How Agents Decide: Goodfire's Eric Bigelow on Critical Tokens, Phase Shifts, & In-Context Learning

LLM "decisions" hinge on stochastic token sampling at critical junctures, and chain-of-thought monitoring is becoming less trustworthy as a safety mechanism.

• Forking paths: Bigelow's 2024 paper resampled 30 rollouts at every token of a model's chain-of-thought, finding outcome distributions can suddenly collapse at single 'critical tokens,' sometimes trivial ones like an open parenthesis.
• Decisions happen at sampling: Bigelow argues model behavior is determined less by hidden 'reasoning' than by the stochastic coin-flip of token sampling, which then gets locked in via in-context self-consistency.

Themes, sentiment and reasons are AI-written paraphrases, checked so they never quote an episode. They may contain errors. Not investment advice. Episodes remain the property of their creators.