Multi-Adapter Representation Interventions via Energy Calibration

Representation intervention has emerged as a promising paradigm for aligning large language models toward desired behaviors without modifying model weights. Existing methods typically apply a fixed intervention uniformly across all inputs. However, we find that the appropriate intervention direction and strength vary substantially across samples, and such indiscriminate intervention leads to degradation of general capabilities on benign inputs. To address these challenges, we propose Multi-Adapter Representation Interventions via Energy Calibration (MARI). Specifically, we introduce a competitive multi-adapter mechanism in which specialized experts capture non-linear correction patterns and adaptively determine the appropriate intervention direction and strength for different samples. Furthermore, we design an energy-based gating module that leverages internal propagation dynamics to distinguish inputs that are applicable for intervention. Extensive experiments across diverse model families and parameter scales demonstrate that MARI achieves state-of-the-art alignment performance. Our method significantly improves performance on TruthfulQA, BBQ, and safety benchmarks, while maintaining and even improving general capabilities on tasks such as MMLU and ARC. Our code is available at https://github.com/V1centNevwake/MARI.

Large language models are often adjusted to make them show desirable attributes, such as being more truthful, fair, or safe. A common approach is to slightly change the model’s internal behavior while it is answering a question. However, existing methods often apply the same change to every input. This is too crude: some questions need a strong correction, some need a different kind of correction, and many ordinary questions do not need any correction at all. Applying interventions indiscriminately can therefore improve the target attribute on some examples, but also damage the model’s general ability on benign inputs. This paper proposes MARI, a more selective way to guide language models. Instead of using one fixed correction for all inputs, MARI uses multiple small specialist modules that can provide different corrections for different questions. It also includes a gate that first checks whether an intervention is needed. If the input does not appear to require correction, the model is left unchanged and answers normally. Experiments show that MARI improves language models on truthfulness, bias, and safety benchmarks while preserving, and sometimes improving, their performance on general reasoning tasks. This suggests that model behavior can be guided more precisely by intervening only when necessary and adapting the correction to each input, rather than applying the same intervention to everything.