Goodfire: a handful of forking words decide where a reasoning chain lands
The company resamples every step of a model's chain of thought to find the tokens that actually change the outcome. It has also made its research agent Silico generally available.
Goodfire, which works on interpretability, has published a method for examining the reasoning chains of language models: Forking Paths Analysis.
The logic is straightforward. As a model thinks through a question step by step, you re-run the chain many times from each point onward. You then collect the answers all those alternative paths arrive at and turn them into a distribution. That shows where along the chain the answer actually changes.
What comes out is this: not every part of a long chain of thought carries equal weight. A few forking tokens decide the outcome, and the model's answer often sits a single word away from a different one. The company argues this matters for oversight of reasoning, because watching a few critical points may be enough rather than the whole chain. Goodfire reports the method yields the same information at far lower compute than earlier approaches; that efficiency figure is the company's own measurement.
Goodfire has also made its research agent Silico generally available. By the company's account, the platform reproduced three studies from the interpretability literature — Interpretability in the Wild, ROME and Mixing Mechanisms — without access to the authors' code. Silico is offered as a paid service.
The company's work is on the Goodfire research page.
