acceptodds
Under review as a conference paper at ICLR 2027

When Does a Language Model Commit? Answer Preference Before Explicit Answers

Abstract

An answer can be predictable before it is generated even when the model's preference later changes. We measure when preference stabilizes using a language model's continuation probabilities over a specified finite answer set. We define commitment as the first high-margin preference for the eventual answer after which the preferred answer remains unchanged until a specified text event. Across 352 controlled arithmetic responses, preference stabilizes before an explicit judgment. Among 155 responses whose initial preference differs from the final answer, 136 stabilize at the token that completes the printed total or second comparison subtotal. In parity, the outcome word follows four or five tokens later. Initial agreement also differs from stability: 108 of 197 initially agreeing responses later reverse preference. Hidden-state readouts, matched interventions, and held-out stopping comparisons test how this preference is represented and used. Aligning preference with numerical output separates changes in the preferred answer from the time taken to express a judgment.

Primary area:
interpretability and explainable AI
Keywords:
Large Language Models, Interpretability, Chain-of-thought Reasoning, Early Stopping

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_faU9OpenReview ↗
Rating
3(+2)
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper looks at arithmetic tasks such as judging even/odd from adding 4 numbers. They focus on Qwen for the main analyses and activation based analyses. They use a method to define when the model has settled on an answer: i.e. log odds of the yes/no answer exceeds 2, and the model doesn't change its preference after this point. They show that this usually occurs at around the time the model writes the total of the sum, but tokens before it writes its final yes/no answer. They also show that hidden state probe (though not specifically at the commitment point) decode the yes/no answer near the final answer with high AUROC. They show that activation patching at points is sufficient to change the log odds in favour of the donor answer. They show a modest advantage for early stopping based on their method compared to a fixed prefix baseline. Their conclusions surround when commitment occurs in these kind of arithmetic procedures.

Critical strengths

S1: The authors define a procedure for detecting early commitment during a reasoning trace S2: The authors use a combination of analyses including hidden state probes and activation patching

Critical weaknesses

W1: I don't find the question the authors address sufficiently important or the contribution significant in terms of understanding what it means when the model has committed to an answer W2: the activation based analyses are not specifically related to the commitment point, so are somewhat disconnected from that part of the paper. W3: The paper relies on a single Qwen model for the activation based analyses and for the main coverage W4: Minor but I found the paper difficult to follow in terms of the method and needed to refer to the appendix.

Questions for authors

Q1: Why is it surprising that the model commits to the yes/no answer at a point when it has emitted the total answer of the sum? Why is this meaningful in terms of understanding the computations that the model is performing? Q2: Why aren't the activation based analyses specifically targeted at the commitment point? Why do the activation based analyses help us understand what is going on at the time of commitment? Q3: Why was only Qwen used for the activation based analyses? Q4: Can the authors please enhance the readability of the main text?

3 comments

Thank you for the direct questions. They point at real gaps in the submission, and the revision changes the paper on each of them. Changes are in blue in the revised PDF.

Q1 / W1. Why is it surprising that the model commits when it has written the total, and why is this meaningful?

It would be unsurprising if it were guaranteed. It is not, and the cases where it fails are the informative ones. The revision adds three observations.

First, writing the correct total does not settle the answer. SmolLM3-3B writes a correct total in 242 responses, and 44 of them end with the wrong yes/no verdict. A reader of the text sees a correct number followed by a wrong answer. The trajectory shows what happened. In 43 of the 44, the model preferred the correct answer when it finished the total, and the preference flipped at one later token, a wrong parity word such as "parity = odd" after "total = 144" (Section 5.5, Figure 3b, Table 57).

Second, whether commitment coincides with the number depends on how the output is formatted. For the same questions in two formats it coincides in 31/39 responses in one format and 1/39 in the other. In a 120B-parameter model the preference turns more than 100 tokens before the final answer, inside the reasoning channel, when the number is first written there (Section 5.2).

