LLMs often write their reasoning after the decision. Here's what that means.

Chain-of-thought prompting became one of the most widely adopted techniques in applied AI because the intuition behind it is simple and compelling: show your work. If you ask a model to reason step by step before answering, the displayed thinking should produce a better answer and explain how the model got there.

A paper published in June 2026 by researchers at the University of Groningen challenges that intuition at its foundation. Their finding: language models reach a commitment boundary — a precise moment where the answer is locked in — before the reasoning chain finishes. Everything generated after that point does not causally influence the output. The model has already decided. What follows is text that reads like continued reasoning but does not change the answer.

They call this epiphenomenal chain-of-thought. And across graduate-level benchmarks and math competition problems, they found models can be shortened by up to 55% — cutting away reasoning generated after commitment — with negligible impact on accuracy.

The implications run in two directions. One is about cost. The other is about trust.

What the paper found

The researchers studied what they call extended thinking or large reasoning models: systems like GPT-o1 and DeepSeek-R1 that generate long internal reasoning chains before producing a final answer. These models are increasingly common in deployments where careful reasoning is expected to matter, from legal analysis to scientific problem-solving to financial modeling.

The core question: does the reasoning in these chains actually cause the final answer, or does the model commit to an answer at some point and then continue generating reasoning as continuation text?

To answer this, they used two complementary methods.

The first is early exit analysis. They systematically truncated reasoning chains at different points — at 10%, 20%, 30% of the chain, and so on — and measured whether removing the tail changed the output. If reasoning after a certain point is truly causal, removing it should change answers. If the model has already committed, removing it should not.

The second is attention probe analysis. They trained lightweight classifiers on the model's internal attention states to detect when the model's internal representation of the final answer stopped changing. This gives a view inside the model, tracking not what it says but what it is building toward internally.

The two methods converged on the same finding: there is a sharp transition point — the commitment boundary — where the model's internal state locks. After this transition, the attention state no longer shifts, and truncating the remaining reasoning leaves outputs unchanged.

What epiphenomenal means here

The word epiphenomenal comes from philosophy of mind. An epiphenomenon accompanies an event without causing it. The classic example: your feeling of pain does not cause you to pull your hand away from a hot stove; the neural reflex does. The pain is real — you experience it — but it is not in the causal chain.

Applied to LLM reasoning: the text generated after the commitment boundary is real. The model produced it. It reads like thinking. But it did not drive the decision, because the decision was already locked in when the text began to be generated.

This is different from saying the reasoning is wrong, or that chain-of-thought is useless. Reasoning generated before the commitment boundary may well be causal. Early steps likely do shape internal representations and narrow the answer space. The problem is that what is generated before and after the commitment looks identical from the outside. There is no visible signal in the text itself that says "this step mattered" versus "this step was decorative."

How the 55% figure works

Across GPQA (graduate-level questions in biology, chemistry, physics) and AIME25 (a math competition benchmark), the researchers found that models can be shortened by 55% on average before accuracy degrades meaningfully.

This means the commitment boundary typically arrives before the halfway point of the reasoning chain. The model spends more tokens reasoning after it has decided than before.

The 55% figure is an average across tasks and model families. It varies. On simpler tasks, the commitment arrives earlier — sometimes within the first quarter of the reasoning chain. On genuinely hard problems, the commitment arrives later and less sharply. The key result is not a single precise number but a consistent pattern: in large reasoning models, a substantial portion of generated tokens fall after the answer is already committed.

The practical reading: if a reasoning model generates 1,000 tokens to answer a question, roughly 550 of those tokens may be produced after the answer is locked. They cost the same to generate as the tokens that mattered. They add latency. They may not add accuracy.

The cost implication

Reasoning models are expensive. GPT-o1 and DeepSeek-R1 generate far more tokens per query than standard chat models, and token cost is proportional to the number of tokens generated, regardless of whether those tokens are causal.

If 55% of reasoning tokens fall after the commitment boundary, that is 55% of the reasoning budget that produces no accuracy benefit. For teams running reasoning models at volume — automated analysis, large-batch document review, high-frequency decision support — this represents a real and recoverable cost.

Early-exit strategies that detect the commitment boundary and halt generation have been explored before in the context of standard language models. This research suggests they deserve much more attention for reasoning models specifically, where the post-commitment tail is substantially longer.

The researchers' attention probes, which classify commitment from internal states, could in principle be operationalized as a stopping condition. The model generates until its internal state locks, then stops. No framework or model modifications required: just a lightweight classifier watching the attention states.

This does not require waiting for commercial providers to ship early-exit APIs. Teams running open models can implement this themselves. The inference overhead of the probe classifier was not reported as significant.

