Models make up their minds
long before they say so.

What is the commitment boundary?

When a Large Reasoning Model solves a hard problem, it generates a Chain-of-Thought before writing its answer. In this monologue, the model hedges, tries alternatives and re-checks its statements. But is it really all necessary to produce the answer? Turns out that it is not, and that we can precisely locate the step at which the model becomes sure of the final answer.

We call this step i* the commitment boundary, falling typically around the midpoint of the CoT. Beyond this moment, the models often engage in what we defined epiphenomenal reasoning: on the surface, they seem to hedge, re-check, verify their guess, but they are not. The confidence of their guess remains unaltered.

How?

For each sentence in the reasoning trace, we truncate the CoT to the end of each sentence and force the model to commit to an answer. Comparing that answer to the one the full trace produces gives a step-by-step confidence signal about the moment in which the model becomes sure.

Before the commitment boundary, we get wrong or uncommitted answers. After the boundary (even if the model is writing "wait, let me re-check...") we get the same answer consistently at each truncation. The post-boundary tail of hedging and re-verification is epiphenomenal: it looks like deliberation, but it has no causal role in producing the answer.

We trained a small causal attention probe to recognize this moment online instead of post-hoc with over 90% detection accuracy on in-distribution data.

A real example: MATH-500, question 10

Models: gpt-oss-20b · gemma-4-26B-it · Qwen3-14B  |  Benchmarks: MATH-500 · AIME 2025 · GPQA Diamond · ZebraLogic

The trace below is one of sixteen independent reasoning chains sampled from gpt-oss-20b on this question. Each point is one sentence; its y-position is the model's actual confidence in its final answer if reasoning were stopped there, measured by truncating the trace and forcing an answer.

This trace has four distinct mid-guesses: first 20 (step 20, very brief), then 200 (steps 14–15, 24), then 2000 (step 25), and finally 2200 (step 38), each held with high confidence before being revised. Along the reasoning trace, the model correctly eliminates each candidate (due to wrong digit count, not divisible by 3, ...) and at step 39 commits to 2220, where P(final) jumps from 0.36 to 0.995 in a single step.

The commitment boundary falls at step 39 out of 73 — 53% into the trace. The four mid-guesses before it are genuine: each is held with high P(wrong) while P(final) stays near zero, and each gets explicitly refuted in the text. After step 39, P(wrong) collapses to near zero and stays there for all 34 remaining sentences.

What we found

• confidence along the reasoning trace is strongly bimodalalmost every step shows near-zero confidence (such as in the no CoT baseline) or near one (like at the end of the trace); the commitment boundary is thus a sudden transition at step i*, in correspondence of which the answer snaps to the correct one. This spike of confidence increment is on average 4.6× larger than the second-largest step improvement.

  Figure — normalized step confidence distribution (gpt-oss-20b · MATH-500)

  

• model family matters a lotgemma-4-26B commits after only 13–23% of its reasoning tokens; gpt-oss-20b takes 43–68%; Qwen3-14B falls in between.

  Figure — commitment boundary position i* across models and datasets

  

• models produce mid-guesses before the commitment boundary i*they exhibit high-confidence on incorrect intermediate guesses that later get revised. In the running example, the model seriously considers 22020 as potential answer before ruling it out.

• mid-guesses are not noisethe same problem explored in independent traces produces the same wrong guesses at the same relative positions. Traces with more mid-guesses are also systematically longer, with median trace length increasing monotonically with the number of distinct hypotheses explored. Across MATH-500 as a whole, independently sampled traces for the same problem share mid-guesses with a median Jaccard similarity of 0.71, and those mid-guesses appear at nearly identical relative positions (median standard deviation 0.005, versus a random-permutation baseline of 0.011). The model has a characteristic exploration structure for each problem — it is not generating freely.

• post-boundary reasoning is merely epiphenomenalwe randomly corrupted numeric values in AIME 2025 traces on either side of the boundary. At 20% corruption, pre-boundary perturbations reduced answer survival to 61%; post-boundary perturbations left it at 95%. At 50% the gap persists: 27% vs. 72%. The hedging terms ("wait", "let me check", "but actually") appear at exactly the same rate before and after i*. Surface uncertainty does not indicate causal uncertainty.

• the boundary is detectable from activations alonea small causal attention probe trained on hidden states at a single transformer layer classifies each step as no-guess, mid-guess, or final-guess with over 90% detection accuracy on in-distribution data, generalizing to all three out-of-distribution benchmarks without retraining. Used as an online exit signal, it consistently outperforms fixed-percentage truncation at every operating point.

Paper and citation

Beyond the Commitment Boundary: Probing Epiphenomenal Chain-of-Thought in Large Reasoning Models. Daniel Scalena, Sara Candussio, Luca Bortolussi, Elisabetta Fersini, Malvina Nissim, Gabriele Sarti. 2026.

@inproceedings{scalena2026commitment,
  title     = {Beyond the Commitment Boundary: Probing Epiphenomenal
               Chain-of-Thought in Large Reasoning Models},
  author    = {Scalena, Daniel and Candussio, Sara and Bortolussi, Luca
               and Fersini, Elisabetta and Nissim, Malvina and Sarti, Gabriele},
  year      = {2026}
}