Third, the alternative that early-answer studies worry about, that the model knows its answer from the start and the reasoning is decoration, is not what happens here. One token before the number is written, the preference is correct in 56%, 50% and 59% of responses in the three models, which is near chance. At the number it is correct in 99.6% and 97.5% for the two small models, and in 100% one token later for the large one (Figure 3a). When we edit the hidden state so that a different number is written, the verdict follows the new number.

On why this matters, we see two reasons. It says what an early-answer lead contains. In the parity tasks, for responses whose preference changes, the lead before the answer word consists of the four or five tokens written after the number that already decides the answer, and not of a stretch of hidden computation. It is also useful. Returning SmolLM3's preference at the number instead of its final verdict raises held-out accuracy from 77/96 to 91/96, and from 55/64 to 63/64 on a second set that was used for nothing else, while generating 18% fewer tokens.

Q2 / W2. Why are the activation analyses not targeted at the commitment point, and what do they tell us about it?

You are right about the submission. Its patching was done near the answer and not at commitment. The revision adds interventions at the commitment token and at neighbouring tokens, in two model families (Section 5.4, Appendices Q and W).

We move the hidden state at that token toward the state of a response with the opposite answer, while keeping linear predictions of the input numbers and of text progress fixed. The control is a random edit of the same size with the same component along the linear answer readout. Applied at the commitment token, the first edit changes the final answer in 5/32 responses for Qwen3-4B and 5/20 for SmolLM3. The random edit changes 0/32 and 0/20, and leaves every continuation identical to the unedited response token for token.

This tells us two things about commitment. In Qwen, the answer is easy to redirect one token before commitment (20/32 responses), harder at commitment (5/32), and almost fixed one token after (1/32). The edited text explains why. Before the number is complete, the edit changes which number is written, and the verdict follows it. Once the number is on the page, the same edit rarely changes the answer. The observational finding, that preference settles at the number, therefore has a causal counterpart: the written number carries the decision. We also learn that the linear readout predicts the preference without being the lever that sets it, since moving the state only along the readout direction never changes the preferred answer.

We do not want to overstate this. The cohorts are small: each model alone gives p = 0.0625 at the commitment token, and 0.002 when the two are pooled. SmolLM3 does not show the same drop across commitment (3/20, 5/20, 4/20), and edit sizes differ somewhat between positions. The paper states these limits.

Q3 / W3. Why was only Qwen used for the activation analyses?

That was a limitation of the submission. SmolLM3-3B, from a different family, now has its own readouts, transfer tests and commitment-relative interventions, reported next to Qwen3-4B in Sections 5.3 to 5.5. The timing measurement now also covers MMLU-Pro and GPT-OSS-120B, in addition to the three further models and three multiple-choice benchmarks that were already in the submission (lines 207 to 210).

Q4 / W4. Readability of the main text.

We rewrote the abstract, the introduction, Sections 5.2, 5.4 and 5.5, the discussion and the conclusion in plainer language. The introduction now states the two measurement problems in words and ends with a list of contributions. Section 3 has a worked example that follows Figure 1 token by token: which strings are scored, why an early agreement at token 32 is not commitment, and where commitment falls. The method can now be followed without the appendix. A new Figure 3 shows the main result for three models in one panel, and the conclusion states three findings, each tied to one results section.

A note on the summary.

In case it helps the discussion, the hidden-state readouts are not evaluated by AUROC against the correct answer. They predict the model's own log odds, and we report the Pearson correlation with that quantity (lines 347 to 349 of the submission, Figure 4a in the revision). This matters for Q2, because a readout of the model's own preference can be compared directly with the effect of editing the state, which is what the new Section 5.4 does.

Dear Reviewer faU9,

Thank you for reading the response and the revision, and for updating your assessment. If any part of the revised main text is still hard to follow, please tell us which one and we will revise it.

