CORTEXA
← Browse

Sagar Dangal

1 paper indexed

arxivcs.AI2026-07-09

Attention Degradation, Function Token Anchoring, and the Limits of Attention-Based Intervention in Large Language Models

Sagar Dangal, Manoj Shakya

Mean cross-positional attention degradation is widely reported in transformer interpretability, yet whether it causally limits contextual retrieval remains untested. We present six coordinated experiments across GPT-2, LLaMA-3.2-1B/3B, OPT-1.3B, and distilgpt2. We first character…

View free PDFSource page