acceptodds
Under review as a conference paper at ICLR 2027

AXIOMSHIFT: Altering Axioms to Evaluate LLM Reasoning At Knowledge Boundaries

Abstract

Mathematical reasoning has become one of the central proving grounds for state-of-the-art AI, with recent systems solving several long-standing open problems. However, it remains unclear whether a model’s answer follows from the given premises or prior knowledge encoded in its parameters, or if the model even recognizes when the provided premise lacks enough information to determine an answer. On conventional benchmarks, these differences are indistinguishable. We introduce AxiomShift, a 700-question benchmark that isolates two distinct boundaries of model knowledge. First, it distinguishes supplied from parametric knowledge by perturbing a single axiom in each evaluation world, forming counterfactuals that test whether a model answer is attributable to prior knowledge or the given premise. The second boundary tests whether a model recognises and respects what can be known from given premises, through withholding or contradicting information required for the derivation, which produces genuinely unanswerable questions. Across 26 models, we identify four distinct model failure types. As model accuracy increases, simple numerical errors decline. However, prior leakage itself demonstrates a sharp split, with leakage on answerable problems decreasing (12.6% to 1.9%), while rising on unanswerable problems (12.1% to 15.5%). Taken together, our findings suggest model scaling improves ability to follow supplied premises, however, does not reliably improve recognition of knowledge boundaries.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
AI Security, Data Sovereignty, Reasoning, Knowledge, Foundation Model, Prior, LLM

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_KkXgOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors test whether models follow rules that differ from familiar math and science rules. Some questions can be solved using the supplied information. Others have missing or conflicting information, so the model should say it cannot give a definite answer. More accurate models do better on the solvable questions, but they still sometimes give textbook answers when they should decline to answer.

Critical strengths
  1. The paper tests whether models can follow new rules and recognise when information is missing. A model could do one well and still struggle with the other.

  2. The authors report how often models decline to answer and whether their wrong answers match textbook answers. This gives a clearer picture of the mistakes than overall accuracy alone.

Critical weaknesses
  1. An answer close to the textbook answer could also come from a calculation mistake. I need more evidence before accepting the authors' explanation. The example that names a method without showing a calculation also does not establish that checking the steps would miss the error.

  2. The main analysis of questions without a definite answer uses different sets of completed questions for different models. The authors include some checks, but I still want this analysis repeated on the same questions for every model.

  3. I found the definitions and results difficult to follow. A complete example would help explain the scoring. The authors should also make clear what each percentage counts.

  4. I want to know whether the wording of the questions affects the results. Models might behave differently if they are explicitly told that the supplied rules replace ordinary textbook rules.

Questions for authors
  1. How do you distinguish a model using textbook rules from one making a calculation error? Do the conclusions change if you require closer numerical matches?
  2. Does the main result hold when every model is evaluated on the same questions?
  3. Could you explain the scoring through a worked example and clarify what each percentage measures?
  4. Do the results change when the prompt explicitly tells models to follow the supplied rules over textbook knowledge?
3 comments
AuthorsOpenReview ↗

Response to Reviewer KkXg

Response to Reviewer KkXg

Thank you for the clear and constructive review. We address each point below. All changes are in the revised PDF.

W1 & Q1: Textbook rules versus calculation errors

We agree that matching the prior is not enough on its own, so we added three pieces of evidence.

First, in a new analysis, we manually checked reasoning traces behind all 142 prior-matching answers of Qwen3-30B-A3B. In 128 of them the trace uses the prior directly. In the other 14 it does not say which value was used (new manual check, end of Section 4.3.2). We have included the full list of the 142 checked questions and their responses to our rebuttal version of the supplementary material, under the name Qwen3-30B-A3B-Instruct_manual_check.xlsx.

Second, in a new experiment we conducted, we measured prior preference from token probabilities (details can be viewed at Appendix D of rebuttal pdf). At the point where the model starts its final answer, we read off the probability it gives to the prior and to two decoy numbers next to it. A model that reaches the prior through an arithmetic slip has no reason to favour it over its neighbours. Yet on the questions where Qwen3-4B gave the prior, it already favours the prior over both decoys before any reasoning on 82% of them, against 42% on its other questions. Qwen2.5-Math-7B shows the same pattern (63% against 47%). The two reasoning models build the preference up while reasoning: Qwen3-4B-Think favours the prior over ABSTAIN on none of the unanswerable questions before reasoning, and on 34% once 60% of its trace is written.

