Abstract
In Large Language Models (LLMs) generation, there exist knowledge conflicts and scenarios where parametric knowledge contradicts knowledge provided in the context. Previous works studied tuning, decoding algorithms, or locating and editing context-aware neurons to adapt LLMs to be faithful to new contextual knowledge. However, they are usually inefficient or ineffective for large models, not workable for black-box models, or unable to continuously adjust LLMs' sensitivity to the knowledge provided in the context. To mitigate these problems, we propose CSKS (Continuously Steering Knowledge Sensitivity), a simple framework that can steer LLMs' sensitivity to contextual knowledge continuously at a lightweight cost. Specifically, we tune two small LMs (i.e. proxy models) and use the difference in their output distributions to shift the original distribution of an LLM without modifying the LLM weights. In the evaluation process, we not only design synthetic data and fine-grained metrics to measure models' sensitivity to contextual knowledge but also use a real conflict dataset to validate CSKS's practical efficacy. Extensive experiments demonstrate that our framework achieves continuous and precise control over LLMs' sensitivity to contextual knowledge, enabling both increased sensitivity and reduced sensitivity, thereby allowing LLMs to prioritize either contextual or parametric knowledge as needed flexibly.
Introduction
Large Language Models (LLMs) often face a "decision-making dilemma." They possess vast parametric knowledge (facts learned during training), but they are also frequently provided with contextual knowledge (information in the prompt, often via Retrieval-Augmented Generation or RAG). When these two sources conflict—for example, if a user provides a document with updated facts that contradict the model's outdated training—LLMs can be surprisingly "stubborn," ignoring the provided context in favor of their internal beliefs. Conversely, they might be too "compliant," blindly following misinformation in the prompt.
Existing solutions like fine-tuning the large model are computationally expensive, while decoding-time tricks often lack nuance or fail to work on "black-box" models where weights cannot be modified. The CSKS framework addresses these issues by providing a mechanism to modulate this sensitivity precisely and efficiently.
Methods
The core of the CSKS framework is the principle of Proxy Steering. The process involves two main stages:
- Building the Proxies: Two small "proxy" models (e.g., 7B parameters) are fine-tuned. A positive model ($P$) is trained to be highly faithful to contextual information, while a negative model ($N$) is trained to prioritize its internal parametric knowledge.
- Contrastive Steering: During generation, the framework calculates the difference between the output distributions (logits) of both proxies: $(D_P - D_N)$. This difference represents the directional signal required to move from stubbornness toward compliance.
- The Steering Wheel ($\alpha$): This signal is added to the original distribution of the large target model ($L$), scaled by a hyperparameter $\alpha$. The final token is sampled from the resulting distribution: $\text{softmax}[D_L + (D_P - D_N)\alpha]$.
By adjusting $\alpha$, users can continuously tune the model's behavior. An $\alpha > 0$ increases sensitivity to context (making the model more "compliant"), while $\alpha < 0$ decreases sensitivity (making the model more "stubborn" or conservative).
Experimental Setup
The authors evaluated CSKS using several benchmarks, including ECQA for proxy tuning and MuSiQue and PopQA for synthetic evaluation. They also utilized DynamicQA, a real-world dataset featuring evolving facts (Static, Temporal, and Disputable conflicts) to test practical efficacy.
To measure performance, a new "Sensitivity Score" was introduced, accounting for three dimensions of conflict difficulty:
- Perturbation Degree: The extent of deviation from the truth.
- Contextual Detail: The richness of the provided context.
- Popularity: The frequency of the fact in the training corpus.
Key Results
CSKS significantly outperformed baselines like CAD and COIECD. On Llama-3-70B, it improved the Sensitivity Score by +30.26, and on Qwen2.5-72B, by +47.67. A key finding is the "3B Sweet Spot": a 3B proxy model is nearly as effective as a 7B proxy at steering a 72B model, offering high efficiency.
The framework's unique ability for bidirectional control allows it to not only make models more faithful to context but also less sensitive, which is critical if the input context is suspected of containing misinformation. However, there is a reasoning trade-off; extremely high values of $|\alpha|$ (typically $> 1.5$) can degrade general reasoning capabilities.
Conclusion
CSKS provides a lightweight and effective "steering wheel" for LLMs, enabling continuous adjustment of their sensitivity to contextual knowledge without retraining. While it offers superior performance and black-box compatibility, future work is needed to automate the calibration of $\alpha$ and mitigate reasoning degradation at high steering intensities.
Summary updated