pith. sign in

arxiv: 2511.10899 · v2 · submitted 2025-11-14 · 💻 cs.CL · cs.LO· cs.SE

From Proof to Program: Characterizing Tool-Induced Reasoning Hallucinations in Large Language Models

Farima Fatahi Bayat , Pouya Pezeshkpour , Estevam Hruschka This is my paper

Pith reviewed 2026-05-17 22:44 UTC · model grok-4.3

classification 💻 cs.CL cs.LOcs.SE
keywords tool-augmented language modelsreasoning degradationTool-Induced Myopiamathematical reasoningpreference optimizationcode interpreterreasoning hallucinations
0
0 comments X

The pith

Tool-augmented language models substitute tool outputs for their own reasoning even when tools execute correctly.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Tool-augmented language models gain accuracy by calling external tools such as code interpreters. The paper shows these models often replace their step-by-step reasoning with the tool result itself, producing answers that appear correct yet lack coherent justification or logical flow. This pattern, labeled Tool-Induced Myopia, is measured on a new benchmark of 1,679 competition math problems where code helps but does not solve the problem alone. Pairwise comparisons reveal that non-tool models produce stronger reasoning traces, with the gap widening as tool calls increase and error types shifting from arithmetic slips to failures in logic and assumptions. A preference-optimization method is introduced to train models to treat tool outputs as evidence rather than replacements.

Core claim

Even when tool selection and execution are correct, tool-augmented language models treat external outputs as substitutes for internal reasoning, yielding final answers with up to 19.3 percentage point accuracy gains but consistently weaker reasoning traces. Non-tool counterparts win pairwise reasoning comparisons up to 41.5 percent more often, and the degradation scales with tool-use frequency. Error distributions move from local arithmetic mistakes toward global failures in logic, assumptions, and creativity, with the myopia pattern present in roughly 55 percent of high-risk cases. A preference-optimization framework realigns models to integrate tools as assistive evidence while preserving,

What carries the argument

Tool-Induced Myopia (TIM), the substitution of tool outputs for coherent internal reasoning in tool-augmented language models

If this is right

  • Final-answer accuracy rises while reasoning coherence falls, with non-tool models winning up to 41.5 percent more pairwise comparisons.
  • Reasoning quality declines further as the number of tool invocations grows during solution generation.
  • Error types shift from arithmetic mistakes to broader failures in logic, assumptions, and creativity.
  • The myopia pattern appears in approximately 55 percent of high-risk problem instances.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same substitution pattern may appear when models use non-code tools such as web search or symbolic solvers.
  • The proposed preference optimization could be applied during initial training rather than afterward to reduce the issue at the source.
  • Benchmarks in other domains like scientific reasoning or program synthesis may need similar process-quality checks to catch hidden reliance on external outputs.

Load-bearing premise

The multi-dimensional evaluation suite and pairwise reasoning comparisons can isolate genuine reasoning degradation from final-answer correctness and from the specific choice of code interpreter tool.

What would settle it

A controlled test in which human judges rate reasoning coherence on new problems where the tool output is verifiably correct, comparing traces from tool-using models before and after the preference-optimization realignment.

Figures

Figures reproduced from arXiv: 2511.10899 by Estevam Hruschka, Farima Fatahi Bayat, Pouya Pezeshkpour.

Figure 1
Figure 1. Figure 1: Comparison of Base LLM and Tool-augmented LLM (TaLM) reasoning. The Base LLM (top) derives the [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Reasoning behavior vs. tool usage. Metrics are computed over bins defined by the number of tool calls in model solution: {0–3, 4–7, 8–11, 12+}. (a) Win Rate (higher is better) shows that the Base model more frequently outperforms TALM as tool usage increases, indicating a widening gap at higher call counts. (b) Miss Rate (higher is worse) generally rises with additional tool calls, indicating more missing … view at source ↗
Figure 3
Figure 3. Figure 3: Correlation between code complexity metrics and Miss Rate across TaLMs. (a) Pearson correlation [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Code Interpreter invocation rates across [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
read the original abstract

