acceptodds
Under review as a conference paper at ICLR 2027

Correct Answers, Invalid Traces: What Verifiable Grade-School Math Reveals About Chain-of-Thought Traces

Abstract

Chain-of-thought traces are widely read as records of how models reach their answers, informing debugging, agent auditing, and claims about reasoning. Testing this interpretation is difficult because natural-language thinking traces are rarely mechanically verifiable. We revisit it in iGSM, a synthetic grade-school mathematics benchmark designed to study thinking traces and used to support claims of learned reasoning and planning. Crucially, iGSM exposes the exact quantities and dependencies that a correct solution should use, allowing generated traces to be checked programmatically step by step and enabling us to test whether correct answers are reliably accompanied by valid traces. We first evaluate models trained exclusively on valid, minimal traces. Answer correctness and trace validity nearly coincide in distribution but decouple out of distribution: on the hardest instances, 31.6% of correct answers have invalid traces, over half of which pass all syntactic and arithmetic checks but fail semantic dependency checks. We then intervene on trace supervision. Non-minimal training traces induce non-minimal outputs, while re-asking the same problem with a different query reveals computations inherited from the original query, weakening minimality as evidence of selective planning. Shuffling tokens in 10% of training trace sentences preserves near-clean accuracy even out of distribution despite no trace passing verification. Swapped training traces likewise retain high in-distribution accuracy. We discuss the implications of these findings for chain-of-thought monitoring and interpretation in the context of AI safety.

Primary area:
interpretability and explainable AI
Keywords:
chain-of-thought, reasoning traces, trace validity, faithfulness, iGSM, synthetic grade-school math, out-of-distribution generalization

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_rZHoOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

This paper looks at whether a correct answer from a model actually comes with a valid chain-of-thought trace. To test this, the authors use iGSM, a synthetic grade-school math benchmark from Ye et al. (2025a). Because every quantity and dependency in an iGSM problem is known, each step of a generated trace can be checked automatically.

They retrain the small GPT-2-style models from Ye et al. on problems up to 15 operations and test them on harder problems up to 23. On problems like the training ones, correct answers almost always come with valid traces. On harder problems the two drift apart: at 23 operations, about a third of correct answers (31.6%) have invalid traces, and more than half of those look fine on the surface and only fail when the dependencies are checked.

The second part of the paper changes what the models are trained on. Models trained on longer, non-minimal traces produce non-minimal traces themselves. When a problem is re-asked about a different quantity, some traces still include steps that only the original question needed, which the authors take as a sign that minimal traces are not enough to show the model is planning. Models trained on traces from unrelated problems still get about 82% of in-distribution problems right, and models trained with the first part of each trace scrambled do as well as the clean model, even though none of their traces pass the checker.

The results come from 4,096 test problems per difficulty level, mostly from one training run, with extra runs and seeds in the appendix, a no-trace baseline, and fifteen worked example traces. The authors present this as a warning for anyone reading model traces as evidence of reasoning or using them for safety monitoring, while noting they do not test whether traces are causally faithful.

Critical strengths
  1. The setting makes the question checkable. Most work on CoT faithfulness has to judge natural-language traces by hand or with another model. Here every step can be checked exactly, and the paper reports the full counts (Tables 8 and 9), so the reader can see which failures drive the numbers. Testing this on iGSM, the benchmark originally used to argue for learned reasoning and planning, makes the result more relevant than it would be on a purpose-built task.

  2. Some of the interventions are well designed. The shuffled-prefix models are the clearest case: with 30% of trace sentences scrambled, the model still reaches about 75% at op = 15, against about 25% with no trace at all, so the trace is clearly doing useful work while never being valid. The re-asking experiment is also a nice idea, since in standard iGSM the question always targets the last node, and it can't be told apart from simply computing toward the end. These are the parts of the paper I found most convincing.

Critical weaknesses
  1. The clean model is much weaker than the one in Ye et al. out of distribution. At op = 23, Ye et al. report 76.8% with a valid trace, the model here gets 39.7% (Table 10). With Ye et al.'s numbers, 31.6% of correct answers having invalid traces is not even possible, so part of the effect may simply be undertraining rather than distribution shift. The 200k-step run also lowers this share at op = 20-22, which doesn't fit with the claim that doubling training doesn't change the picture.

  2. The 31.6% looks less striking once broken down (Table 9, op = 23). Almost half of these traces (352/750) don't parse, so they are not plausible-looking at all. Most of the remaining ones use a wrong input that has the same value as the right one, or a wrong value that later gets multiplied by zero, so the answer comes out right anyway. The checker also seems strict in some cases: in the running example, using A instead of Q counts as an error, even though the problem says Q equals A. It would help to know how many of these cases have answer 0 or are just this kind of substitution.

Questions for authors
  1. Can you report the main analysis on a model closer to Ye et al.'s performance (their checkpoints, or a longer/better-tuned run)? If the share of correct answers with invalid traces drops a lot there, the distribution-shift story becomes much weaker.

