference techniques in NLP, setting a precedent for future research. Future developments could explore the integration of AdaCAD with larger LLMs and more complex tasks, investigating the broader implications of dynamically balancing context and intrinsic knowledge. Furthermore, the method's application could be expanded to other domains such as dialog systems and interactive AI, where context-aware and accurate responses are crucial.

Conclusion

AdaCAD’s fine-grained adaptive decoding introduces a robust mechanism to balance conflicts between contextual and parametric knowledge dynamically. By utilizing Jensen-Shannon divergence to infer the degree of adjustment, AdaCAD consistently outperforms existing decoding methods across diverse datasets and tasks. This adaptive strategy holds promise for enhancing the reliability and contextual relevance of outputs generated by large-scale LLMs.