Steering Conversations via Logit Bias in LLMs

ACL ARR 2025 February Submission8139 Authors

Abstract: Recent advancements in open-domain conversational agents highlight challenges in creating dynamic chatbot personalities. This paper explores Logit Bias as a novel mechanism for customizing LLM outputs, enabling seamless personality shifts without relying on static, dataset-constrained training. Unlike fine-tuning or prompt tuning, this method personalizes interactions without additional training, offering a flexible and efficient alternative. Through extensive experiments, we show that this approach effectively modifies model behavior while maintaining overall performance, influencing conversational quality and linguistic properties. This scalable solution allows for dynamically adaptable language models, meeting user expectations across diverse applications without requiring fine-tuning.

Paper Type: Short

Research Area: Dialogue and Interactive Systems

Research Area Keywords: Dialogue and Interactive Systems, Generation, Special Theme Track

Contribution Types: Model analysis & interpretability, Approaches low compute settings-efficiency

Languages Studied: English

Submission Number: 8139