  2. Of the 750 correct-but-invalid cases at op = 23, how many have answer 0, and how many same-value substitutions are definitional (like Q = A in the running example) rather than coincidental?

  3. For the shuffled models, does the intact suffix pass the checker if the shuffled prefix is treated as given? This would show whether "no trace passes verification" reflects the reasoning or just the scrambled block.

  4. The swapped models score below the no-trace baseline in distribution (81.7% vs 87.6%, Table 2). How does this fit with presenting them as evidence that invalid traces still help? Similarly, with 3.7% and 9.6% of re-asked traces padded (Table 5), most stay minimal, so why conclude that minimality fails?

Suggestion (future work): training with queries that do not always target the sink would separate a learned "compute toward the sink" habit from a lack of planning more cleanly than re-asking.

4 comments
AuthorsOpenReview ↗

Response to Reviewer rZHo (part 1 of 2): W1 and W2

We thank the reviewer for appreciating our contributions including that the setting makes every step of a trace checkable and reports the full counts, and the design of the shuffled-prefix and re-asking experiments. An overview of the new experiments is in our global comment. Our response is in two parts, W1 and W2 here, and Q1 to Q4 in part 2.

W1. The gap between Ye et al.'s and our models in accuracy.

RESPONSE: Our clean model was trained with Ye et al.'s published recipe, and we could not close the gap despite repeated training runs. Doubling the training to 200k steps raises the share of correct answers with a valid trace at op 23 from 39.7 to 41.3%, and a 4x larger Qwen3-architecture model trained with the same recipe reaches 45.5%. The 200k run moves the share of correct answers with invalid traces by 2 to 5 points at op 20 to 23 and leaves the pattern unchanged.

The main question is whether this decoupling of correct answers and trace validity is an artifact of our training. Ye et al. did not release their checkpoints, so we could not examine them. However, Zhang, Neubig and Yue (ICML 2026) have released checkpoints of 100M parameter Qwen2.5 style models that were pre-trained from scratch for 10B tokens on GSM-Infinite, a close variant of iGSM with the same step format, together with a trace validity checker. We evaluated their two pre-trained models, trained on op=2 to 10 and on op=2 to 20, on their test problems using their checker. The table gives, per op, their answer accuracy and P(trace valid | correct). Averaged over op as in their paper, our evaluation reproduces their Figures 3 and 12.

op910111213141516171820
trained on op 2 to 1095.6 / 97.692.9 / 97.074.6 / 83.926.5 / 74.13.1 / 0
trained on op 2 to 2098.9 / 97.398.3 / 99.497.0 / 95.197.9 / 93.595.0 / 88.289.1 / 68.071.7 / 50.881.5 / 61.064.6 / 56.657.4 / 48.549.7 / 54.1

Where accuracy is above 95, P(valid | correct) is 88 to 99. Where accuracy falls to between 50 and 90, it drops to between 48 and 68, including inside the op=2 to 20 model's own training range. Models trained by others, on a related benchmark and scored with their own checker, show the same decoupling, so it is not an artifact of our training.

References: Zhang, C., Neubig, G. and Yue, X. (2026). On the Interplay of Pre-Training, Mid-Training, and RL on Reasoning Language Models. ICML 2026.

W2. The op 23 breakdown.

RESPONSE: We agree the headline needs the breakdown next to it, and we expand Table 9 further. At op 23 there are 750 correct answers with an invalid trace.

Case when Answer is 0. 300 of the 750 have answer 0, against 22.7% of all correct answers. Removing every problem whose answer is 0 leaves 450 of 1,836 correct answers with an invalid trace i.e., 24.5%.

Same-value substitutions. Of the 234 traces in which a step uses a wrong input with the same value as the right one, 3 use a quantity the problem defines as equal to the right one, 171 use an unrelated quantity that happens to have the same value, 109 of them a shared 0, and 60 differ in more than one input. The case of the running example is therefore rare.

Further, we do not agree that such a substitution is valid. That the problem defines Q as equal to A does not make A a substitute for Q. The step has to use the quantity that the problem's sentence names, and using another quantity that happens to have the same value is a different computation. Moreover, the training traces never make such a substitution, so a model that makes one at inference has deviated from its training.

Traces that do not parse. These traces are not malformed text. In 338 of the 352 a letter is bound twice, 194 at a Define statement, 127 as a clause-local symbol and 17 both, and the other 14 use a letter that was never defined. They read like valid traces, which is why we count them as invalid rather than as noise. One trace at op 23, for instance, defines Parrot Paradise's Shark as w at its second step, with w = 6, then at step 14 writes "Define Trout's Bone as w; so w = Z = 7." and two steps later computes the answer as "K = E * w = 7 * 7 = 3". The inputs and the arithmetic are right and the answer is correct, but w names two quantities, so the trace does not parse.

