Multi-Adapter Representation Interventions via Energy Calibration

Researchers propose MARI, a novel method for aligning large language models through adaptive representation interventions that adjust correction strength per input rather than applying uniform fixes. The approach combines multi-adapter experts with energy-based gating to maintain general model capabilities while improving alignment on safety and truthfulness benchmarks.

The research addresses a fundamental limitation in current alignment techniques for large language models. While representation intervention—modifying model activations without changing weights—has proven effective for steering model behavior, existing approaches treat all inputs identically. This one-size-fits-all strategy creates a critical trade-off: stronger interventions improve safety performance but degrade the model's general reasoning abilities on unrelated tasks. MARI resolves this tension through architectural innovation. The multi-adapter mechanism employs specialized experts that learn task-specific correction patterns, allowing the system to dynamically calibrate intervention strength based on input characteristics. The energy-based gating module acts as an intelligent filter, identifying which samples actually require intervention by analyzing internal model dynamics rather than relying on external classifiers. This represents meaningful progress in a crucial challenge facing the AI industry: achieving robust alignment without sacrificing model utility. The breadth of evaluation—spanning TruthfulQA, BBQ bias benchmarks, and general knowledge tasks like MMLU—demonstrates the method's practical viability across diverse use cases. For developers building aligned systems, MARI offers a path toward more nuanced control over model behavior. The published code enables rapid adoption and further iteration. The work exemplifies how careful engineering of intervention mechanisms can overcome apparent capability-alignment trade-offs, suggesting that future alignment methods may achieve both objectives simultaneously through smarter adaptation rather than stronger constraints. This approach aligns with industry trends toward interpretability-driven alignment techniques.

• →MARI uses adaptive multi-adapter experts to customize intervention strength per input rather than applying uniform corrections.
• →Energy-based gating distinguishes inputs requiring intervention by analyzing internal propagation dynamics.
• →The method improves safety benchmarks while maintaining or improving performance on general knowledge tasks.
• →Competitive multi-adapter mechanism captures non-linear correction patterns across diverse sample types.
• →Published code availability enables broader adoption and validation within the research community.