Back to AI Research

AI Research

Multi-Adapter Representation Interventions via Ener... | AI Research

Key Takeaways

  • Multi-Adapter Representation Interventions via Energy Calibration (MARI) is a research framework designed to improve how we align Large Language Models (LLMs...
  • 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).
Paper AbstractExpand

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 this https URL .

Multi-Adapter Representation Interventions via Energy Calibration (MARI) is a research framework designed to improve how we align Large Language Models (LLMs) with desired behaviors—such as truthfulness and safety—without needing to change the model's underlying weights. While existing methods often apply a single, uniform "fix" to every input, this paper demonstrates that such broad interventions can be ineffective and may even degrade a model's general reasoning capabilities. MARI solves this by making interventions adaptive, ensuring that the model only receives the specific type and strength of correction it needs for a given task.

The Problem with One-Size-Fits-All Interventions

Current alignment techniques often rely on the "linear representation hypothesis," which assumes that complex concepts like truthfulness can be corrected by applying a single, static vector to the model's internal activations. The authors found that this approach is fundamentally limited. Their analysis shows that the optimal direction and strength of an intervention vary significantly from one input to the next. When a model is forced to use a single, global intervention, it often over-corrects, which can disrupt the model's ability to handle benign or general queries, leading to a decline in performance on standard benchmarks.

How MARI Works

MARI introduces two primary innovations to handle this variability:

  • Competitive Multi-Adapter Mechanism: Instead of one global editor, MARI uses multiple lightweight "experts." During training, these experts compete to solve specific types of inputs, allowing them to specialize in different correction patterns. At inference time, the model uses an entropy-based router to automatically select the most confident expert for the specific query, ensuring the intervention is tailored to the input.

  • Energy-Based Gating: To prevent the model from intervening when it isn't necessary, MARI includes an energy-based gate. This module measures how a small "probe" update propagates through the model. If the signal suggests the input is benign, the gate suppresses the intervention entirely, allowing the model to rely on its original, frozen parameters. This protects the model's general capabilities by ensuring that interventions are only triggered when they are truly beneficial.

Results and Performance

The researchers tested MARI across various model families and scales. The results show that MARI achieves state-of-the-art performance on alignment-focused benchmarks, including TruthfulQA, BBQ, and safety-related tasks. Crucially, unlike previous methods that often sacrifice general intelligence for better alignment, MARI maintains or even improves the model's performance on general reasoning tasks like MMLU and ARC. By combining specialized expert routing with a selective gating mechanism, MARI provides a more reliable and precise way to steer LLMs toward desired behaviors.

Comments (0)

No comments yet

Be the first to share your thoughts!