Counting only failures that no value coincidence can explain leaves 397 of 2,376 correct answers at op 23, i.e. 16.7%. In distribution, at op 15, 0.5% of correct answers have an invalid trace under the paper's full count. The strictest count out of distribution is therefore still more than thirty times the loosest count in distribution.

Finally, note that the cases this strict count excludes are not noise. A trace in which a step uses a wrong input of the same value, or whose error a multiplication by zero removes, reaches the right answer by coincidence and is still semantically invalid.

AuthorsOpenReview ↗

Response to Reviewer rZHo (part 2 of 2): Q1 to Q4

Q3. Suffix validity of the shuffled models.

RESPONSE: Yes, for the 10% shuffled model the intact suffix passes the checker for most correct answers. We split each generated trace at the earliest point from which every sentence is well formed, bound every letter the suffix uses but never defines to any quantity of the problem with that value, trying all bindings, and run the full dependency, arithmetic and query check on the suffix. Share of correct answers whose suffix passes:

op1520212223
clean model, whole trace valid99.592.587.578.668.4
10% shuffled, suffix valid given the block99.594.891.187.181.6
30% shuffled, suffix valid given the block82.357.948.240.334.5

So the reviewer's reading is correct. For the 10% shuffled model the shuffled block is almost always a single sentence, and what follows it is a valid derivation. The model writes the shuffled block and then a suffix. The suffix is valid at least as often as the clean model's whole trace at every level, and more often from op 20 onwards, 81.6% against 68.4% at op 23. The corruption stays in the block and does not spread into the derivation that follows. The suffix has exactly one letter whose definition was lost in the shuffled block in 2,454 of the 2,463 correct answers at op 23, so only one letter has to be matched to a quantity and the rest of the suffix is checked exactly. For the 30% model the block is about three sentences, accuracy falls, and the suffix is valid for only a third of correct answers at op=23. "No trace passes verification" therefore reflects the shuffled block, not the reasoning. We will add this analysis to the appendix.

Q4. Swapped traces and re-asking.

RESPONSE: The reviewer is right that 81.7% for the swapped traces model below 87.6% for the no-traces model does not show that "invalid traces help". The sentence in the introduction that a trace can help the model while being far from a valid derivation refers to the shuffled models, where the 30% shuffled model reaches 75% at op=15 against 25% with no trace. The swapped result shows something different, that a trace can be unrelated to the problem and the answer still be right, since accuracy survives when every training trace belongs to another problem. We will make the attribution to the shuffled models explicit.

For re-asking, 3.7% and 9.6% are indeed small and we do not claim most traces are padded. Our point is that the rate on the original queries is 0.4% and 1.3%; i.e. re-asking increases padding by 7 to 9 times. Further, nine in ten of the extra parameters lie on the original chain to the sink: 114 of 128 at op=15 and 142 of 157 at op=20. Parameters on that chain are only 24 to 28% of the unnecessary parameters available in these problems. If the model chose its unnecessary parameters at random, about one in four of them would be on that chain. In this case, nine in ten of the ones the model adds are on that chain. On the standard query the parameters the question needs are exactly the chain to the sink, so a minimal trace there cannot show that the question selected the parameters. The probing experiment run for Reviewer y8kb also supports our claim. At the end of a re-asked question, the probe on the clean model's hidden state still marks as necessary one in ten of the parameters on the original chain that the new question does not need. This is against 0.3% for unrelated ones, and when the model does write an unnecessary parameter, it is one the probe had already marked as necessary, in 65%, 50% and 58% of cases at op=15, 20 and 23.

Q1. Checkpoints.

RESPONSE: Please see our answer to weakness 1. Ye et al. have not released any checkpoints. We did evaluate the model checkpoints of Zhang, Neubig and Yue (2026), that show the same decoupling wherever their accuracy drops.

Q2. Answer-0 and definitional substitutions.

RESPONSE: Answered under weakness 2. Of the 750 correct answers with an invalid trace at op 23, 300 have answer 0, and 3 of the 234 same-value substitutions are definitional.

Suggestion (future work).

RESPONSE: Thanks for this suggestion, we will investigate this further.

We hope the new evidence on models trained by others, the recount under the reviewer's own accounting, and the validity of the shuffled models' traces after the shuffled block address the concerns behind the review.

Thank you for the detailed response and the new experiments.

My two main concerns are mostly answered. The Zhang et al. checkpoints show the decoupling is not only from your training run, and the op 23 breakdown is what I asked for.

Two things I would still fix in the abstract. It gives 31.6% alone, but the strict count is 16.7%. And for the 10% shuffled model the rest of the trace is valid in most cases (81.6% at op 23), so “no trace passes” comes from one shuffled sentence.

I am raising Soundness and my rating to 3.

Thank you for raising the score and the comments on the abstract. We will incorporate them in the next version of the paper.

