Thoughtology: Patterns in LLM Reasoning скачать в хорошем качестве

Thoughtology: Patterns in LLM Reasoning

Last week at Yandex Hall, Sara Vera Marjanovič, PhD Fellow at the University of Copenhagen and NVIDIA intern, presented her paper “Thoughtology: Patterns in Reasoning”. She explored how Large Reasoning Models like DeepSeek-R1 generate reasoning traces and what these reveal about their internal processes. The talk highlighted recurring reasoning patterns, the phenomenon of “rumination” in complex tasks, and the discovery of a reasoning “sweet spot.” The research shows that longer, repetitive reasoning chains often correlate with incorrect answers, emphasizing that more compute does not always mean better performance and pointing to the need for more efficient reasoning strategies. The event took place on February 17 at Yandex Hall, a public space for Armenia’s IT community by Yandex Armenia. We look forward to seeing you at our upcoming events!