Multi-Adapter Representation Interventions via Energy Calibration

Researchers introduced MARI, a multi-adapter representation intervention method for LLM alignment.

MARI uses specialized adapters plus energy-based calibration to decide when and how to intervene in model representations.

The paper reports stronger truthfulness and safety results on benchmarks like TruthfulQA and BBQ.

It also preserves or improves general task performance on models and tests including MMLU and ARC, reducing the usual alignment trade-off.