Why the geometry of LLMs reasoning might matter? June 15, 2026 reasoningllmoptimal-transportinterpretability