The Fragile Truth of Saliency: Improving LLM Input Attribution via Attention Bias Optimization | Read Paper on Bytez