一作ACL 2026Poster· Natural Language Processing
G r AI n S : Gradient-based Attribution for Inference-Time Steering of LLM s and VLM s
Duy Nguyen · Archiki Prasad · Elias Stengel-Eskin · Mohit Bansal
摘要
Inference-time steering provides a lightweight alternative to fine-tuning large language models (LLMs) and vision-language models (VLMs) by modifying model activations without updating weights. However, existing methods often rely on a global intervention vector, overlook token-level causal influ...