Ask a language model a question, let it reason step by step, and it eventually writes something like "so the answer is…". Eric Bigelow, a researcher at the interpretability company Goodfire, has spent years asking what happens before that point, and where the choice is really made. His answer, given on The Cognitive Revolution podcast, published October 10, 2026: the decision "really happens during sampling."
Rerunning the reasoning at every word
A language model writes one token at a time. A token is a word or a piece of one. At each step the model produces a probability for every possible next token, and the system then draws one at random according to those odds. That draw is called sampling.
Bigelow's Forking Paths experiment exploits this. In Forking Paths in Neural Text Generation, published in December 2024, he and his co-authors took a GPT-3.5 model's step-by-step reasoning chains and stopped them at every token. From each stopping point they let the model finish the reasoning again and again, about 30 times, and recorded which final answer each run reached. The paper is more specific than that shorthand: from one base response, the team sampled 30 continuations for each alternative token the model gave at least a 5% chance. It covered seven tasks, from symbolic puzzles and math to question answering and constrained story writing. The method is expensive, costing roughly a million generated tokens for each response analyzed.
Bigelow said the most striking view comes when the results are laid out as a time series: how likely each final answer is at each point in the reasoning. Often the odds sat steady, say 50% for one answer and 30% for another. Then at some token "suddenly the distribution changes really dramatically," usually collapsing onto one final answer as the model, in his words, "decides" what to say. He put the word in scare quotes himself.
Some of these turning points were what a human judge might expect, such as the first time an answer is mentioned. Others were not. In one example from the paper, the reasoning wrote "kilowatt hours" followed by an open parenthesis before "kWh". Resample from exactly that point, Bigelow said, and runs where the next token was not an open parenthesis but "some other just arbitrary word" ended with a different final answer. The paper also reports that other reasoning traces stay stable or drift gradually rather than jumping.
A coin flip, then consistency
Bigelow draws a broader lesson. Even when models plan their responses, he said, what they say at each moment is "really determined by like the flip of a coin." He describes reasoning as a form of in-context learning: the model adapts to text it has already produced, much as it adapts to a prompt. Once a token is drawn, the model builds a self-consistent line of reasoning around it.
His example is the current year. If a model's reasoning early on assumes it is 2024 rather than 2026, its answer to a question such as who the current head of state is could change completely. In that case there is a ground truth, and the model is hallucinating. But Bigelow said many uses have no single right answer. When he asks a model what to do next in a research project or what his experimental results mean, he might want to see "this whole tree of possibilities that it could have generated," and what would have changed had it assumed A instead of B.
He initially thought this made models very unlike people. In his example, a person who misspoke and said "Billy fell into a hole" instead of "Billy fell into a spaceship" would correct themselves and carry on with the story; a language model might commit to the slip and tell a very different story. But he now sees forks in human life too. He cited research on rationalization in which people choose the stimulus they prefer from a set, the experimenter swaps it, and people then explain why they chose something they never picked.
Where the forks fall
The host, Nathan Labenz, pressed on the timing. A turning point that came only at the end, after all the reasoning, would be strange; so would one at the start, followed by many tokens of reasoning toward a foregone conclusion. Bigelow said both happen, along with forks in the middle. In the earlier study there were more near the beginning and end, varying by task, and "every input behavior can tell its own story."
Late collapses are common, he said, when the model simply does not know. Uncertainty holds steady until a single point where the model "becomes utterly certain," often exactly where it writes "the answer is." His example is the last-letter task, which asks a model to join the final letters of four words. According to the paper, the names Forrest, Juanito, Allan and Candice should produce "tone." The GPT-3.5 model struggled, sometimes even writing Python code to solve the puzzle, and the alternative paths at the answer token produced a scatter of letters drawn from the words. It was, Bigelow said, "just kind of making it up on the fly."
That work is now about two years old, and Bigelow said today's reasoning models behave somewhat differently. They often run a "linearized search," listing many possibilities in their reasoning before going back to pick one. In his forking-paths analyses of such models, the uncertainty curves are often smoother. But forks remain: when none of the listed options looks more promising, the model eventually says "the answer therefore is" something that could have been different.
Smooth averages, sharp cases
Part of Bigelow's inspiration came from research on grokking, in which a neural network suddenly starts to generalize after long training. He credited Neel Nanda with the idea that phase transitions are everywhere if you look closely at individual tasks, even when learning curves averaged over a large dataset look smooth. Bigelow argues in-context learning is similar: aggregate results can look "deceptively" smooth while a single story or problem shifts sharply after one token or sentence.
The grokking study by Nanda and colleagues shows the pattern in a controlled case: in small transformers learning modular addition, a sharp jump in test accuracy rested on circuits that had been forming gradually beneath it. That finding is limited to those arithmetic tasks.
