On the Role of Verifiers and Thinking Traces in Reasoning Models
Abstract
Most of the recent successes of LLMs came from the application of Reasoning Models. I will provide a perspective on reasoning models in terms of verifier-based test-time scaling methods taken to post-training. From this perspective, post-training can be viewed as laboriously compiling the verifier signal into the model weights via guessed solutions to synthetic problems. Unlike standard LLMs, reasoning models also emit the so-called “thinking traces” on the way to guessing the solution. The literature has ascribed several properties to these traces–including that they provide the end user a window into the LLM’s “thinking”, and that the length of the traces is proportional to the complexity of the problem at hand. I will share results from our recent research that call these claims into question, and provide an alternate view of the role of intermediate tokens in the reasoning models.
Speaker
