TL;DR
This paper introduces CSKS, a novel framework that uses small "proxy" language models to continuously steer the sensitivity of large language models (LLMs) to contextual knowledge. This allows LLMs to dynamically balance their internal knowledge with external information, crucial for applications like retrieval-augmented generation where context quality varies.
What Problem Does This Paper Address?
Large Language Models (LLMs) possess vast amounts of parametric knowledge, but this knowledge can become outdated or incorrect. To address this, methods like retrieval-augmented generation (RAG) inject external, contextual information. However, LLMs can struggle to reconcile conflicting information between their internal knowledge and the provided context, leading to errors or hallucinations. Existing methods for improving contextual sensitivity are often inefficient, not applicable to black-box models, or lack the ability to precisely and continuously adjust this sensitivity, which is vital when dealing with external information of varying reliability.
Methods
The core of CSKS involves training two small "proxy" language models (LLMs) that are fine-tuned to be either highly sensitive to contextual knowledge (P model) or to prioritize their own parametric knowledge (N model). The difference in the output distributions of these two proxy models is then used to adjust the output distribution of a larger target LLM. This is achieved by adding a scaled differential term, controlled by a hyperparameter $\alpha$, to the target LLM's logits. By varying $\alpha$, the framework can continuously steer the LLM's sensitivity, either amplifying the influence of contextual information ($\alpha > 0$) or suppressing it to favor parametric knowledge ($\alpha < 0$). This process does not require modifying the weights of the large target LLM, making it efficient and applicable to black-box models.
Datasets
- Training Datasets for Proxy Models: The ECQA dataset was used to construct two fine-tuning datasets for the proxy models.
- Evaluation Datasets:
- Synthetic QA Data: A custom dataset was created by perturbing existing QA datasets (MuSiQue and PopQA) along three dimensions: degree of perturbation, contextual detail, and popularity. This allows for fine-grained evaluation of knowledge conflict resolution.
- Real-World Knowledge Conflicts: The DynamicQA benchmark was used to evaluate CSKS in more realistic scenarios involving evolving real-world information and different conflict types (static, temporal, disputable).
- Multi-hop Reasoning: The 2WikiMultiHopQA dataset was used to assess the impact of CSKS on complex reasoning abilities.
Key Insights and Ideas
- Proxy Models as "Steering Wheels": The use of small, adaptable proxy models to guide larger, more "stubborn" LLMs is a key innovation. These proxies act as specialized "experts" to provide a strong and reliable signal for contextual faithfulness.
- Continuous and Bidirectional Control: The hyperparameter $\alpha$ offers precise, continuous control over contextual sensitivity, allowing for both increased and decreased reliance on external information. This is a significant advantage over methods that offer discrete adjustments.
- Robustness of the Framework: Even without the "negative" proxy model (CSKS w/o negative), the framework shows competitive performance, highlighting the robustness of using proxy knowledge. However, the negative model further enhances the ability to distinguish between contextual and intrinsic knowledge.
- Trade-off Between Sensitivity and General Capabilities: While strong steering ($\alpha$ far from 0) can significantly enhance contextual sensitivity, extreme values can lead to a noticeable degradation in general reasoning abilities. This suggests the importance of finding a balanced $\alpha$ range.
- Efficiency and Black-Box Applicability: The framework's ability to steer LLMs without modifying their weights makes it efficient and applicable to black-box models, broadening its practical utility.
Key Results
- Superior Performance: CSKS consistently outperformed baseline methods (Prompt, CAD, COIECD) across various evaluation datasets (MuSiQue, PopQA, DynamicQA) and dimensions, achieving substantial average sensitivity score improvements (e.g., +30.26 for LLaMA-3, +47.67 for Qwen2.5).
- Precise Control: Experiments demonstrated that CSKS can precisely control contextual sensitivity. For instance, on DynamicQA, as $\alpha$ increased, overall accuracy monotonically improved. Conversely, negative $\alpha$ values suppressed contextual influence.
- Proxy Model Size: Smaller proxy models (even 0.5B) showed a noticeable impact on the 72B target model's sensitivity, with 3B proxies matching the performance of 7B proxies, validating the efficiency of the approach.
- Impact on Reasoning: While extreme steering values slightly impacted multi-hop reasoning performance (e.g., on 2WikiMultiHopQA), moderate $\alpha$ values allowed for effective context sensitivity adjustment while largely preserving sophisticated reasoning capabilities.
Limitations and Future Work
- Optimal $\alpha$ Calibration: The paper acknowledges that determining the optimal $\alpha$ value in real-world scenarios, where a balance between knowledge adaptation and general abilities is crucial, can be challenging. Future research could explore methods for automatically or adaptively determining $\alpha$.
- Generalization to Diverse Architectures: While CSKS showed strong transferability across different LLM families (Llama, Qwen, Gemma), further exploration across a wider range of architectures could be beneficial.
Summary updated