Bigelow found a similar lesson in a simple prompt. He began the Forking Paths project by asking for a story about a boy named John. The stories looked different on the surface, but about half gave John a dog and an adventure, and many ended at a pond or waterfall, with either the dog or John saving the day. Only by resampling the same prompt many times, he argued, can researchers see how "bumpy or smooth" this landscape is.
What "uncertain" means here
Labenz asked what a belief is for such a system. Bigelow said it partly reduces to the probabilities of the tokens a model will generate, with nuance about how those change across inputs and during reasoning. His mental model comes from Bayesian cognitive science, the approach laid out in the textbook Bayesian Models of Cognition: a mind holds latent hypotheses with prior probabilities and reweights them as evidence arrives. A common criticism of that framework, Bigelow said, is that it is unclear how the brain implements such inference. With language models, "we don't have that excuse," because researchers can look inside and intervene. A 2026 study he co-wrote did this in a narrow setting, feeding small Llama base models sequences of numbers and finding curved internal structures that track the inferred mean and spread.
He stressed a caveat about the word "uncertainty" in the forking work. By the end of a reasoning chain, the model is nearly certain, because the answer follows from what it has already written. Saying a model is 50-50 between A and B at some point means only that resampling from there produces each answer half the time. It "doesn't mean that it really represents these full rollouts."
Whether models internally anticipate where they will end up is an open question. Bigelow said it has proven hard to decode outcome distributions from a model's hidden states, though that may be because the data are so expensive to collect. His current work suggests models can be decoded "to some extent" a few steps ahead, and that sampling can be guided that way. A 2025 paper he co-wrote points in that direction on a small scale: with a 3-billion-parameter Llama model and a handful of selected questions, simple probes on its internal activations predicted rollout distributions better than baselines, and steering grew less effective once the model had committed to an answer. Bigelow doubts a model knows the full story it might tell. He wonders whether it resembles a mathematician who can see "the shape of the proof" before working through the steps.
Cost is the practical obstacle. In the follow-up Forking Fast, Bigelow's team modeled the answer curves as smooth in most places and jumpy at a few points, so fewer samples are needed. The paper, which studied 100 multiple-choice questions on two 8-billion-parameter models and collected 1.77 billion tokens, reports that smoothing improves low-sample estimates but can blur forks, sometimes estimates the largest forks worse, and produces uncertainty ranges that are too narrow. Bigelow still considers reading hidden states the most efficient route, if it can be made to work.
Show users the fork
Do the sudden cliffs bother him? Labenz said they felt "weird and like wrong"; he usually experiences his own thinking as "gradually landing the plane." Bigelow said many people are unsettled at first. But if users want varied outputs and also want each line of reasoning to be coherent rather than nonsense, he thinks forking "might just be an inevitable property of having these two key ingredients."
He does think some of it could be trained away, especially sharp forks at tokens like that parenthesis. Those may reflect odd correlations in internet text, where writing "kilowatt hours (kWh" happens to line up with something semantic. Forks at points people can interpret would be better.
Mostly he wants forks surfaced to users. He recalled an older OpenAI interface, no longer available, where users could hover over a token to see the alternatives the model might have produced. He also pointed to Loom, an interface that keeps many continuations of a model's text as a navigable branching tree. He wants something similar built into standard chat tools, at the level of meaning rather than single words, so that a sharp or gradual "left turn" is visible. Today's chat interface, he said, has barely changed since the Eliza days of 50 years ago or more.
Some of the variety may already be shrinking. Bigelow said studies have found that outputs become less diverse the more models are post-trained. In his recent experiments with Qwen3.6-35B-A3B, an open-weight model he described as heavily trained with reinforcement learning, two-thirds of requested stories were about a clock maker. He added that a quick, small-scale test of Astra produced fairly different stories. He also regrets that the sampling parameters behind the Claude API and the latest GPT models are now hidden. He suspects the main reason is to prevent reverse engineering and distillation rather than a new sampling method, though he said he would be glad to learn otherwise.
Retire the parrot, keep the dice
Labenz noted that commentators had recently been writing a eulogy for the "stochastic parrot" meme, the claim that language models merely stitch together memorized text. Bigelow said the metaphor "should probably be put to rest." He likened it to the Chinese room argument and asked what must be inside that giant rulebook to make conversation possible. He pointed to work such as Othello-GPT, in which a transformer trained only on move sequences developed internal representations of the board that causally shape its moves. At some point, Bigelow argued, learning a world model is more efficient than learning a lookup table.
The stochastic part stays, in his view. Unless models halt whenever they are less than fully certain, they must sometimes flip the coin and follow a path. A model that always gives the same output for the same input would be "the most extreme version of mode collapse," with no diversity of perspective.
He thinks that holds even for looped transformers, which repeat their internal layers to reason in hidden activations before writing anything. Labenz and Bigelow discussed Astra as a model that may be bringing this approach into production. Bigelow has run forking-paths analyses on the open-source looped models he could find, including the Ouro series. After the hidden reasoning, as they write out their reasoning chains, the models are still uncertain between answers, and those answers still change at particular tokens.