acceptodds
Under review as a conference paper at ICLR 2027

A Study of Naturally Occurring Mathematical Errors in Real-World Theoretical Research with Frontier Models

Abstract

Recent advances in artificial intelligence have enabled frontier models to play a growing role in theoretical research. AI for mathematics, in particular, has attracted widespread attention. Researchers increasingly turn to proprietary models such as Fable 5 and GPT-5.6 to investigate research-level open questions. However, these models remain prone to mathematical errors, making it important to understand how such errors arise and what they imply for the responsible use of AI in theoretical research. Existing studies often examine mathematical errors through evaluations on selected problems, particularly in competition- and Olympiad-level mathematics. Systematic documentation and analysis of mathematical errors encountered during sustained, real-world theoretical research with frontier models remain limited. In this paper, we present 65 source-linked mathematical error records drawn from frontier-model investigations of 12 research-level problems in quantum information, quantum error correction, quantum complexity, and related fields. Each record links an erroneous claim or inference to its original source, relevant context, and a checkable witness explaining the mathematical defect. Where the evidence permits, we also examine visible reasoning summaries alongside subsequent outputs. Using an established classification framework, we find that logical and reasoning errors dominate our corpus, accounting for 68.2% of error labels. Overall, our study provides an empirical analysis of mathematical errors encountered in real-world theoretical research with frontier models.

Primary area:
foundation or frontier models, including LLMs
Keywords:
mathematical reasoning, mathematical errors, frontier models, AI-assisted theoretical research, empirical error analysis

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 4 reviews, above 9% of the others.
Reviewer_gjzMOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents 65 source-linked mathematical errors produced by frontier models across 12 research-level problems in quantum information, quantum error correction, quantum complexity, and related fields. The authors also provide a comprehensive error analysis to investigate why these errors arise.

Critical strengths
  1. The collection of mathematical errors is a valuable resource, particularly as frontier models become increasingly capable and their remaining limitations harder to characterize.
Critical weaknesses
  1. The paper’s fit for an AI conference is unclear in its current form. Although the collected examples are valuable, the analysis remains largely descriptive, offering limited insight into the underlying mechanisms that produce these errors or actionable strategies for preventing or reducing them.

  2. The authors could strengthen the contribution by developing a benchmark for automatic mathematical error detection using the collected examples. Alternatively, these examples could serve as seeds for generating additional training data, enabling an investigation into whether targeted training reduces similar errors in LLMs.

Questions for authors

NA

Reviewer_YH6eOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

In the paper, authors present 65 source-linked mathematical error records drawn from frontier-model investigations of 12 research-level problems in quantum information, quantum error correction, quantum complexity, and related fields.

Critical strengths

The paper concluded some indexes for evaluating mathematic errors. The largest strength is proposing special ideas that combine various evaluation indicators.

Critical weaknesses
  1. The error corpus is small and discipline-restricted, with only 65 records drawn exclusively from quantum information, quantum error correction, and quantum complexity. It cannot support robust generalization of the reported error patterns to other mathematical or theoretical domains.
  2. Process-level diagnosis of error origins is incomplete, as the analysis depends only on exposed reasoning excerpts rather than full internal chain-of-thought traces of the frontier models. This prevents reliable discrimination between intrinsic reasoning failures and surface-level generation artifacts.
  3. The study adopts an existing error taxonomy originally built for competition mathematics, without validating its suitability for characterizing research-level mathematical inference. It remains unclear whether this framework captures unique failure modes of open-ended theoretical inquiry.
  4. There is no controlled benchmark comparison across the same frontier models between competition problems and these research tasks. Therefore, the observed shift toward logical/reasoning errors cannot be definitively attributed to the research nature of the problems rather than confounding factors such as prompt design, task duration, or human researcher filtering.
  5. The sampling of errors is opportunistic and observational, collected passively during the authors’ own research workflows rather than from a prespecified, standardized sampling protocol. Selection bias may skew the relative frequency of error types reported in the corpus.
Questions for authors
  1. How did you design your error sampling criteria and mitigate selection bias?
  2. Could the low rate of problem-understanding errors come from researchers’ careful question framing rather than model behavior?
  3. Without full chain-of-thought traces, how do you separate genuine logical flaws from summarization artifacts?
  4. Are there recurring quantum-specific failure modes poorly captured by the competition-derived taxonomy?
  5. What features of quantum reasoning may shape this error distribution, and which fields will you prioritize for expansion?
  6. Might researcher background or prompting choices substantially change the observed error distribution?
  7. Does your dataset distinguish easily caught minor errors from insidious ones that can mislead mathematicians?
  8. Will future corpus expansion include controlled experiments to disentangle task, prompt and model factors?
Reviewer_Jd4uOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This project seeks to measure frontier models on research-level mathematics tasks in an open ended setting. The authors collect a dataset of 65 mathematical errors that frontier models make (mostly regarding the field of quantum computing), classify them into a discrete set of major error types, and calculate the relative rates of those errors. For each error, they provide a witness that the LM did, in fact, make an error. They provide illustrative examples of several of the concrete errors.

Critical strengths

It seems like the ratio of different kinds of errors is different for this kind of research than in other evaluations and competition math, so if it is true it would be quite useful to know.

