The Fragile Truth of Saliency: Improving LLM Input Attribution via Attention Bias Optimization
NeurIPSSpotlight2025
TL;DR
Input saliency aims to quantify the influence of input tokens on the output of large language models (LLMs), which has been widely used for prompt engineering, model interpretability, and behavior att…
Opening excerpt from the authors’ abstract. source
Read the paper
Topics
large language model interpretability language model optimization attention llm rag
← All NeurIPS 2025 Spotlight papers · Browse the whole archive