In-Distribution Steering: Balancing Control and Coherence in Language Model Generation
This addresses the issue of degraded text plausibility in activation steering methods for real-world applications, though it appears incremental as it builds on existing steering techniques.
The paper tackled the problem of balancing control and coherence in language model generation by introducing In-Distribution Steering (IDS), which adapts steering strength based on input data distribution, resulting in strong accuracy on classification tasks and coherent text without collapse.
Activation steering methods control large language model (LLM) behavior by modifying internal activations at inference time. However, most existing activation steering methods rely on a fixed steering strength, leading to either insufficient control or unadapted intervention that degrades text plausibility and coherence. We introduce In-Distribution Steering (IDS), a novel method that adapts steering strength based on the input data distribution in representation space. IDS dynamically adjusts interventions according to how far a given input lies within the distribution, enabling adaptive intervention and generation stability during text generation. Experiments demonstrate that IDS achieves strong accuracy on classification tasks while producing coherent text without collapse, making IDS particularly well suited for real-world applications.