Reviewer_WW1DOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies whether chain-of-thought (CoT) traces genuinely constitute valid derivations of a model’s final answers. Using the programmatically verifiable iGSM benchmark, the authors separate answer correctness from trace validity. They show that, especially under distribution shift, models can produce correct answers while emitting traces that violate the problem’s dependency structure. Experiments with swapped, shuffled, and non-minimal traces further demonstrate that answer correctness, trace validity, and trace minimality can come apart.

Critical strengths
  1. The paper turns the difficult question of whether a CoT is “correct” into a mechanically testable problem involving dependency structure, ordering, and arithmetic constraints.

  2. The swapped-trace, shuffled-trace, non-minimal-trace, and re-asked-query experiments probe different aspects of trace semantics and minimality. The results provide consistent evidence that correct answers and valid traces can become decoupled.

  3. The paper distinguishes syntactic/arithmetic failures from semantic dependency failures and shows how invalid traces become more common as reasoning depth moves outside the training distribution.

Critical weaknesses
  1. The experiments primarily use a 124M-parameter GPT-2-style model on a synthetic grade-school mathematics benchmark. It is unclear whether the same effects hold for stronger and more recent reasoning models. The authors should consider evaluating newer open-weight models, such as Qwen3.5, or models with substantially different scales and architectures.

  2. The verifier mainly checks whether the model follows the dependency graph specified by the generator, rather than whether it provides any mathematically equivalent valid derivation. A model may produce the correct answer using an alternative intermediate variable, an equivalent algebraic transformation, or a different but valid computation order, yet still be marked invalid if it does not align with the prescribed dependency structure. The paper does not explain how such cases should be handled. The authors should clarify whether the checker accepts equivalent solutions and, if not, acknowledge that the reported validity measures specification-graph consistency rather than mathematical derivation validity in the broader sense

Questions for authors

Same as Weakness

3 comments
AuthorsOpenReview ↗

Response to Reviewer WW1D

We thank the reviewer for appreciating our contributions including that trace validity is made mechanically testable, that the swapped, shuffled, non-minimal and re-asked experiments give consistent evidence of decoupling, and the separation of syntactic from semantic failures. An overview of the new experiments is in our global comment.

Weakness 1. Larger and more recent models.

RESPONSE: We did not evaluate the Qwen3.5 pretrained model, for the following reasons.

Evaluating pretrained open-weight models directly has three problems. We have no control over their training data, so we cannot say which problems are in or out of distribution. They write long free-form reasoning even at small op and often get the answer wrong. And the iGSM checker cannot parse that reasoning, so trace validity cannot be scored. There is also a reason of principle: the claims we test, Ye et al.'s and our own, are about models trained from scratch on iGSM, so a larger architecture trained the same way is the right test of whether they hold at scale.

So we trained the same recipe from scratch on the Qwen3 0.6B architecture, which has 492M parameters, which is 4x the GPT-2 model size, with the identical vocabulary. It has a clean arm and a 10% shuffled arm, both evaluated on the same problems as the paper. Greedy evaluation on 4,096 problems per level, reported as accuracy / P(valid | correct).

op1520212223
clean model, GPT-2 124M98.8 / 99.583.3 / 92.577.8 / 87.567.8 / 78.658.0 / 68.4
clean model, Qwen3 492M99.0 / 99.788.5 / 91.482.7 / 84.675.1 / 76.068.4 / 66.5
10% shuffled, Qwen3 492M99.1 / 088.6 / 083.4 / 074.5 / 067.1 / 0

While the larger model is more accurate at every level, its trace validity percentage decays in the same way as the GPT-2 model. So the decoupling is not a property of the 124M model. The 10% shuffled arm matches the clean arm within noise while writing no valid trace at any of these levels, as for GPT-2.

Weakness 2. What the checker accepts.

RESPONSE: Trace validity in the paper means consistency with the problem's dependency graph, as the reviewer says. In iGSM, that means that the checker does not reject any valid derivation. The checker accepts any computation order, since any topological order of the necessary quantities passes. It accepts any algebraic form of a step, since the letters that name quantities can be chosen freely and a multi-input sum or product may be split into intermediate steps in any grouping. It does not need to handle alternative intermediate variables, because each quantity in an iGSM problem is defined by exactly one sentence, so the only valid way to compute a quantity is from the quantities that the sentence names. A step that uses any other quantity is not an equivalent derivation but a different computation that happens to give the same value.

The checker is lenient in one known direction, since it does not check which operation combines the parameters a step uses, and Example 9 in the appendix shows a trace that passes with a wrong answer. Any leniency in the checker raises the measured validity, so the invalid-trace shares we report are lower bounds.

We hope the larger-model replication and the clarification of what the checker accepts address the reviewer's concerns.

Reviewer_WW1DOpenReview ↗

Official Review

Thanks for your response. I will keep my positive score.

Thank you.

