acceptodds
Under review as a conference paper at ICLR 2027

LinAlg-Bench: When LLMs Stop Computing — Structured Hallucination, Sign Drift, and Depth-Gated Collapse in Matrix Arithmetic

Abstract

Frontier LLMs execute linear algebra flawlessly — until, at a measurable depth, they stop computing and start fabricating. We introduce LinAlg-Bench, a diagnostic benchmark of nearly 1,200 SymPy-verified problems across nine task types and 3×3–5×5 matrices, evaluated on ten frontier models at temperature zero (over 11,000 graded outputs, with results stable across three complete runs). Failure is not gradual but structural: models near ceiling on shallow tasks collapse to near zero on 5×5 eigenvalues — and the collapse is spectrum-conditional, with over 40% aggregate accuracy on integer spectra versus under 15% on irrational ones, implicating required numerical depth rather than matrix size. A three-stage forensic pipeline reveals that failures dissociate cleanly by task: eigenvalues fail by constraint-aware confabulation — fabricated values still match the matrix trace in nearly half of cases; determinants fail by sign accumulation, with a dimensiongated complete collapse that is absent at 4×4 yet dominant at 5×5. Forcing an O(n3) algorithm yields no recovery, chain-of-thought prompting rescues only the one model already strongest on the task, and chained shallow operations with total arithmetic matched to one eigenvalue computation remain near-ceiling — together implicating depth-of-state, not operation count, strategy, or prompting, as the bottleneck. LinAlg-Bench frames this working-memory limit as a falsifiable hypothesis and provides the instrument to test it.

Primary area:
datasets and benchmarks
Keywords:
large language models, benchmark, mathematical reasoning, linear algebra, hallucination, working memory, execution depth, evaluation, error analysis, fabrication

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_2864OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper introduces LinAlg-Bench, a diagnostic benchmark of 1,194 SymPy-verified linear algebra problems across nine tasks on 3X3, 4X4, 5X5 matrices. The goal is to measure mathematical execution depth (and when it breaks). An LLM-as-a-Judge is used to categorize the types of errors made and the analysis (after testing on 10 LLMs) shows a shift in the type of error as the challenge increases.

Critical strengths
  • It is important to have granular, controled-difficulty, measurements of LLM skills. This work suggests another angle for that.
  • The method seems easy to extend further as models become stronger (6X6 matrices, more advanced matrix operations, etc).
Critical weaknesses
  • There seems to be inconsistencies in the data/claims throught the paper. For example, figure 2 claims that arithmetic errors persist - but the total number of problems per dimension is not the same, so in %, arithmetic errors drop substantially. The "Declared-pass protocol" is an obvious bias upwards for the 2 frontier models - why not just take relative % instead of assuming pass on problems which we not attempted? Keeping only parseable boxed answers creates another data bias. "Judge-leniency" looks like an ad-hock crutch. Etc.

  • Writing quality: The abstract and introduction contains many terms which are not defined and not widely known. It sells results without explaining what was done. Almost all citations are in the 2 paragraphs of previous works - and missing obvious papers (e.g. Charton 2020 is cited but Charton 2022 "Linear Algebra With Transformers" is not). Figure captions not self-contained with no clear message. Figure fonts too small for human eyes. Looks very "Claudeish".

Questions for authors
  • In table 1 the core challenge design is described. Why do you say that each level strictly adds a "computational requirement"? e.g. Gaussian elimination requires multi-step memory. Determinant calculation doesn't necesseraly require tracking a similar number of steps, and the steps are different. The generality of the method depends on the ability to order the challenges by difficulty, but that's not obvious to do.
Reviewer_1kYvOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

Authors introduce LinAlg-Bench, designed to diagnose LLM failures in unaided linear algebra computation. The benchmark comprises 1K+ SymPy-verified problems across 9 tasks and 3 matrix sizes (3 by 3 - 5 by 5), including an eigenvalue subset varying matrix symmetry & spectrum type. Authors evaluate ten models and propose a forensic error taxonomy, identifying determinant/eigenvalue degradation and constraint-consistent fabricated answers, with a depth-dependent working-memory bottleneck as the proposed explanation.

Critical strengths
  • Diagnostic benchmark design. Exact ground truth & regenerable problems make the benchmark useful for controlled evaluation, while varying matrix dimension exposes failures that single-size evaluations would miss.

  • Factorial eigenvalue evaluation. Separating symmetry & spectrum type provides a useful distinction between problems with easily searchable integer answers and those requiring more demanding numerical computation.

  • Detailed failure analysis beyond final-answer accuracy. The findings that fabricated eigenvalues often preserve the trace but not the determinant is somewhat interesting -- showing how incorrect answers can satisfy simple verification checks.