Third, in a new experiment, we used tighter numerical tolerances (Appendix C). Halving the tolerance keeps 91.5% of prior-leaking answers across all 26 models (1,568 of 1,714), and 82.4% equal the prior exactly. Therefore out conclusions are unchanged.

On the example that names a method without a calculation: we agree it does not show that checking the steps would miss the error. We no longer rest the claim on it, and we have softened the section title to “Stated reasoning traces do not fully reveal leakage” (Section 4.3.1).

W2 & Q2: The same questions for every model

We repeated the analysis on the 330 questions completed by all 26 models (new analysis, Appendix B), and the main result holds. Between 10% and 90% accuracy, answerable leakage falls from 14.5% to 1.5% (r = −0.71). Unanswerable leakage does not fall: 14.9% to 19.5% (r = +0.20), and the slope’s 95% CI includes zero. The full-set values are 12.6% to 1.9% and 12.1% to 15.5%. Therefore, the conclusion drawn from this common question set is almost identical to the one we have made. All runs draw from the same shuffled list of 700 questions, so shorter runs are also a fair random draw across the seven worlds (Section 3).

W3 & Q3: A worked example and what each percentage counts

The box in Section 4.3.1 now walks through the full scoring of one question. It shows the question, how aexta^{\mathrm{ext}} and ainta^{\mathrm{int}} are computed, the tolerance τ(a)=max⁡(0.02, 0.01∣a∣)\tau(a)=\max(0.02,\,0.01|a|) with the two windows it gives, and how real replies to this question are classified into each failure type. The tolerance is also stated in the Evaluation Protocol (Section 3). The per-model table (Appendix E) defines each column’s denominator: accuracy is over completed questions, leakage is over committed answers on prior-distinguishing questions, and failure to abstain is over unanswerable questions.

W4 & Q4: Telling models that the supplied rules replace textbook rules

In a new experiment we conducted, we tested this directly (Section 4.2, Figure 8 in Appendix J). We reran the 210 unanswerable questions on four models, replacing the closing sentence with one of two stronger instructions. The first tells the model to use only the information given, even where it contradicts the real world. The second also forbids any value or default assumption from prior knowledge, and names filling in a missing quantity explicitly. The first instruction has no effect. The second lowers failure to abstain from 52.6% to 46.9%, yet 31.8% of the answers models still commit are the prior, against 36.9% under the original prompt. Prior gap-filling is not mainly an artifact of weak wording, while weak wording can have some slight confounding effects.

Thank you for the additional experiments. They substantially address my main concerns, and I am raising my recommendation to weak acceptance. The tasks are still narrow, but they show that solving problems correctly and knowing when not to answer are different abilities and to approach them soundly.

Thank you for the follow-up and for raising your recommendation. We are glad the additional experiments addressed your main concerns. We agree that the task scope could be made wider. We will further expand the worlds and tasks in the camera-ready version.

Reviewer_uYdXOpenReview ↗
Rating
1
Confidence
3(+1)
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes a benchmark for evaluating LLM reasoning behavior at so called 'knowledge boundaries', situations where the model is not given the requisite information to solve the problem, or when the problem requires the use of information that runs counter to prior assumptions. The paper separates prior vs non-prior knowledge and deducible vs non-deducible knowledge, and evaluates LLM behavior on both. They show that prior leakage remains even when performance improves. Stronger reasoning LLMs are likely to fill in missing information, and weaker LLMs are likely to ignore facts that run counter to prior knowledge.

Critical strengths
  1. The paper's figures do a good job at explaining the taxonomy of the various error types discussed in the paper.
Critical weaknesses
  1. The benchmark consists of an incredibly specific set of tasks with limited variation in its content and counter factual interventions. Therefore the results have very little generality and cannot be considered robust evaluations of LLM reasoning behavior.

