Real Alpha
Watch mode

The Cognitive Revolution · 2 min instead of 2 h 1 m · Oct 10, 2026

Resampling every token shows where answers collapse

EB

Eric Bigelow

Researcher, Goodfire

  • Nathan LabenzHost · Host, The Cognitive Revolution

    Founded AI video company Waymark and red-teamed GPT-4 before its release.

About this segment

Bigelow describes resampling about 30 rollouts at every token of a reasoning chain. Uncertainty over the final answer can collapse at one token, even an open parenthesis.

From “How Agents Decide: Goodfire's Eric Bigelow on Critical Tokens, Phase Shifts, & In-Context Learning”, Oct 10, 2026. Watch the full episode

“there will be certain points in reasoning that the model was initially sort of at a 50% chance of this answer and 30% chance of this answer.”
EBEric Bigelowat 7:55 · The Cognitive Revolution7:55

Takeaways

  1. Resample about 30 rollouts per token to map uncertainty over final answers
  2. Distributions can shift sharply at single, seemingly arbitrary tokens
  3. This echoes the phase transitions seen in grokking and training dynamics
Watch 2:14 from 6:30

In this episode

Full episode

  1. 6:30Resampling every token shows where answers collapse2:14

  2. 32:33Reasoning models enumerate options, then pick from the soup2:55Up next

Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

  1. 6:36

    but i'll give a quick overview of um a project that i think is really one of the most informative projects and how i think about how lms work at a basic level and that's uh my project forking paths and neural text generation and for that basically so this was two years ago with at the time it was one of the gpt 3.5 models and what we do is we take uh reasoning chains we give it a chain of we give it a question and a chain of thought prompt and have it reason through you know step by step and then every single token in its reasoning we resample all these completions so we resample like 30 different rollouts at every single token and then from every one of those rollouts we see what final answer it landed on and you can aggregate these into various distributions and then do analysis on these distributions but I think the most striking part of this is if you

  2. 7:32

    arrange it like a multivariate time series, you see these really interesting dynamics that are happening. And I kind of take inspiration from this, from some of the learning dynamics work that was coming out around that time, like around grokking and things like this, where there are these really sharp phase transitions in training, and the model just seems to suddenly learn something. And it turns out that you actually have these during reasoning and in context learning as well, where over the course of reasoning, there will be certain points in reasoning that the model was initially sort of at a 50% chance of this answer and 30% chance of this answer.

  3. 8:09

    And then there's some point that it hits where suddenly the distribution changes really dramatically. And often the distribution just kind of collapses onto a final answer as the model sort of, you might say, decides what it's going to say next. Although I think I use the word decides here with maybe scare quotes around it because I think there's some nuance to what that means but one of my big takeaways from this and um I think this also just kind of falls out of thinking about the architecture but it's nice to have results that really support this too is