1 h
Seminar, Hosted by Simons Institute for the Theory of Computing — The Future of Language Models and Transformers
DeepSeek-R1 Thoughtology: Thinking about LLM Reasoning
IVADO - Mila - McGill University
Recording
Abstract
Large reasoning models such as DeepSeek-R1 work through multi-step reasoning chains before answering, and because R1's chains are publicly visible they open a new field of study, "Thoughtology". Starting from a taxonomy of the building blocks of R1's reasoning, Siva Reddy's analyses examine the effect and controllability of thought length, how the model handles long or confusing contexts, cultural and safety concerns, and cognitive phenomena such as human-like language processing and world modelling. The findings include a "sweet spot" of reasoning beyond which extra inference time impairs performance, and a tendency to ruminate on formulations the model has already explored. He also presents VinePPO, a reinforcement-learning algorithm that improves reasoning.