1 paper across 1 session
We develop a theoretical model of information flow constraints in LLMs that predicts LLM failures on global reasoning tasks.