  2. It is unclear how deference to prior knowledge is defined and determined when reading the outputs of the evaluated LLMs. The paper mentions that the prior answer is determined by the answer taken from 'train-world axioms', but there are no mentions of what those are in the paper.

  3. The only significant finding seems to be that prior leakage induced errors reduce at a lower rate compared to other error types. While interesting, the contributions are somewhat lacking.

Questions for authors
  1. How is prior knowledge determined in your method? In particular for the qualitative analysis, how do you assess whether substituted or inserted constants are the result of prior knowledge vs simple mistakes?
4 comments
AuthorsOpenReview ↗

Response to Reviewer uYdX (1/2): summary, presentation, generality

Response to Reviewer uYdX (Part 1/2)

Thank you for taking the time to review our paper, and for the honest feedback. We have revised the manuscript substantially in response, and we address each of your points below. All changes are highlighted in the revised PDF.

1. On the summary

Thank you for the summary, we think it captures the main thrust of the paper well. There are two small places where we think our original wording may have given a slightly different impression from what we intended, and we would like to clarify them.

First, the distinction we draw is between answerable and unanswerable questions rather than “deducible and non-deducible knowledge”. A question is answerable when the supplied axioms pin down exactly one value, and unanswerable when they are contradictory, admit several values, or leave out something the derivation needs (Section 2 PROBLEM FORMULATION). The two framings are close in spirit, but we wanted to make sure the terminology lines up with the paper.

Second, on weaker models: we agree that stronger reasoning models tend to fill missing information with the prior, and we see why the paper may have suggested that weaker models “ignore facts that conflict with prior knowledge”. Our data actually point the other way. Only 7.3% of the weakest five models' committed answers on answerable questions match the prior (Table 1). Most of their errors are numbers that match neither the correct answer nor the prior. To further investigate your point, we ran a new experiment that shows overriding the supplied axiom with the prior turns out to peak at models with middle capability (Section 4.2, new Figure 7 in Appendix J). A model has to be capable enough to compute/retrieve the familiar prior answer before it can fall back on it. We have made this clearer in the rebuttal pdf. That is why our work does not support your claim of “weaker models ignore facts that conflict with prior knowledge”

2. Presentation

We agree that the original manuscript was harder to follow than it should have been, and we apologize for that. It likely contributed to some of the points above. We have reworked the writing throughout, and we would be glad to hear about any passage that is still unclear.

3. Weakness 1: Generality and variation of tasks

To address your the concern on generality of tasks, and to further diversify the question set, we produced an additional 3 worlds in the fields of Biology, Astronomy and Chemistry, with 300 new questions. The three new worlds follow the same format. On the four models they give results close to the original seven (Section 4.2, Appendix F and H). Appendix H also shows one prior-leaking reply from each new world.

The worlds differ in the kind of change, not only in content:

  • Some replace a memorized table or constant: the genetic code (codon), combining powers (valence), Mendel’s one-half share (inheritance ratio) and the ion product of water (ion product).
  • Some change a physical law: the equivalence of gravitational and inertial mass (two-mass) and the inverse-square law of gravity (gravity power).
  • Some change the curvature of space (polygon, SAS).
  • One changes an economic law, and one removes an order axiom.

Within them, the 66 question families each use their own derivation (45 in the original seven worlds, 21 in the new three). Inputs range from a single formula or one printed constant to a 64-codon table or a 125-fact order relation. Answers include both integer counts and continuous quantities. Every world has all three forms of unanswerability.

(Continued in Part 2/2.)

AuthorsOpenReview ↗

Response to Reviewer uYdX (2/2): prior answers, simple mistakes, contributions

Response to Reviewer uYdX (Part 2/2)

4. Weakness 2 and your question: how prior answers are determined, and how we separate them from simple mistakes

We have clarified “training-world axioms” in Section 2. To better illustrate the process, we added a box in Section 4.3.1 that walks through the full scoring of one question. The grading tolerance τ(a)=max⁡(0.02, 0.01∣a∣)\tau(a)=\max(0.02,\,0.01|a|) is stated in Section 3, under Evaluation Protocol.