One result arrived after our reply and concerns your Q1. On 200 new questions, with the rule fixed before any response was generated, SmolLM3 again prefers the correct answer when it writes the total and loses it at the parity word. 39 responses print a correct total and end with a wrong verdict, and the preference is correct at the total in 38 of them. Returning the preference at the total gives 190/200 against 159/200 (Appendix X).

Reviewer_8AxuOpenReview ↗
Rating
2
Confidence
4(+1)
Informativeness
1
Presentation
2
Soundness
1(−1)
Read review
Paper summary

This paper studies the timing when LLM fosters stable preferenes.

The authors defines the non-revertal timing as the commitment time, based on sequentially computing the log prob of the continuations to the prompt.

The core observation is the final answer predictibility & stablility are independent to the statement. As I absorb, before LLM generates the final answer, it might already have reached the crucial state; also, if the LLM comes up with the final answer, it might still be uncertain of it.

Critical strengths

Interesting discovery over the reasoning of LLMs. This work researches in the reasoning stablity and the correlation to the final answers. This potentially help the community better understand the reasoning mechanism and further help the LLM-RL community.

Critical weaknesses
  1. The conclusion section is hard to comphrehend. I suggest the authors make the conclusion part more direct and more related to the core body part.

  2. The experiments are mainly conducted on author's synthesized data, ARC challenge, CommonsenseQA and OpenBookQA. The correctness on these benchmarks are high and may potentially biased to test cases that are polluted or memorized. I suggest using HLE/ALE (just for example) that small LLMs cannot achieve half of the full scores to exclude the potential dataset bias.

Questions for authors
  1. Though I did not put this in the weakness, I wonder the whether the conclusions of this paper can be extrapolated to 10^2 B or larger scale LLMs? As nowadays the community would find that conclusions/biases are different in 10^0/1/2/3 B scale models.
3 comments

Thank you for the review and for the concrete suggestions. Each one led to a change in the paper. Changes are in blue in the revised PDF.

W1. The conclusion is hard to comprehend and should be more direct and closer to the body.

We rewrote it. The conclusion now states three findings, each tied to one results section. First, an early correct guess is not a settled answer: more than half of the responses that start with their final answer abandon it at least once (Section 5.1). Second, the preference turns to the final answer when the decisive number is first written, in models with 3B, 4B and over 100B parameters, while exact coincidence with our commitment time depends on the output format (Section 5.2). Third, this is useful: SmolLM3 has the right answer when it writes the number and loses it one word later, and reading the answer at the number raises held-out accuracy from 77/96 to 91/96 with 18% fewer tokens (Section 5.5).

The abstract and introduction were rewritten in the same way. The introduction now ends with a list of contributions, and Section 3 has a worked example that follows Figure 1 token by token.

One clarification on the summary. We do not claim that predictability, stability and the written statement are statistically independent. We show that they are different quantities that can come apart. The new SmolLM3 result is the clearest case: the answer the model prefers at the number and the answer it finally writes differ in 48 of 242 responses, and in 43 of these the preferred answer is the correct one.

W2. Benchmarks with high accuracy may be polluted or memorised; use tasks on which small models score below half.

We added MMLU-Pro (Section 5.2, Appendix R, Table 51). We chose it over HLE because HLE is gated and many of its questions are open-ended, whereas our measurement needs a fixed set of candidate answers. MMLU-Pro has up to ten options and meets your criterion: Qwen3-4B answers 18/48 and SmolLM3-3B 17/48 of the held-out questions, with every attempted question in the denominator. Commitment is still measurable there. 29 and 27 responses commit, 11 and 12 of them to a wrong final answer, so the measurement does not depend on easy or correct answers.

On memorisation we would add two points. Low accuracy alone does not show that a benchmark is unseen, so we do not rely on it for that. The stronger protection is that every arithmetic question in the new studies is generated programmatically from random operands (192 new questions per model, plus 64 more for the interventions). On these questions the answer is not recalled. One token before the total is written, the preference of the two small models is correct in 56% and 50% of responses. At the token that completes the total it is correct in 99.6% and 97.5%.

