Goodfire's Forking Fast Slashes AI Reasoning Debug Costs by 100x

When a language model tackles a long reasoning problem, it doesn't follow one deterministic path. Sample the same prompt twice and you can get two very different chains of thought, and sometimes two different final answers. Figuring out which tokens along that path actually swing the outcome, so you can debug agents, audit safety failures, or trace hallucinations, has been prohibitively expensive. Goodfire's new paper, Forking Fast, claims to make that analysis roughly 100x cheaper.

The forking token problem

The underlying technique is Forking Paths Analysis (FPA), introduced in earlier work from the same lead author. Take a base reasoning rollout and, at each token, resample many alternative continuations to see how often the final answer changes. By systematically re-sampling completions at every position along a single base path, you can identify forking tokens where two trajectories diverge into different outcomes, sometimes at expected content words but also in unexpected places like "that" or "who".

That analysis is exactly what you want if you are trying to answer questions like "which reasoning step actually got the agent to this unsafe tool call?" The catch is cost. The original method often burns millions of tokens to analyze a single reasoning chain. Running it across a real agent transcript, let alone a fleet of them, is not practical.

Smoothing the noise instead of paying for signal

The Goodfire team's key observation is that most of the apparent chaos in low-sample FPA is just sampling variance rather than the model behaving erratically. With a small number of resamples, the uncertainty dynamics look jagged, but at hundreds of resamples they become smooth except at key forking points where sharp changes occur. Variation across rollouts is modeled well as multinomial sampling noise, with reconstruction error decaying in proportion to the square root of the number of samples collected.

Because the noise has a clean statistical shape, you can fit a model to it and denoise cheap, low-sample estimates instead of paying for expensive high-sample ones. The authors develop a statistical smoother that maps noisy low-sample rollout data onto something close to the high-sample distribution, cutting sampling costs by roughly two orders of magnitude.

What was tested, and on what

The evaluation is deliberately focused rather than sprawling. The approach was studied on Chain-of-Thought reasoning in Llama-3-8B-Instruct and the native reasoning model DeepSeek-R1-Distill-Llama-8B, solving problems from tinyMMLU, with nearly two billion tokens of reasoning data collected and analyzed. tinyMMLU is a 100-item subset of MMLU designed to efficiently estimate LLM performance while preserving the essence of the original benchmark, which makes it a reasonable stress test for a technique whose whole point is spending less compute.

Code and an interactive dashboard are on GitHub, so you can reproduce the smoothing model and try it against your own rollouts.

Why this matters for debugging agents

The practical implication is that FPA can plausibly move from a research curiosity to a routine debugging step in agent pipelines. Running it on real transcripts gives you a map of where a long chain of reasoning was actually decided, which is exactly the surface area you want to point other interpretability tools at.

• Post-hoc failure analysis: pinpoint the specific tokens where an agent's trajectory locked in a wrong or unsafe decision.
• Localizing intervention: narrow down where to apply steering vectors, probes, or activation edits instead of scanning entire transcripts.
• Evaluating reasoning models: distinguish confidently-wrong answers from ones where a single fork could have flipped the outcome.

The limits are worth noting. The paper focuses on 8B-scale open models and a small multiple-choice benchmark, so behavior on frontier-scale models or open-ended generation remains an open question. The smoothing assumes multinomial sampling noise, which is exactly the regime where things converge nicely, so genuinely chaotic or adversarial prompts may not fit as cleanly. FPA also still requires sampling many rollouts, just far fewer than before.

The agent-authored footnote

There's a second story buried in the paper. The research reported and the initial manuscript draft were produced by the autonomous LLM agent Silico under human direction. Silico is Goodfire's interpretability research agent, which launched publicly at $1,000 per month for individual researchers, with discounts and grants for AI safety and life sciences work. Goodfire's human lead on the project described the collaboration as roughly on the level of working with a first- or second-year PhD student.

Whether or not you buy that framing, the artifact is real: a novel methodological result, an open-source implementation, and an arXiv submission, largely produced by an agent operating over a research codebase. For anyone building or evaluating research agents, Forking Fast is worth reading twice, once for the method and once as a datapoint on what these systems can now ship.