Q1 (prior leakage vs simple mistakes). A wrong answer can land on the prior by accident, so we checked this in three ways.

  1. Manual check. We added an experiment in which we manually read the reasoning traces behind all 142 prior-matching answers of Qwen3-30B-A3B. In 128 of them, the trace uses the prior directly. In the other 14, the trace does not say which value was used, the result is discussed at end of Section 4.3.2. We have included the full list of the 142 checked questions and their responses to our rebuttal version of the supplementary material, under the name Qwen3-30B-A3B-Instruct_manual_check.xlsx.
  2. Prior preference test (Appendix D). In a new experiment we added, at the point where the model starts its final answer, we read off the probability it gives to the prior and to two decoy numbers next to it. A model that reaches the prior through an arithmetic slip has no reason to favour it over its neighbours. Result is presented in a new section in appendix D.
    • On the questions where Qwen3-4B gave the prior, it already favours the prior over both decoys before any reasoning on 82% of them, against 42% on its other questions. Qwen2.5-Math-7B shows the same pattern (63% against 47%).
    • The two reasoning models build the preference up while reasoning. On unanswerable questions, Qwen3-4B-Think favours the prior over ABSTAIN on none of the questions before reasoning and on 34% once 60% of its trace is written.
  3. Stricter matching (Appendix C). In a new analysis we conducted, halving the tolerance keeps 91.5% of prior-leaking answers (1,568 of 1,714), and 82.4% equal the prior exactly. Result is presented in appendix C.

5. Weakness 3: contributions

We understand the concern, but we see the contribution as broader than a single finding. The paper offers three things:

  1. A novel benchmark: 700 (now increased to 1000 during rebuttal) questions across seven worlds and 45 families, with generators, seeds, per-question model replies, and a GUI so others can reproduce it. This is supported by reviewer 5BLe “Programmatic construction and more comprehensive result reports provide a solid foundation for further analysis.”
  2. A framework for analyzing failure modes masked by ordinary model accuracy measurements: characterize model behaviors around two knowledge boundaries and classify their behavior errors into four disjoint failure types, which ordinary accuracy simply collapses into a single number. This is supported by reviewer KkXg “This gives a clearer picture of the mistakes than overall accuracy alone.” and reviewer 5BLe “This paper does not merely compare the accuracy differences between conventional tasks and counterfactual tasks, but also considers situations such as generating a unique answer given a premise, lacking necessary information, having multiple answers, and contradictory premises.”
  3. A distinct failure mode, prior leakage (gap-filling & override). When required information is withheld, stronger models abstain more often, but when they do commit, most of their answers are the prior. This is also supported by reviewer 5BLe “In particular, separating the analysis of "prior coverage of provided conditions" and "prior completion of missing conditions" helps understand the different performances of the model on the two types of knowledge boundaries”

Thank you again for your time. We hope these clarifications and revisions address your concerns, and if so, we would be grateful if you would consider revisiting your assessment. We are happy to answer any further questions.

Dear Reviewer uYdX,

Thank you again for your review. As the discussion period is closing in 2 days, we wanted to check whether our response has addressed your concerns. In brief, we added three new worlds in biology, astronomy and chemistry (Appendix H), a manual check of the reasoning traces behind 142 prior-matching answers, a token-probability test of prior preference (Appendix D), and a stricter-tolerance analysis (Appendix C). We also reworked the presentation throughout.

We would be grateful for any further feedback, and we are happy to clarify anything that remains unclear.

First of all, thank you for the response. I have assessed your rebuttals but I am still concerned about the generality of your synthetic tasks. You have added 3 new topics but the possible variations/perturbations within each task remain incredibly specific. I do not think the general capabilities of an LLM can be discriminated based on its ability to discern a dozen or so variations to problem contents such as the curvature of space, equivalence of gravitational and inertial mass etc. There needs to be a more systematic, comprehensive method to test this.

For this reason, I have raised my confidence on my initial assessment.

