Steering Language Models Before They Speak: Logit-Level Interventions
This addresses the need for consistent and fine-grained control in LLM steering for applications such as user-adaptive communication and toxicity mitigation, representing a novel method for a known bottleneck.
The paper tackles the problem of steering language models for specialized applications like style-sensitive rewriting and toxicity mitigation by proposing a training-free inference-time logit intervention method, achieving up to +47%p accuracy and 50x f1 improvement in evaluations across datasets.
Steering LLMs is essential for specialized applications such as style-sensitive text rewriting, user-adaptive communication, and toxicity mitigation. Current steering methods, such as prompting-based and activation-based approaches, are widely used to guide model behavior. However, activation-based techniques require deep access to internal layers, while prompting-based steering often fails to provide consistent or fine-grained control. In order to address these limitations, we propose a training-free inference-time logit intervention for controllable generation. Our approach utilizes a statistical token score table derived from z-normalized log-odds of labeled corpora to shift the decoding distribution. Empirical evaluations across three diverse datasets focusing on writing complexity, formality, and toxicity demonstrate that our method effectively steers output characteristics, confirming its broad applicability and task-agnostic nature. Our results show that statistically grounded logit steering can achieve large, consistent, and multi-task control gains: up to +47%p accuracy and 50x f1 improvement.