The relative prevalence of relative errors types is useful by itself, to some degree.

I especially liked the worked examples -- I could imagine training myself on these examples to catch similar ones when doing research myself.

Another strong point is that each error has a witness.

Critical weaknesses

The motivation of understanding what kinds of errors current frontier models make on open-ended research tasks is clear and worthwhile, but this paper does not live up to that motivation in 2 ways.

  1. It is not clear how the data was collected, and therefore it is difficult to interpret the results.
  2. There are relatively few data points (65 error examples), and they come from a narrow cluster of related fields.

With respect to point 1, this paper frames itself as comparing to existing datasets such as MATH, but MATH can be used to get an absolute error rate.

It is not clear what this evaluation has over a more comprehensive eval targeting research-level math, such as Humanity's Last Exam. You could answer that HLE is isolated problem solving rather than open ended exploration, but it is not clear to me a priori that this discrepancy should cause a difference in the LM's error rate given that we are conditioning on situations where the model is attempting to solve a question with an actual answer. That would need to be argued.

Also, this paper gives relative rather than absolute error rates, which is a significant scope restriction, because it gives no insight how suspicious the researcher should be of any given reasoning step.

Questions for authors
  • How were these errors generated?

  • What was the criterion for inclusion?

  • Was error discovery manual or automated?

    • If automated, is it using the same models you are evaluating?
    • If so, do you have a metric ameliorating or at least measuring that source of bias?
  • How many conversations were the errors pulled from?

  • What was the granularity of chosen reasoning steps?

  • If you want to evaluate models on research-level math problems, why not just look at the performance on humanity's last exam? If you are interested in the relative rate of errors, you could run the models on the dataset and then classify those errors and then calculate the relative ratios.

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

This paper presents a dataset of 65 mathematical error records produced by applying frontier models to 12 real-world theoretical research problems from quantum-related fields. For each error record, the dataset provides the original source and relevant context, as well as a "witness" report that describes the mathematical defect alongside checkable evidence such as a counterexample.

The paper studies an example each from four error categories represented in the dataset, and compares the distribution of errors across categories with that of an existing dataset covering competition- and Olympiad-level mathematics (Yoshida et al., 2026). The authors identify that logical and reasoning errors are the largest category in both datasets, but observe a higher proportion of logical implication errors and a far lower proportion of problem understanding errors in their own dataset compared to Yoshida et al. (2026). They suggest that Yoshida et al. (2026) contains a higher concentration of problem understanding errors because they arise from errors in the interpretation of a problem, and since Yoshida et al. (2026) covers many distinct problems (each requiring a new interpretation), it presents more opportunities for interpretation errors. Meanwhile, they suggest that their own dataset contains a higher concentration of logical implication errors because the source theoretical research problems require many intermediate claims and deductions, which present more opportunities for errors in how intermediate results are interpreted and applied.

The paper also provides a supplementary analysis of cases where an error is visible in the reasoning itself or as a discrepancy between reasoning and a subsequent output.

Critical strengths
  • High-quality, verifiable error records from an under-explored setting. Each record links an error to its original source and a checkable witness (e.g., a counterexample or explicit derivation), and naturally occurring errors from sustained research-level work are expensive to collect. This makes the dataset a valuable resource, even if the conclusions currently drawn from it are limited.
Critical weaknesses
  • Small and uncontrolled dataset, which undermines the comparison with Yoshida et al. (2026) that forms the paper’s main empirical result. The dataset contains only 65 error records, all from quantum-related fields, and they are heavily clustered: 21 come from a single research problem (channel additivity violations), so the records are not independent, and no uncertainty is reported for any of the percentages. More fundamentally, the two collections differ in model (frontier models versus Llama-3.3-70B-Instruct), task (research problems versus MATH) and collection process (errors the authors happened to identify during their own research versus failures selected by an automatic answer check). Any of these could account for the difference in distributions, so the observed differences cannot be attributed to the research setting specifically. To me, the dataset is therefore too small and too uncontrolled to support the comparative conclusions drawn from it.
  • Limited explanation of how errors arise in practice, despite this being the paper’s second stated contribution. The paper delivers mainly a distribution over error categories, and its mechanism analysis is thin. Only 12 of the 65 error records support error-to-reasoning alignments, and the paper's explanations for the observed distribution are untested (e.g., Section 4.5’s suggestion that sustained investigations create more opportunities for invalid inference). The problem is compounded by the unavailability of complete chain-of-thought traces, which means the process-level analysis must rely on provider-visible reasoning summaries; the authors themselves note that this limits the ability to reconstruct error formation in detail. The paper’s claims about how errors arise are therefore not adequately supported by the available evidence.
Questions for authors
  • (Additional analysis) Does the category distribution still differ from Yoshida et al. (2026) if records are weighted per research problem, or if the channel additivity records are excluded?
  • (Clarification) How were errors detected, and what were the inclusion criteria? Is there any estimate of how many errors were missed? Errors that are easier to spot (e.g., arithmetic) may be over-represented relative to subtler ones.
  • (Clarification) Which models produced which errors? A per-model breakdown would show whether the distribution reflects frontier models generally or is dominated by one model.
  • (Clarification) Will index.json and the per-record evidence folders be publicly released?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.