Question. Can the conclusions be extrapolated to models at the 100B scale or larger?

We ran the official GPT-OSS-120B checkpoint, a mixture-of-experts model with 116.8B total and 5.1B active parameters, on 64 parity and 64 modular-product questions (Section 5.2, Appendices S and V). Three conclusions carry over and one does not.

What carries over is the following. All 128 responses commit before the final verdict, on average 38 and 34 tokens earlier. An early correct guess is again unreliable: all 70 responses that start with their final answer abandon it later. And the preference turns to the final answer when the decisive number is first written. In parity, one token before the total first appears the preference equals the final answer in 38/64 responses, which is near chance. At the token that completes the total it does so in 52/64, and one token later in 64/64 (Figure 3a).

What changes is where this happens. This model reasons in a separate analysis channel, and the number first appears there, a median of 113 tokens before the final verdict. Measured against the formatted line in the final channel, commitment is therefore early and not simultaneous: 116 of 128 responses commit before that line and none exactly at it. The token-level coincidence that we reported for Qwen3-4B is thus specific to models and formats that write the number once. The underlying pattern, that preference follows the first written result, holds at this scale.

This is one model, and we do not claim a scaling law. The paper states this limit in Section 5.2 and in the discussion.

Dear Reviewer 8Axu,

One experiment finished after our reply, and it bears on your concern about dataset bias and on your question about other models. We posted the full result under the general response and summarise here what concerns your review.

Your W2 asked whether the results could come from polluted or memorised test cases. In the new check, a script first fixed the rule and then generated 200 new parity questions from random operands, excluding every operand combination used before. These questions did not exist before the run, so the result cannot come from memorised benchmark items. SmolLM3's final verdict is correct in 159 of the 200. The answer it prefers when the total is written is correct in 190 (exact p = 3.1 × 10^-6), which repeats the main result of Section 5.5 on new data (Appendix X).

Your question asked whether the conclusions hold at other scales. Next to the 120B model in our reply, the same run covers ten models from 1.5B to 4B parameters in seven families. In all six models that answer mostly correctly, the preference rises at the token that completes the total. The other four answer near chance, and their preference never settles. We report those four as plainly as the others (Appendix X).

The three changes that your review asked for are all in the revised PDF: a conclusion rewritten around three findings, MMLU-Pro as a benchmark on which both small models score below half (18/48 and 17/48), and the 120B model. We would be grateful to know whether they address your concerns, or what remains unclear, while the discussion is open.

Reviewer_bfz1OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper aims to investigate when a Large Language Model's preference for the final answer becomes stable during reasoning. The core observation is that the answer can already be read from the model's probability distribution before the model has explicitly stated the answer in words.

Critical strengths
  1. The experiments include strict reality checks to prevent overhyping the results. The authors included control tests to make sure they didn't exaggerate the observations.

  2. Preference is defined before fitting a probe, while commitment requires the complete subsequent winner trajectory. Such a observation is interesting.

Critical weaknesses
  1. While this paper helps us pinpoint exactly when the model's answer preference stabilizes, there is still very little evidence showing what new solutions or methods this discovery could actually inspire. The North Star of interpretability work should focus on how to derive effective empirical methods on proxy tasks from the mech analysis results [1]

  2. The main observation was only built on a very narrow set of math problems. The paper claims to pinpoint exactly when the model makes up its mind. However, they only strictly proved this using one specific model (Qwen3-4B) on math problems.

  3. The residual activation patching mentioned in the paper replaces the entire residual state, which may simultaneously contain various types of information such as the answer, question content, and the text progress.

  4. The tool used for readout is not verified to be generalizable. Linear probe only works if it is trained on the exact prompt format it is testing. When tested on an unseen prompt format, the performance might drop.

  5. Dynamic stopping does improve accuracy in at least one clean arithmetic condition, but the policy comparisons are exploratory, reuse an examined test split, and exclude candidate-scoring cost. These results support the authors' careful distinction between preference timing and practical stopping, but leave the benefit of dynamic monitoring under realistic compute budgets unclear.

