Shailesh Rana

This paper analyzes the internal decision-making process of large language models by tracking how the answer score changes across multiple internal computational steps (trajectories), finding that model correctness is often unstable.