Reviewer_y8kbOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper revisits iGSM (Ye et al., 2025a), a synthetic grade-school math benchmark in which every generated chain-of-thought trace can be checked step by step against the problem's dependency graph. The authors use this to separate two questions that are usually conflated: whether the final answer is correct, and whether the trace is a valid derivation of that answer. They train GPT-2-sized models following the iGSM-med recipe and report four findings: (1) A model trained only on valid, minimal traces produces correct answers that almost always come with valid traces in distribution. Out of distribution the two come apart: at op = 23, 31.6% of correct answers come with invalid traces, and about half of those pass every syntactic and arithmetic check. (2) Training on non-minimal traces makes the model's outputs non-minimal too. Re-asking the same problem about a different quantity shows that the model computes parameters leading to the original final quantity even when the new query doesn't need them. This weakens the claim that minimal outputs are evidence of planning. (3) Models trained on traces with 10% of the prefix sentences shuffled match the clean model's accuracy, even out of distribution, while never producing a valid trace. (4) Models trained on traces taken from unrelated problems still answer 82% of in-distribution problems correctly. All of these findings suggest that the validity of CoT and the correctness of the final answer are decoupled.

Critical strengths
  1. The main contribution of this work is that it builds on top of the iGSM paper and exposes the non-causality between the generated CoT and the predicted final answer. Through a wide range of intervention like swapping, shuffling, and providing non-minimal CoT, this work provides multiple evidence that all suggests that not all sentences within a generated CoT are driven by the goal of finding the answer.

  2. Because iGSM traces can be checked exactly, the paper can break failures down precisely (Table 3, Table 9). It separates parse failures and reference errors from semantic failures, where a step uses the wrong inputs. This is more convincing than studies that judge traces only by their effect on the answer.

  3. The re-asking and widened-world experiments are a clever way to separate "selecting the parameters the question needs" from "computing toward the sink of the dependency graph," two explanations that the standard iGSM evaluation cannot tell apart.

  4. In the age of LLMs, I really appreciate the authors' effort to make this paper readable for a human reviewer like me. The AI-use statement says AI tools helped with drafting and editing, but the text clearly shows human revision and judgment. The sentences are concise and the language is professional. That is unlike other submissions I have reviewed, which are full of very long sentences or casual, unprofessional language, typical signs of text produced by an LLM with little human intervention.

Critical weaknesses
  1. The abstract does not define its key terms. I cannot tell from the abstract what a minimal trace and a non-minimal trace are, or what "selective planning" means. Minimality is only defined in Section 3 (lines 226–228), and "selective planning" is never defined anywhere in the paper. The abstract tries to summarize every experiment, which leaves no room to explain the concepts the reader needs. Please define these terms, or cut some experimental detail from the abstract to make room.

  2. The abstract, introduction and experiments section present the experiments in different orders. The abstract says "We first evaluate models trained exclusively on valid, minimal traces. Answer correctness and trace validity nearly coincide in distribution but decouple out of distribution." However, Section 4 starts with swapped traces (Table 1) and shuffled traces (Table 2). The experiment on valid minimal traces and the discussion of how correctness and validity decouple out of distribution only appear at lines 314–366. The three places also disagree with each other:

    • abstract: clean → non-minimal/re-ask → shuffled → swapped
    • introduction: clean → non-minimal/re-ask → swapped → shuffled
    • Section 4: swapped → shuffled → clean → non-minimal → re-ask
  3. The introduction contains too much general background. Lines 037–073 cover the history of chain-of-thought prompting, RL-trained reasoning models, multi-agent systems, and the METR investigation and its press coverage. This is how CoT is used and interpreted in the LLM community, and most of it is not needed to motivate this work. As a result, the introduction runs past page 2.5. Please condense lines 037–073 into at most two paragraphs. The paragraph at lines 074–079, which states the assumption the paper tests, should stay. The "Our Results" part (lines 081–135) also largely repeats the abstract and could be shortened.

  4. The paper diagnoses the problem but offers little toward a fix. My biggest concern is that the work focuses only on exposing the problem. The authors already have an exact step-level checker, so they could at least take a first step, for example:

    • training with a reward or auxiliary loss based on trace validity, and checking whether out-of-distribution trace validity improves;
    • testing whether the checker, used as a monitor, catches the out-of-distribution failures and at what cost in accuracy.
  5. The analysis is not deep enough to explain what drives the model's behavior. The paper designs experiments showing that answer correctness and trace validity come apart, but it does not try to explain what makes the model write a given sentence in its trace. For the experiments that control the minimality of the training traces, a highly relevant result from the original iGSM paper is mentioned only in passing (lines 087–088, 143–144) and never discussed. Ye et al. (2025a) probe the hidden state at the end of the question and find that the model already encodes whether each parameter is needed to compute the final answer. So the model probably relies on this encoded "plan" to produce minimal traces. When the training traces are made non-minimal by adding unnecessary computations, the model no longer needs this plan, because it can simply solve every parameter in topological order. An experiment that would add depth is to probe the hidden state at the end of the question and test whether the model still encodes the full set of necessary parameters:

    • for the models trained on non-minimal traces;
    • most importantly, for the re-asked questions. Ye et al.'s probe was trained only on standard questions, where "needed for the query" and "on the path to the sink" are the same set. Probing re-asked questions would show directly which set the model encodes, and whether the padded traces happen even though the correct set is encoded. This would settle the planning question that Section 4 leaves open.
Questions for authors

See weakness above. Please consider addressing the Weakness 4 & 5 and I'm giving my initial score of 3 assuming these issues will be properly addressed during rebuttal. I believe this work is generally well written and can be improved to be a conference-ready paper.