[1] A Pragmatic Vision for Interpretability, https://www.alignmentforum.org/posts/StENzDcD3kpfGJssR/a-pragmatic-vision-for-interpretability

Questions for authors

Please check my weakness section.

3 comments

Thank you for the careful review and for noting the control tests. We answer each weakness below. Changes are in blue in the revised PDF.

W1. Little evidence of what new methods this discovery could inspire.

We accept this criterion. The essay you cite asks that an interpretability claim be validated by what it lets one predict or do on an objective task, and the revision adds such a test (Section 5.5, Figure 3, Appendix U).

The preference trajectory shows where a model's errors enter. SmolLM3-3B writes a correct total in 242 of 256 responses, yet 44 of these end with a wrong verdict. In 43 of the 44, the model's own yes/no preference is correct at the moment the total is completed. It stays correct until the model writes a parity word, which is the wrong one in 43 cases ("parity = odd" after "total = 144"). After that word the preference is wrong in all 44. The measurement therefore places the failure at one generated token, after the model already preferred the right answer.

This gives a rule that needs to know where the result line is, but not how the number maps to the answer. When the total line closes, return the answer the model currently prefers. On the held-out TEST set it answers 91/96 questions correctly against 77/96 for full generation (exact p = 0.003) and generates 17.9% fewer tokens. On 64 further questions that played no part in any choice it gives 63/64 against 55/64 (p = 0.02).

We want to be clear about the status of this result. The rule (the event rule with margin 0) was in our predefined policy set, but the frozen selection did not choose it for SmolLM3. The selection criterion allowed at most 5% disagreement with the full-generation answer, so it rejected a rule for correcting that answer: on DEV the rule disagreed on 4 of 32 questions and was right on all 4. We therefore label the SmolLM3 result post hoc, and the 64 untouched questions are its check. For Qwen3-4B, margin 0 was the frozen choice from the start.

W5. Stopping comparisons were exploratory, reused an examined test split, and excluded candidate-scoring cost.

That description of the original experiment is accurate, and the submission said so (lines 398 and 424 to 425). The new experiment uses separate TRAIN (64), DEV (32) and TEST (96) questions, freezes all rules on DEV, and measures wall-clock time including tokenisation, generation, candidate scoring, readout, parsing and synchronisation (Section 5.5, Appendix N, Tables 33 to 40).

For Qwen3-4B the event rule keeps the accuracy of full generation (95/96 on TEST, 31/32 on the timed subset) and is 1.150 times faster [1.142, 1.160] with exact scoring and 1.181 times [1.171, 1.192] with a linear readout. The same table shows where monitoring does not pay. Scoring at every step saves 5.45% of tokens in SmolLM3 but is slower overall (0.892 times), because it needs about six scoring calls per request. Checking once, at the event the trajectory analysis points to, is what makes the rule cheap. We did not time the margin-0 rule separately for SmolLM3. It generates the same tokens as the printed-total rule, which runs 1.181 times faster, and adds one scoring call.

W2. The main observation rests on a narrow set of math problems and one model.

The submission did measure timing on three further models (Qwen2.5-3B, Qwen2.5-7B, GLM-4-9B) and three multiple-choice benchmarks (lines 207 to 210). You are right, though, that the central alignment result and all activation analyses rested on Qwen3-4B and arithmetic. The revision adds three things (Section 5.2, Appendices R, S and V). SmolLM3-3B, a different family, has its own trajectories, readouts and interventions. MMLU-Pro adds harder questions on which both small models are below half (18/48 and 17/48). There, 29 and 27 responses commit, including 11 and 12 with a wrong final answer, so commitment is not limited to easy or correct answers. GPT-OSS-120B adds a model with 116.8B total parameters.