The trust implication

The cost story is about efficiency. The trust story is about something harder to recover from.

In regulated industries, a standard expectation for AI systems is that decisions must be explainable. The practical implementation of that expectation, in many deployments, is showing the model's reasoning chain. The assumption behind this is that the reasoning chain is a causal account: it shows why the model reached the conclusion it reached.

The commitment boundary research undermines that assumption directly. If much of the reasoning chain is generated after the answer is locked, the chain is not a causal account. It is a plausible story constructed around a decision that was made by a different mechanism.

This matters most in three deployment contexts:

In healthcare, AI systems making triage or diagnostic recommendations are often required to show reasoning so that clinicians can review and override. If the visible reasoning is post-hoc, clinician review is reviewing a rationalization, not a cause.

In lending and financial services, credit decisions supported by AI must be explainable under FCRA in the US and similar frameworks elsewhere. A reasoning trace that presents post-commitment text as an explanation may not meet the spirit of that requirement.

In legal AI, systems summarizing documents or flagging clauses increasingly show their reasoning to allow attorney review. If the flags were committed internally before the reasoning was written, the attorney's review of the reasoning may not locate where the actual decision happened.

None of this means reasoning model deployments are invalid in these contexts. It means the assumption that "we show the reasoning chain, therefore the system is explainable" needs to be revisited. The chain is necessary but may not be sufficient as a causal account.

What teams building with reasoning models should know

The research does not say chain-of-thought is useless. It says the relationship between displayed reasoning and causal reasoning is more complicated than assumed, and that the complication has both an engineering consequence and a governance consequence.

Three things worth doing differently:

Do not treat reasoning traces as causal accounts. Use them as supplementary context and as a check against obvious errors. Do not present them to end users or auditors as proof of how the decision was made. The displayed reasoning describes a path; it may not be the path that was actually taken internally.

Explore commitment-boundary-aware inference. If you run open models, consider implementing attention probes or early-exit classifiers. If you use APIs, monitor whether providers offer reasoning-length controls and test whether shortened outputs maintain accuracy for your use case. The 55% number suggests meaningful cost recovery is available.

Invest in pre-commitment interpretability. The researchers argue that monitoring internal states before the commitment boundary is more informative than analyzing the full reasoning chain. This is a harder technical problem than reading the chain, but it is the right problem. For high-stakes applications, the goal should be understanding what the model was building toward in its early reasoning steps, not reading what it wrote to fill the page after deciding.

Where this matters most

The finding is most consequential for teams using extended thinking models in deployments where:

• Accuracy is important and cost is a constraint — the 55% recovery is real

• Decisions must be auditable or explainable — the post-hoc reasoning concern is direct

• Errors carry real harm — knowing that the visible reasoning may not reflect the actual decision process changes how you should design human review

Tasks where chain-of-thought has always looked most impressive — long mathematical derivations, multi-step logic puzzles, formal proofs — are also the tasks where the commitment boundary arrives latest and the reasoning before commitment is most substantive. That is, the best-case scenarios for reasoning transparency are genuinely different from the average case. The worst scenarios, in terms of the commitment-to-reasoning ratio, may be common tasks where the model commits quickly and then generates text to justify.

The honest caveat

The methodology uses proxy measures. Early exit analysis and attention probes tell you about the behavioral and representational signatures of commitment, but not about the underlying computational mechanism. The researchers are not able to open the model and watch decisions form; they are inferring commitment from what changes and what does not.

Results vary across model families and task types. The 55% average is not a constant. On hard, genuinely novel problems, the commitment arrives later and the reasoning before commitment is likely more substantive. On simpler or more familiar tasks, the commitment may arrive even earlier.

The models tested — GPT-o1, DeepSeek-R1 — are large and frontier-class. Whether smaller or differently trained models show the same pattern is not confirmed. The commitment boundary may be more pronounced in models specifically trained to generate long reasoning chains, since these models have learned to fill the reasoning budget regardless of when the answer was determined.

The paper

"Beyond the Commitment Boundary: Probing Epiphenomenal Chain-of-Thought in Large Reasoning Models" is by Daniel Scalena, Sara Candussio, Luca Bortolussi, Elisabetta Fersini, Malvina Nissim, and Gabriele Sarti. Published June 2026, available at arxiv.org/abs/2606.13603. Code at github.com/DanielSc4/Commitment-Boundary.

The core finding is not that AI reasoning is broken. It is that the relationship between what a model says it is thinking and what is actually driving its outputs is less direct than most deployments assume. That gap is worth knowing about before you ship.