Multi-Adapter Representation Interventions via Energy Calibration

2026-05-27

AI Summary

This work addresses the limitation of existing representation intervention methods for large language models, which employ uniform strategies that often degrade general capabilities on benign inputs. To overcome this, the authors propose MARI, a multi-adapter representation intervention framework that innovatively integrates competitive nonlinear mixture-of-experts adapters with an energy-based gating mechanism derived from the model’s internal propagation dynamics. This design enables sample-adaptive selection of both intervention direction and intensity. Notably, MARI achieves behavior alignment without requiring fine-tuning, significantly improving performance on alignment benchmarks such as TruthfulQA, BBQ, and safety evaluations, while preserving or even enhancing general-purpose capabilities on tasks like MMLU and ARC.

Abstract

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.

Problem

• representation intervention
• large language models
• adaptive intervention
• alignment
• general capabilities

Innovation

• representation intervention
• multi-adapter mechanism
• energy-based gating
• adaptive alignment
• large language models