Reproduce “When Attention Closes: How LLMs Lose the Thread in Multi-Turn Interaction” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Vardhan Dongre, Joseph Hsieh, Viet Dac Lai, Seunghyun Yoon, Trung Bui, Dilek Hakkani-Tür
The authors investigate whether models retain information about instructions after direct attention to them is lost. They measure attention to goals and impose sliding-window masks, while probing residual activations. In Mistral-7B, recall on a 20-fact task fell to 11.2% by turn 50, versus 100% under default inference; this intervention does not establish that residual information causes recall failures in ordinary conversations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.