4 comments
AuthorsOpenReview ↗

Response to Reviewer y8kb (part 1 of 2): Weakness 5

We thank the reviewer for the careful reading and for highlighting the exact step-level breakdown, the re-ask and widened-world experiments, and the readability of the paper. An overview of the new experiments is in our global comment. Our response is in two parts, weakness 5 here, and weaknesses 4, 1, 2 and 3 in part 2.

Weakness 5. Hidden-state probe on non-minimal models and re-asked questions.

RESPONSE: We ran the probing experiment the reviewer mentioned. Following Ye et al. (2025a), we freeze the model, add a rank-8 update on the input embedding and a linear head at the end of the question, and train the probe to predict whether a parameter is necessary. The probe is trained on standard questions and tested on the op=15, 20 and 23 test problems, first with their original question and then with the question re-asked as in Table 5 in the paper.

The table reports the percentage of necessary parameters the probe calls necessary / the percentage of unnecessary parameters it calls unnecessary. A parameter counts as necessary when the probe's score for necessary exceeds its score for unnecessary, with no threshold tuned. Probes are trained with Ye et al.'s recipe, 100,000 steps on a fresh stream of problems.

cleannon-minimal90% non-minimaluntrained network
op 1598 / 9994 / 9496 / 9971 / 76
op 2094 / 9990 / 9593 / 9971 / 80
op 2389 / 9988 / 9691 / 9972 / 83

The clean model encodes the necessary set well above the untrained network. The model trained only on non-minimal traces encodes it less well, as the reviewer anticipates, but only by 1 to 5 points. The model trained with one minimal trace in ten is within 2 points of the clean model.

On re-asked questions, every parameter is in one of the four groups: either necessary for the new question, on the original chain but no longer needed, dependent on the asked quantity, or unrelated. The table below reports, for the clean model, the share of parameters in each group that the probe calls necessary. The ideal values are 100, 0, 0 and 0.

op 15op 20op 23
probe trained on standard questions
necessary for the new question929190
on the original chain, no longer needed121010
dependent on the asked quantity0.00.10.1
unrelated0.30.30.2
probe trained on re-asked questions
on the original chain, no longer needed556
unrelated111

The model marks 90% to 92% of the parameters the new question needs as necessary, but it also marks as necessary one in ten of the parameters on the original chain that the new question no longer needs, against 0.3% of unrelated parameters. A probe trained on re-asked questions still marks 5% to 6% of those original-chain parameters as necessary, against about 1% of unrelated ones. So the path to the sink is part of what the model encodes at the end of the question, and this is not an artifact of the probe.

Finally, this explains the padded traces of Table 5 in the paper. There, re-asking makes the clean model compute parameters the new question does not need, and nine in ten of those parameters lie on the original chain to the sink. We took those traces and, for each unnecessary parameter the model computed, checked what the probe had marked on the model's state at the end of the question, before any step was written. In 65%, 50% and 58% of cases at op=15, 20 and 23 respectively, the probe had already marked that parameter as necessary, against 12%, 10% and 10% for all original-chain parameters. So the extra steps are not due to a correct plan written out carelessly. On a re-asked question the model's plan still includes parameters from the original chain to the sink. This answers the planning question the reviewer raised.

AuthorsOpenReview ↗

Response to Reviewer y8kb (part 2 of 2): Weaknesses 4, 1, 2 and 3

Weakness 4. RL with a validity-gated reward, and the checker as a monitor.

RESPONSE: We ran both experiments the reviewer suggested, RL with the checker in the reward, and the checker as a monitor. For the RL experiment, we train the clean model with GRPO on op=16 to 20 problems with an outcome reward of answer correctness alone. We also run a second RL variant in which, following the process-gated reward of Zhang et al. (2026), the reward is answer correctness gated by checker validity. We use a standard GRPO configuration, 256 prompts with 16 rollouts each per step, 140 steps over 10,240 problems. We also run a control in which the clean model continues pretraining on the same 10,240 problems with their reference traces, for the same number of passes. We use greedy evaluation on 4,096 problems per level, reported as accuracy / P(valid trace | correct answer).