In all three models the same thing happens when the total is first written (Figure 3a). One token earlier the preference is correct in 56%, 50% and 59% of responses. At the token that completes the total it is correct in 241/242 responses for Qwen and 236/242 for SmolLM3, and for GPT-OSS in 64/64 one token later. In GPT-OSS this happens in the analysis channel, a median of 113 tokens before the final verdict.

What does not carry over is the exact coincidence between our commitment time and the number. It depends on the output format (31/39 against 1/39 for two formats of the same questions), and the abstract now says so. We also do not claim to pinpoint when the model makes up its mind internally. Commitment is defined on output probabilities for stated answer strings (Section 3), and the tasks remain controlled arithmetic plus one harder benchmark.

W3. The residual patch replaces the entire state, which mixes the answer, the question content and text progress.

We agree for the original experiment. The new intervention (Section 5.4, Appendices Q and W) addresses this in two ways. First, the edit toward the opposite-answer donor is projected so that fitted linear predictions of the four input numbers, their sum and three measures of text progress stay unchanged. Second, the control is a random edit with the same norm and the same component along the answer-readout direction, instead of an arbitrary donor.

At the commitment token the donor edit changes the final answer in 5/32 Qwen and 5/20 SmolLM3 responses. The matched random edit changes 0/32 and 0/20 (exact p = 0.002 pooled over the two models, 0.0625 for each alone). Across the three tested positions the counts are 38/156 against 0/156, and all 156 random-edit continuations are identical to the unedited response token for token. The edited text shows the route: in 31 of the 38 changes the edit altered the number the model wrote, and the verdict followed the new number.

Two limits remain. The projection keeps fitted linear predictions fixed, which is weaker than removing all information about the question or progress, and the fits for the input numbers are weak (Tables 49 and 50). The cohorts are also small.

W4. The readout is not verified to generalise and may drop on an unseen prompt format.

It does drop, and the submission reported it: correlation with the model's log odds fell from 0.948 within a condition to 0.445 when that prompt was left out of training (lines 353 to 358). The commitment measurement itself does not depend on the readout, because preference is defined from output probabilities before any probe is fitted.

The revision adds a format that is excluded from training altogether, in both families (Section 5.3, Appendix O, Table 41). Training on a broader set of other conditions raises correlation on the unseen format from 0.462 to 0.584 for Qwen and from 0.155 to 0.535 for SmolLM3, with paired gains of 0.122 [0.098, 0.148] and 0.380 [0.225, 0.519]. Transfer is still partial, and for SmolLM3 the broader training lowers correlation on the original format from 0.903 to 0.854. We report both.

Ethics flag.

The review selects "Flag for Ethics Review: Yes" without a stated reason, so we are not sure what to address. The study uses public models and public or programmatically generated questions. It involves no human subjects and no personal data, and our use of AI tools is disclosed in the AI Use Statement and in the submission form. If there is a specific concern, we would be glad to answer it.

Dear Reviewer bfz1,

One experiment finished after our reply, and it bears on two of your points. We posted the full result under the general response and summarise here what concerns W1 and W5.

On W1, we had labelled the SmolLM3 accuracy gain post hoc, because the rule was not the one that our frozen selection picked. We have now tested it prospectively. The rule was fixed in the script before any response was generated, and it ran on 200 new questions that share no operand combination with the earlier ones. Returning the answer that the model prefers when the total line closes gives 190/200 against 159/200 for full generation (exact p = 3.1 × 10^-6), with 18% fewer tokens (Appendix X).

On W5, our reply said that we had not timed this rule separately for SmolLM3. We have now done so. In the new run the scoring call recomputed the whole prefix instead of reusing the generation cache, and the rule ran at 0.996 times the speed of full generation. In that implementation it saves tokens but not time. With the cached scoring of the paper's timing experiment we estimate about 1.15 times, and Appendix X says that this combination was not timed directly.

The same run covers ten models from seven families. The preference rises at the written total in all six models that answer mostly correctly. Two of them settle only at the parity word, and there the rule costs accuracy, so the Discussion now says that the reading point has to come from each model's trajectory.

