Learn · Intermediate

Monte Carlo rollouts: estimating whether a reasoning step actually changes the answer

Monte Carlo rollouts estimate whether a reasoning step actually changes the chance of getting the right answer by sampling many possible continuations. They matter because a chain of thought can look lucid and important while contributing little to the final result.

A rollout is simply one played-out future. In a board game, start from the current position, let plausible moves unfold and record whether the game is won. Repeat this many times. The fraction of wins estimates how promising the position was. That is the Monte Carlo idea: use many random samples to estimate a quantity that would be too difficult to calculate exactly.

For language-model reasoning, the “position” can be a partial written solution. Suppose a model says, first, “Let x be the number of red balls,” then performs algebra. Is the first sentence truly doing work, or is it a plausible-looking label that the model could replace without changing its odds of solving the problem? A researcher can sample many continuations after that sentence and compare them with continuations after an altered or omitted version. If the expected reward changes, the sentence has measured advantage.

The recent Legibility is Not Interpretability paper uses this basic idea to study chain-of-thought evaluation. It defines a step’s importance as its advantage: the change in expected reward from including the step, estimated through Monte Carlo rollouts. It then asks whether LLM judges can identify the same important steps from the text alone. The answer is uncomfortable. Strong judges beat a prevalence baseline but remain well below a noise ceiling, especially for correct answers.

This is a lesson about LLM-as-a-judge, not an indictment of written explanations. A judge can identify an obvious mistake in a wrong derivation, much as a teacher can spot a line where arithmetic fails. It is harder to distinguish the one consequential move in a correct-looking solution from a fluent but redundant sentence. Surface legibility is therefore not the same as causal importance.

The word “advantage” can be misleading if it sounds like a moral judgment. It is only a counterfactual estimate. How much better do expected outcomes become when this piece of context is present? A high positive advantage means the rollouts succeeded more often with the step; a near-zero advantage means it did not materially alter success under the experiment’s setup. A negative advantage means the step made the continuation less likely to succeed.

There are several ways to construct the comparison, and each has tradeoffs. One can delete a step, replace it with a neutral placeholder, swap in another plausible step, or begin continuations before and after the step. Deletion may make text unnatural; replacement may introduce a different signal. The right intervention depends on the question. The important discipline is to state it explicitly rather than treating a model’s prose as a transparent window into computation.

Monte Carlo methods also carry uncertainty. A small number of samples can make a lucky path look important. A model’s sampling temperature, the reward function, stopping rule and prompt all affect the estimate. The paper itself notes that its labels are rollout estimates and therefore contain noise. Rollouts can be expensive too: evaluating every sentence in a long trace may require thousands of generated continuations.

Still, the method gives researchers a valuable upgrade over “this sentence sounds central.” It treats reasoning evaluation as an experiment. If changing a step does not change outcomes, the step may be commentary, decoration or a downstream reflection of a decision made elsewhere. If it does change outcomes, the researcher has evidence of functional importance—even if not a complete mechanistic explanation.

The broader lesson is practical: explanations should be tested like components. Readability is useful for people, but causal relevance needs intervention and measurement. Monte Carlo rollouts are one of the clearest ways to make that distinction when the system is stochastic and the full space of possible futures is too large to enumerate.

Ground Truth. (2026, September 7). Monte Carlo rollouts: estimating whether a reasoning step actually changes the answer. Ground Truth. https://groundtruth.day/learn/monte-carlo-rollouts-for-evaluating-reasoning.html

@misc{groundtruth:monte-carlo-rollouts-for-evaluating-reasoning,
  title  = {Monte Carlo rollouts: estimating whether a reasoning step actually changes the answer},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {sep},
  url    = {https://groundtruth.day/learn/monte-carlo-rollouts-for-evaluating-reasoning.html}
}

Topics: evaluation · reasoning · monte-carlo · chain-of-thought · interpretability