op20232528
clean model, before RL83.3 / 92.558.0 / 68.438.7 / 40.922.0 / 7.6
outcome reward92.7 / 96.680.3 / 80.963.7 / 60.937.6 / 25.0
validity-gated reward92.2 / 98.579.7 / 93.961.7 / 84.138.0 / 45.2
continued pretraining on the same problems90.7 / 97.577.1 / 90.360.1 / 76.238.3 / 44.4

Accuracy improves consistently and by the same amount under both RL settings. Validity improves far more under the validity-gated reward. The continued-pretraining control, last row, nearly matches the accuracy of the outcome reward with higher validity. At the same time, validity still decays with distance from the training range (45.2 at op=28 under the gated reward). Moreover, the gated reward requires a verifiable trace, which doesn't apply in case of the swapped models. That gate exists only because iGSM traces can be checked mechanically; for natural-language traces there is no checker to gate on.

When the checker is used as a monitor, it accepts an answer only if its trace is valid. Among the answers it accepts, 99.5, 96.0 and 93.7% are correct at op=15, 20 and 23, against 98.8, 83.3 and 58.0% of all answers. This however lowers the accuracy over all problems by 0.5, 6.3 and 18.3 points respectively at these op levels. The accuracy it loses is exactly the correct answers with invalid traces, the quantity the paper measures.

References: Zhang, C., Neubig, G. and Yue, X. (2026). On the Interplay of Pre-Training, Mid-Training, and RL on Reasoning Language Models. ICML 2026.

Weakness 1. Terms undefined in the abstract.

RESPONSE: We will revise the abstract, define the terms there and cut the experiment list to make room. By selective planning we mean choosing which parameters to compute from the question, as opposed to computing toward the end of the dependency graph. The term will also be defined at first use in the text.

Weakness 2. Experiment order.

RESPONSE: We will use the same order in the abstract, the introduction and Section 4, namely swapped, shuffled, clean, then non-minimal and re-ask.

Weakness 3. Length of the introduction.

RESPONSE: We will condense lines 037 to 073 to two paragraphs, keep the paragraph at lines 074 to 079, and shorten "Our Results" as suggested.

We hope the experiments reported under weaknesses 4 and 5 address them. We will upload a revised PDF with these experiments and the changes to the abstract and introduction before the end of the discussion period.

Reviewer_y8kbOpenReview ↗

Thank you for your response

Most of my concerns are addressed by the rebuttal. As I have written in my review "I'm giving my initial score of 3 assuming these issues will be properly addressed during rebuttal", and I believe the revised draft is standing up to this expected state. Therefore I'm maintaining my score of 3 with a confidence of 4.

Thank you. We are glad the revised draft has addressed your concerns.

Comments on the paper

3 comments
AuthorsOpenReview ↗

Global Response: Summary of New Experiments and Analyses

We thank all reviewers for their constructive feedback. In response, we conducted several new experiments and analyses. We summarize the key concerns raised, the reviewer threads, and our responses.

  1. RL training and checker-based monitoring (Reviewer y8kb, Weakness 4). The reviewer asked whether trace verification could be used to improve reasoning through training. We compared outcome-only RL, validity-gated RL, and continued pretraining. Validity-gated RL substantially improves trace validity at comparable answer accuracy. Continued pretraining on the same problems reaches similar accuracy with higher validity than the outcome reward. Under both rewards, validity still decays with distance from the training range. The gated reward also needs a checkable trace, which the swapped models do not provide. Used as a monitor, the checker accepts only answers with valid traces. At op=23, 93.7% of the answers it accepts are correct, against 58.0% of all answers, at a cost of 18.3 points of overall accuracy.
  2. Hidden-state planning probes (Reviewer y8kb, Weakness 5). The reviewer asked whether models trained on non-minimal traces, and models given re-asked questions, internally identify which parameters are necessary for answering the question. We conducted hidden-state probing experiments following Ye et al. (2025). The probes reveal a persistent bias toward the original dependency chain. Moreover, 50 to 65% of the unnecessary parameters appearing in padded traces were already marked as necessary by the probe before generation.
  3. Larger-model replication (Reviewer WW1D, Weakness 1). The reviewer questioned whether our findings generalize beyond the 124M-parameter GPT-2 model to larger, more modern architectures. We trained a 492M-parameter Qwen3-architecture model from scratch, with clean and shuffled-trace variants. The larger model reproduces the decoupling between answer correctness and trace validity.
  4. Independent-model replication (Reviewer rZHo, Weakness 1). The reviewer observed that our clean model underperforms the original iGSM model of Ye et al. and questioned whether the decoupling could be an artifact of undertraining. In response, we evaluated independently trained models released by Zhang et al. (ICML 2026), based on the closely related GSM-Infinite benchmark. These models still exhibit the same decoupling.
  5. Robustness of invalid-trace counts (Reviewer rZHo, Weakness 2). The reviewer questioned whether the high rate of correct answers with invalid traces was mainly due to parsing failures, zero answers, or coincidental numerical equalities. We separately analyzed these cases. Even under a conservative counting that excludes failures explainable by coincidental values, 16.7% of correct answers at op=23 have invalid traces, compared with only 0.5% in-distribution under the normal counting of failures. Moreover, the parsing failures were not due to mangled text, but actual syntactic or semantic mistakes in the traces produced by the model.
  6. Validity of shuffled-trace suffixes (Reviewer rZHo, Question 3). The reviewer asked whether the apparently invalid shuffled traces contain valid derivations after the shuffled prefix. We checked these suffixes, allowing consistent assignments to undefined symbols. For the 10% shuffled model, 81.6% of correct answers at op 23 have suffixes that pass this conditional check. Thus, the corruption remains confined to the initial shuffled block and doesn't affect the subsequent derivation.