Reviewer_5BLeOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces AXIOMSHIFT, a benchmark used to evaluate how large language models handle knowledge boundaries. The paper distinguishes between two types of boundaries: the boundary between internal parameter knowledge and context-provided knowledge, and the boundary between conclusions that can be determined based on given premises and those that should be refused due to insufficient information.The benchmark consists of seven mathematical or scientific domains, 45 types of question templates, and 700 questions. Among them, 490 questions can be answered, and 210 questions require refusal due to contradictory conditions, non-uniqueness of answers, or insufficient information. The authors changed key assumptions to calculate the answers under given assumptions and the "prior answers" in the conventional world, and classified the errors into four categories: non-prior numerical errors, non-prior refusal failures, prior leakage, and incorrect refusal.The authors evaluated 26 models and found that as the benchmark accuracy improved, non-prior numerical errors decreased rapidly, while errors matching the prior answers decreased slowly, resulting in an increase in the proportion of remaining errors. Further analysis showed that the prior matching rate on answerable questions significantly decreased with accuracy, but no similar trend was observed for unanswerable questions. The paper also discussed different forms of prior leakage through output cases and keyword classification.

Critical strengths
  1. This paper does not merely compare the accuracy differences between conventional tasks and counterfactual tasks, but also considers situations such as generating a unique answer given a premise, lacking necessary information, having multiple answers, and contradictory premises. By setting ABSTAIN and differentiating different error types, the benchmark can reveal the different behaviors behind the same overall accuracy rate. In particular, separating the analysis of "prior coverage of provided conditions" and "prior completion of missing conditions" helps understand the different performances of the model on the two types of knowledge boundaries. This is the main conceptual contribution supporting acceptance.
  2. Programmatic construction and more comprehensive result reports provide a solid foundation for further analysis. Seven worlds, 45 types of templates and corresponding calculators enable the generation of examples with clear reference answers. The evaluation of 26 models covers different model families and training types. The author also provides the accuracy rate for each model and Wilson confidence intervals, and in Appendix B, examines the overall trend after including truncation and uninterpretable output. These efforts enhance the transparency of the evaluation. The author also explicitly acknowledges that there are confounding factors such as training data and post-training methods in cross-checkpoint comparisons, avoiding the complete disregard of experimental limitations.
Critical weaknesses
  1. Matching the prior answer is not the same as prior leakage. Ordinary arithmetic errors, reversed proportions, or numerical approximations can also produce answers equal or close to the prior answer, and the paper does not verify that models actually hold the relevant prior in the standard world. Moreover, the tolerance settings are not specified, nor is it shown that the tolerance intervals of the correct and prior answers never overlap (e.g., Table 3 counts 1.23 as leakage toward the prior answer 1.24), even though the entire four-way error classification depends on this scoring step.
  2. The claim that leakage cannot be detected from reasoning traces is under-supported. For example, in Appendix E.1 the model restates the combining capacities of zinc and chlorine as 1 and 2 yet gives a 1:2 ratio; this error is checkable from the numbers in the prompt alone and looks more like an ordinary calculation error than unmonitorable leakage. Section 4.4 relies on keyword classification without reporting manual verification accuracy, so the mechanistic attributions should be considerably softened.
  3. The statistical link between capability and leakage does not support the core claims about scaling and trends. Accuracy on the x-axis and the error rates on the y-axis are dependent by construction, and the 26 models are not a controlled scaling experiment; on unanswerable questions r is only +0.19, yet the abstract calls this “rising” while Section 4.2 says it “does not move at all.” The >100% “elimination ratios” from OLS extrapolation in Table 6 show the linear model is unsuitable near proportion bounds, and since models completed different question sets (e.g., 481 vs. 700), key results should be re-checked on a common set.
Questions for authors
  1. What tolerances are used for numerical correctness and for prior matching? How are outputs falling within both tolerance ranges handled? Did all 644 prior-distinguishing questions pass a distinguishability check under the actual tolerances?
  2. Did you evaluate the same models on the corresponding standard-world questions to confirm they hold the relevant priors? Did you manually sample prior-matching outputs to distinguish genuine prior invocation from ordinary calculation errors?
  3. What is the confidence interval of the regression slope for unanswerable questions in Figure 5? Do the authors claim “leakage significantly increases” or “no reliable decrease is observed”? Does this result hold on the question set completed by all models?
2 comments
AuthorsOpenReview ↗

Response to Reviewer 5BLe

Response to Reviewer 5BLe

We thank the reviewer for the careful reading and the suggestions on improving our work. We have run additional experiments and updated our paper based on your advice. We address your concerns below.