Critical weaknesses
  • Evaluation design seems problemetic -- according to appendix the unscored / unmeasured problems are practically measured and credited as success - this leads to 100% accuracy reported on 5x5 determenants whereas only 16 have been attempted and evaluated.

  • Overstated causal & architectural conclusions. Increasing dimension also changes arithmetic complexity, operation count & root-finding requirements, and determinants can be computed without recursive cofactor expansion.

  • Limited insights from empirical results - it does not show critical insights about how frontier LLM models are trained or should be changed, other than attributing it as missed tool & plugin calls to recover from.

Questions for authors

See above

Reviewer_VMUkOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces a tool-free diagnostic benchmark for linear algebra, comprising 1,194 SymPy-verified problems across nine task types, 3×3 to 5×5 matrices, and ten models. Its eigenvalue task uses a factorial design over symmetry, spectrum type, and dimension. The authors analyze sign errors, abandonment of computation, and fabricated answers that satisfy simple constraints, and use Gaussian-elimination prompts, chained multiplication, integration, and scaffolding experiments to relate failures to depth of state dependence. The diagnostic perspective is useful, but problems with scoring and experimental controls currently weaken the quantitative comparisons and causal interpretation.

Critical strengths
  1. The benchmark provides informative diagnostics beyond an aggregate accuracy score.The 360-problem eigenvalue core uses a balanced 2×2×3 design with 30 problems per cell, helping distinguish associations with dimension, symmetry, and spectrum structure. The observation that incorrect eigenvalues more often satisfy the trace constraint than the determinant constraint is independently interesting: it characterizes the structure of errors and remains valuable even if the proposed depth-based mechanism is stated more cautiously.

  2. The paper includes substantial validation and multiple experimental probes. It reports 2,520 manually checked outputs, three input representations, three runs on a 660-problem subset, and four groups of probe experiments. These efforts provide useful material for assessing the robustness of the observed phenomena. The principal concern is how this evidence is scored and interpreted, rather than an absence of validation or repeated experiments.

Critical weaknesses
  1. Appendix C states that o1 attempted only 258 of 1,194 problems and GPT-5.2 attempted 699, with unattempted problems credited as correct under a capability-ordering assumption. Moreover, o1’s 5×5 determinant result is reported as 175/175 although only 16 were attempted, contradicting the claim that all 5×5 problems were directly evaluated. Although this protocol is disclosed, inferred success cannot substitute for direct evaluation. Observed-only results are needed to establish whether the main comparisons remain valid.

  2. The stated tolerance is 0.011, yet the paper reports accepting deviations of approximately 0.09 and 72 polynomial/interval-only answers. The scaffolding evaluation also requires decimal lists at every level, although L0 does not request them. Gemini’s large improvement already appears at L1, which adds only a decimal-format instruction. A consistent semantic scoring rule should distinguish completed solutions, unfinished root solving, and formatting failures before attributing these gains to computational assistance.

  3. The intervention is applied to previously failed problems. Mistral subsequently solves 7/26, or 7/22 after excluding unanswered calls, but the latter rate is compared with 38.3% on the full benchmark to argue against meaningful recovery. These rates concern different problem sets. Rerunning the original prompt on the same failure subset and reporting paired outcomes would help distinguish a strategy benefit from a retry effect.

  4. The integration analysis excludes 331/680 unparseable Qwen3 outputs, leaving as few as 2/40 responses in some cells, while the control separating depth from coefficient size remains incomplete. These limitations leave alternative explanations unresolved. The authors appropriately describe the working-memory account as a falsifiable behavioral hypothesis, but their stronger claims that competing explanations have been excluded or that the limitation is architectural exceed the evidence.

Questions for authors
  1. For each model × task × dimension, could you report the numbers directly evaluated, correctly solved, and not evaluated? Do the main curves and spectrum-type gaps persist on observed-only data or the eight fully evaluated models? How do you reconcile o1's 16 attempted 5×5 determinants with the main text's direct-coverage statement?

  2. Could you rescore eigenvalue and L0–L4 outputs under a single semantic rule, separately reporting complete numerical or exact solutions, polynomial-only answers, insufficiently precise intervals, and format-only failures? How much of Gemini's L0–L1 gain remains when correct non-decimal answers are accepted consistently?

  3. On the same failed-problem subset used for the Gaussian-elimination intervention, how many problems are recovered by rerunning the original prompt? Please report paired outcomes with all attempted problems in the primary denominator, and clarify what the seven Mistral recoveries imply about a local strategy benefit.

  4. Which controls isolate dependency depth from coefficient magnitude, numerical precision requirements, and answer length? Could you report integration success over all calls, alongside a manually checked analysis of unparseable answers, and reconcile the strong exclusionary claims with the stated behavioral-hypothesis scope?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.