We would be grateful to know whether our replies address your concerns. We would also like to know whether the ethics flag reflects a specific issue, so that we can answer it before the discussion closes.

Comments on the paper

2 comments

We thank the three reviewers for their time. The reviews agree on three things that the submission did not do well enough. It did not show what the finding is good for, it rested on one model and easy arithmetic, and its activation analyses were not placed at the commitment point. We ran new experiments on each and have uploaded a revised PDF. New or rewritten text is in blue, the main text is 10 pages, and Appendices M to W are new. This note summarises the evidence that more than one reviewer asked about, and the individual replies give the details.

First, on what the finding is good for (bfz1 W1, faU9 W1 and Q1, 8Axu W1). The preference trajectory shows where a model's errors enter. SmolLM3-3B writes a correct total in 242 of 256 responses, and 44 of these end with a wrong verdict. In 43 of the 44, the model's own preference was correct when it finished writing the total. It turned wrong at one later token, a wrong parity word such as "parity = odd" after "total = 144". If we read the answer that the model prefers at the total, instead of the verdict it writes at the end, we get 91/96 on the held-out TEST set against 77/96 for full generation, and 63/64 against 55/64 on 64 further questions that were used for no choice, with 18% fewer tokens. On Qwen3-4B the same rule keeps full accuracy and is 1.15 times faster with all monitoring cost included (Section 5.5, Figure 3, Appendix U).

Second, on scope (bfz1 W2, 8Axu W2 and question, faU9 W3). SmolLM3-3B, a second model family, now has its own trajectories, readouts and interventions. MMLU-Pro adds questions on which both small models score below half (18/48 and 17/48). The official GPT-OSS-120B checkpoint adds a model with 116.8B total parameters. In these three models the preference turns to the correct answer when the total is first written (Figure 3a). One token earlier it is correct in 56%, 50% and 59% of responses. At the token that completes the total it is correct in 241/242 responses for Qwen and 236/242 for SmolLM3, and for GPT-OSS in 64/64 one token later. In GPT-OSS this turn happens inside the reasoning channel, a median of 113 tokens before the final verdict (Section 5.2, Appendices R, S and V).

Third, on activation analyses at the commitment point (bfz1 W3, faU9 W2 and Q2). We now edit the hidden state at the commitment token and at its neighbours, in both small model families. Moving the state toward that of a response with the opposite answer changes the final answer in 10 of 52 responses when applied at commitment. A random edit with the same size and the same component along the linear answer readout changes none, and it leaves all 156 continuations identical to the unedited response token for token. In 31 of the 38 answer changes across the tested positions, the edit changed the number that the model wrote, and the verdict followed the new number (Section 5.4, Appendices Q and W).

We also report two limits in the paper as plainly as the positive results. The exact coincidence between our commitment time and the printed number depends on the output format. For the same questions in two formats it holds in 31/39 responses in one and 1/39 in the other, and in GPT-OSS the preference turns in the reasoning channel, long before the formatted answer line. What is stable in the three models is that the preference is correct once the number has been written. The other limit concerns the SmolLM3 accuracy gain. It comes from a rule that was in our predefined set but was not the one our frozen selection picked, because the selection criterion required agreement with the full-generation answer and therefore rejected a rule for correcting that answer. We label this result post hoc, and we checked it on the 64 questions that were not used for anything else.

One question is still open, and we prefer to say so now. We do not yet know how common the turn at the number is in other model families. In a first small check, a 1B Llama model answered all 24 questions correctly while its preference was still at chance when the total was written. That model writes out the parity computation before it answers, so the event at which the preference turns appears to be model-specific, and the trajectory is the tool that locates it. We are now running the fixed rule on 200 new questions and on further model families, and we will report the outcome here whatever it is.