W1 & Q1: Tolerance and overlap

  • We revise our paper and explain the tolerance intervals in Sec 3, line 238-240 of the rebuttal pdf (Evaluation Protocol): an answer matches a reference value aa when it falls within τ(a)=max⁡(0.02, 0.01∣a∣)\tau(a) = \max(0.02,\ 0.01|a|) of it.
  • Our new experiments show that halving the tolerance keeps 91.5% of prior-leaking answers (1,568 of 1,714), more detailed results are presented in Appendix Section C.

W2 & Q2: Manual verification

To further support our claim, we manually verify all 142 prior-matching answers of Qwen3-30B-A3B, one of the models with the highest leakage rate. In 128 cases, the reasoning trace uses the prior directly. In the other 14 cases, the trace does not say which value was used (end of Section 4.3.2). We have also softened the section title to “Stated reasoning traces do not fully reveal leakage”. We have included the full list of the 142 checked questions and their responses to our rebuttal version of the supplementary material, under the name Qwen3-30B-A3B-Instruct_manual_check.xlsx.

W3 & Q3: Trend on unanswerable questions and the common set

(a) Fig 5 confidence Interval & and claim on leakage increase: Thank you for the suggestion. The fitted slope for prior leakage on unanswerable questions is +0.042 per accuracy point, with a 95% CI of [−0.051, +0.136] . The interval includes zero, so the fitted line was misleading, and we have removed it from Figure 5. We have revised our main claim to: prior gap-filling on unanswerable questions shows no reliable decrease with competence. The decline on answerable questions remains clear (slope −0.133, 95% CI [−0.189, −0.077].

(b) Key results checked on a common set: In a new experiment, we extracted answers from the 330 questions completed by all 26 models (Appendix B). The split between answerable and unanswerable holds. On the common set, between 10% and 90% accuracy, answerable leakage falls from 14.5% to 1.5% (compared to 12.6% to 1.9% in original pdf), while unanswerable leakage does not fall, 14.9% to 19.5% (compared to 12.1 to 15.5% in original pdf). We have also stated in the original pdf that “The questions are shuffled before feeding into models, ensuring any run is a fair random draw across all seven worlds.”

(c) Controlled scaling experiment. To address your question on the effect of model size, we ran a new size sweep over the full Qwen2.5 instruct line. We added the 0.5B, 1.5B and 3B models to the 7B–72B models already in the paper, so the sweep covers seven sizes from one family under the same prompt and grading (Appendix I, “A size sweep within Qwen2.5 instruct”). We fit a logistic regression of prior leakage on log size. Per-model numbers are in the same appendix.

  • Accuracy does not rise with size. It stays between 10.7% and 27.1% with no monotone pattern (+3.4 points per tenfold increase in size, 95% CI [−5.9, +12.6], r = +0.39).
  • Answerable questions: leakage rises from 1.6% at 0.5B to 16.2% at 72B (odds ratio 2.27 per tenfold increase in size, 95% CI [1.87, 2.76]).
  • Unanswerable questions: gap-filling rises from 2.0% to 13.0% (odds ratio 2.60, [1.85, 3.64]).

Within this family, larger models land on the prior more often without becoming more accurate. The smallest models rarely compute any answer at all (Qwen2.5-0.5B gets 4 of 487 answerable questions right), which partly explains their low leakage. Restricted to 3B–72B, the rise on answerable questions remains (+8.1 points per tenfold increase, 95% CI [+3.0, +13.3], r = +0.94), while the rise on unanswerable questions has a wide interval that includes zero (+6.1, [−3.1, +15.3]). Model size on its own therefore does not reduce prior leakage.