We will upload a revised PDF with all the changes before the end of the discussion period.

AuthorsOpenReview ↗

Revised PDF uploaded

We have uploaded the revised PDF. All new and changed text is in blue.

  • Section 5 (new, main text): GRPO with an outcome reward and with a validity-gated reward, the continued-pretraining control, and the checker as an inference-time monitor (Reviewer y8kb, Weakness 4).
  • Appendix D (new): D.1 validity of the suffix after the shuffled block (Reviewer rZHo, Q3); D.2 finer breakdown of the correct answers with invalid traces, including the strict count (Reviewer rZHo, W2 and Q2); D.3 linear probes of the hidden state on standard and re-asked questions (Reviewer y8kb, Weakness 5; Reviewer rZHo, Q4); D.4 the pretrained models of Zhang et al. (2026) scored with their own checker (Reviewer rZHo, W1 and Q1); D.5 the same recipe on a four times larger architecture (Reviewer WW1D, Weakness 1; Reviewer rZHo, W1).
  • Abstract and introduction: terms defined, one order of experiments throughout, introduction shortened (Reviewer y8kb, Weaknesses 1 to 3); the two findings on swapped and shuffled traces attributed separately (Reviewer rZHo, Q4); "no valid trace" statements qualified to the whole trace (Reviewer rZHo, Q3).
  • Section 3, "Checking a trace": the role of the dependency check when two quantities share a value, and that the checker accepts any valid derivation (Reviewer WW1D, Weakness 2; Reviewer rZHo, W2 and Q2).
  • Appendix B: pointers to the new experiments and the 200k-step run (Reviewer rZHo, W1).
AuthorsOpenReview ↗

Summary of the discussion period

As the discussion period closes, we would like to thank the reviewers for their insightful comments, which have strengthened our paper.

Reviewer engagement and scores. The reviews highlighted several strengths of the paper. Reviewer rZHo noted, "Testing this on iGSM, the benchmark originally used to argue for learned reasoning and planning, makes the result more relevant than it would be on a purpose-built task." Reviewer y8kb appreciated the wide range of interventions and detailed analysis. Reviewer WW1D noted that "The results provide consistent evidence that correct answers and valid traces can become decoupled." Reviewer y8kb also counted the writing quality as a strength: "In the age of LLMs, I really appreciate the authors' effort to make this paper readable for a human reviewer like me."

All three reviewers responded positively to our rebuttal and new experiments, which addressed the concerns and weaknesses they raised. Reviewer rZHo raised Soundness from 2 to 3 and the rating from 2 to 3, writing that the independent-checkpoint evidence and the requested breakdown mostly answered their two main concerns. Reviewer y8kb, who had conditioned their initial score on Weaknesses 4 and 5 being addressed, was satisfied that "the revised draft is standing up to this expected state." Reviewer WW1D also affirmed their positive score.

New experiments and analysis (new Section 5; Appendix D.1 to D.5; changes in blue):

  1. RL fine-tuning with a validity-gated reward versus outcome-only RL and a continued-pretraining control, plus the checker as an inference-time monitor (requested by y8kb).
  2. Hidden-state probes on non-minimal models and on re-asked questions, which resolve the planning question (requested by y8kb).
  3. Larger model replication: we trained a 4x larger Qwen3-architecture model from scratch, reproducing the decoupling (requested by WW1D).
  4. Third-party independent replication of the decoupling on checkpoints trained and released by an independent group (Zhang et al., ICML 2026), scored with their own checker, addressing any concerns about results being artifacts of our training (rZHo).
  5. A strict recount of the correct-answers-with-invalid-traces data still shows a decoupling (rZHo).
  6. The suffix-validity analysis of shuffled traces (asked by rZHo) showed that for the 10% shuffled model the derivation after the shuffled block is mostly valid. We also addressed their final comments about the abstract.

On the ethics flag. Reviewer y8kb flagged the paper for ethics review; however, the review text raises no specific concern. The work uses only synthetically generated grade-school math data and involves no human subjects or personal data in any form. We have not been able to identify a concern the flag refers to and will gladly address any that the Area Chair or the ethics reviewers raise.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.