The manuscript changed as follows. The abstract, introduction and conclusion are rewritten, and the introduction ends with a list of contributions. Section 3 has a worked example that follows Figure 1 token by token. Sections 5.2, 5.4 and 5.5 report the new experiments, and a new Figure 3 shows the main result for three models. The cost figure moved to Appendix N to stay within the page limit. The Supplementary Material contains the code and the per-response trajectories behind every new number.

In our general response we said that we were running the rule on new questions and on more model families, and that we would report the outcome whatever it was. The run is complete. It is now Appendix X of the revised PDF (pages 91 to 94), with short updates to the abstract, Section 5.5, the Discussion and the Conclusion. The script, the 200 questions, every response and all candidate scores are in the updated Supplementary Material.

One script fixed the rule and every definition before any new response was generated. It created 200 new parity questions, excluding every operand combination used before. It then ran ten public models in FP32 on one A100 GPU: the two models of the paper at the same revisions, and eight further instruction-tuned models, for seven families in total (Llama-3.2-3B, Phi-4-mini, Qwen2.5-3B, Qwen2.5-1.5B, Falcon3-3B, Gemma-2-2B, Granite-3.3-2B and SmolLM2-1.7B).

First, the main use replicates. On the new questions SmolLM3's final verdict is correct in 159 of 200. The answer it prefers when the total line closes is correct in 190 of 200, a gain of 15.5 points [9.5, 21.5], with 38 questions corrected, 7 lost (exact p = 3.1 × 10^-6) and 18% fewer tokens. The errors enter where the paper says they do. 39 responses print a correct total and end with a wrong verdict; the preference is correct at the total in 38 of them, and the written parity word is wrong in 38. Qwen3-4B answers 198 of 200 correctly with or without the rule, and its preference goes from 113/196 correct one token before the total to 194/196 at it.

Second, a note on timing. In this run the rule's single scoring call recomputed the whole prefix instead of reusing the generation cache. That call takes about 0.3 s and cancels the time saved by stopping (0.996 times for SmolLM3, 1.026 times for Qwen), while the printed-total rule ran at 1.183 and 1.187 times, as in the paper. In the paper's timing experiment, which reuses the cache, one scoring call costs about 0.07 s. That would put the rule near 1.15 times for SmolLM3, but we have not timed this combination directly. The accuracy gain does not depend on it: on the 64 timed questions the rule answered 62 correctly against 49.

Third, the turn at the written total is common but not universal, and the new data show where it holds. Six of the ten models give mostly correct final verdicts (79.5% to 99%). In all six the preference rises at the token that completes the total, from 47% to 67% correct one token earlier to 77.5% to 99% at that token. In four of them (SmolLM3, Qwen3-4B, Llama-3.2-3B and Phi-4-mini) it is correct in at least 95% of responses there, as in the paper. In the two Qwen2.5 models it reaches 84.5% and 77.5% and completes once the parity word is written. The other four models answer near chance (49% to 59.5%). They print correct totals but do not turn them into correct verdicts: Gemma-2 and SmolLM2, for example, write "parity = even" in 175 and 196 of 200 responses. Their preference stays near chance, because there is no correct answer to settle on. Across all ten models, once the parity line is written the preference equals the final verdict in all 1,985 responses with an accepted total.

This also settles the 1B Llama case from our earlier note. That model writes the parity as a calculation ("parity = 183 % 2 = 1"). The 3B Llama in this run writes the parity word, and its preference turns at the total (194/200).

Fourth, reading at the total is not a general way to raise accuracy, and the Discussion now says so. It helps where the preference is right at the total and the written word goes wrong afterwards, as in SmolLM3 and also SmolLM2 (122 against 98). It costs accuracy where the preference settles only at the parity word, most clearly in Qwen2.5-1.5B (155 against 196). What transfers is the measurement: for each model it shows where the answer is settled and where errors enter. As a simple illustration, defined after this run, we chose the reading point on the first 64 questions and tested it on the remaining 136. The choice never lowered accuracy in any of the ten models, gained 18 questions in SmolLM3 and 14 in SmolLM2, and saved 5% to 18% of tokens.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.