Dear Reviewer 5BLe,

Thank you again for your detailed review. As the discussion period draws to a close, we wanted to check whether our response has addressed your concerns. In brief, we stated the grading tolerance explicitly and showed that halving it keeps 91.5% of prior-leaking answers (Appendix C), manually checked the 142 prior-matching answers of Qwen3-30B-A3B, revised the claim on unanswerable questions in line with the confidence interval, repeated the key fits on the 330 questions common to all 26 models (Appendix B), and added a size sweep across the Qwen2.5 instruct line (Appendix I).

We would be grateful for any further feedback, and we are happy to clarify anything that remains unclear.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Global Response to all reviewers

Global response to all reviewers

We thank all three reviewers for their careful reading and feedback, and Reviewer KkXg for the updated assessment. Below we summarize the main shared concerns and the new experiments and analyses we added. All changes are highlighted in red in the rebuttal PDF.

1. Is prior leakage distinguishable from simple mistakes? (5BLe, KkXg, uYdX)

We added three independent checks.

  • Manual check (end of Section 4.3.2). We read the reasoning traces behind all 142 prior-matching answers of Qwen3-30B-A3B. In 128, the trace uses the prior directly. The other 14 do not state which value was used. The full list is in the supplementary material.
  • Prior preference from token probabilities (new experiment, Appendix D). At the point where the final answer begins, we compare the probability of the prior with two neighbouring decoys. An arithmetic slip gives no reason to favour the prior. Yet on the questions where Qwen3-4B gave the prior, it already favours it over both decoys before any reasoning on 82% of them, against 42% on its other questions (Qwen2.5-Math-7B: 63% against 47%). The reasoning models build this preference up as they reason.
  • Tighter tolerance (new analysis, Appendix C). Halving the tolerance keeps 91.5% of prior-leaking answers (1,568 of 1,714), and 82.4% equal the prior exactly.

2. How answers are scored (5BLe, KkXg, uYdX)

We now state the tolerance τ(a)=max⁡(0.02, 0.01∣a∣)\tau(a)=\max(0.02,\,0.01|a|) in the Evaluation Protocol (Section 3). We also added a new box in Section 4.3.1 that walks through the full scoring of one question, from both verifiers to how real model replies are scored. The per-model table (Appendix E) states the denominator of every rate.

3. Different question pools across models and the trend on unanswerable questions (5BLe, KkXg)

We repeated the key fits on the 330 questions completed by all 26 models (new analysis, Appendix B). The split holds: answerable leakage falls from 14.5% to 1.5% between 10% and 90% accuracy, while unanswerable leakage does not fall (14.9% to 19.5%). Since the 95% CI of the unanswerable slope includes zero, we removed that fitted line from Figure 5 and revised the claim to prior gap-filling shows no reliable decrease with competence.

4. Narrow task scope (uYdX, KkXg)

We added three worlds in genetics, astronomy and acid–base chemistry, with 300 new questions in the same format (new experiment, Section 4.2, Appendices F and H). They give results close to the original seven: prior leakage appears in all three new worlds, with a pooled rate of 15.2% against 18.8% on the original worlds for the same four models. We will expand the tasks further in the final version.

5. Would stronger instructions remove gap-filling? (KkXg)

We reran the 210 unanswerable questions on four models with two stronger instructions (new experiment, Section 4.2, Figure 8 in Appendix J). Telling models to use only the given information has no effect. Explicitly prohibiting gap-filling lowers failure to abstain from 52.6% to 46.9%, yet 31.8% of the answers models still commit are the prior.

6. The role of model size (5BLe)

We completed the Qwen2.5 instruct line from 0.5B to 72B (new experiment, Appendix I). Accuracy shows no trend with size, while prior leakage rises from 1.6% to 16.2% on answerable questions and from 2.0% to 13.0% on unanswerable ones. Model size on its own does not reduce leakage.

7. Do weaker models ignore conflicting facts? (uYdX)

In a new analysis that we conducted to address this question, placing every numerical answer between the correct value and the prior (new analysis, Section 4.2, Figure 7 in Appendix J) shows that prior-overriding peaks at middle capability. A model must be capable enough to compute the familiar answer before it can fall back on it.

We have also reworked the writing throughout for clarity. We are happy to answer any further questions during the remaining discussion period.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.