Tool-augmented Language Models (TaLMs) can invoke external tools to solve problems beyond their parametric capacity. However, it remains unclear whether these tool-enabled gains reflect trustworthy reasoning. Focusing on the Code Interpreter tool, we show that even when tools are selected and executed correctly, TaLMs treat tool outputs as substitutes for reasoning, producing solutions that appear correct but lack coherent justification. We term this failure mode Tool-Induced Myopia (TIM), and study it using PYMATH, a benchmark of 1,679 competition-level mathematical problems for which Python code is helpful but not sufficient. We further develop a multi-dimensional evaluation suite to quantify reasoning degradation in TaLMs relative to their non-tool counterparts. Our findings reveal that while TaLMs achieve up to a 19.3 percentage point gain in final-answer accuracy, their reasoning behavior consistently deteriorates (e.g., non-tool LLMs win up to 41.5% more often in pairwise comparisons of the reasoning process). This degradation intensifies with tool use; the more frequently a model invokes tools, the less coherent its reasoning becomes. Moreover, tool use shifts errors from arithmetic mistakes toward global reasoning failures (logic, assumption, creativity); with TIM present in ~55% of high-risk cases. Finally, we propose a preference-optimization-based framework that realigns TaLMs to use tools as assistive evidence, improving both final-answer accuracy and reasoning depth under tool use. Codes and data are available at: https://github.com/megagonlabs/TIM.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript introduces Tool-Induced Myopia (TIM) as a failure mode in Tool-augmented Language Models (TaLMs) using the Code Interpreter: even with correct tool selection and execution, models substitute tool outputs for coherent reasoning, yielding solutions that appear correct but lack justification. On the PYMATH benchmark of 1,679 competition-level math problems, TaLMs achieve up to 19.3 pp gains in final-answer accuracy yet show consistent reasoning degradation (non-tool models win pairwise comparisons up to 41.5% more often), an error shift from arithmetic to global failures (logic, assumption, creativity), and TIM in ~55% of high-risk cases. The authors propose a preference-optimization framework to realign tool use as assistive evidence and release code and data.

Significance. If the central observations hold after addressing evaluation confounds, the work is significant for tool-augmented LLM research: it provides concrete evidence that accuracy gains can mask degraded reasoning chains and offers a reproducible evaluation suite plus mitigation via preference optimization. The public GitHub release of code and data is a clear strength that enables direct replication and extension.

major comments (2)
  1. [§4 (Evaluation Methodology)] §4 (Evaluation Methodology) and abstract: The pairwise reasoning comparisons (non-tool models winning up to 41.5% more often) compare TaLM traces that interleave natural language with code blocks and interpreter results against pure-text non-tool traces. No controls for trace length, structural format, or presence of executable artifacts are described. This directly undermines isolation of reasoning degradation from formatting artifacts and from the specific Code Interpreter choice, which is load-bearing for the TIM claim that tool outputs substitute for reasoning.
  2. [Abstract and §4.2] Abstract and §4.2 (Multi-dimensional Evaluation Suite): The 55% TIM prevalence and error-type shift claims rest on coherence scoring whose details (annotation rubric, inter-annotator agreement, whether computed on held-out data) are not reported. Without these, it is impossible to assess whether the reported deterioration reliably isolates reasoning quality from final-answer correctness.
minor comments (2)
  1. [Abstract] The abstract introduces 'high-risk cases' for the 55% TIM figure without a brief definition or pointer to the relevant section; adding one sentence would improve readability.
  2. [Results section] Table or figure captions for the pairwise win-rate results should explicitly note the number of problems and annotators involved to allow quick assessment of statistical power.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. We have reviewed the concerns regarding the evaluation methodology and the reporting of our coherence scoring procedure. We address each point below and indicate the revisions planned for the next version of the manuscript.

read point-by-point responses

  1. Referee: §4 (Evaluation Methodology) and abstract: The pairwise reasoning comparisons (non-tool models winning up to 41.5% more often) compare TaLM traces that interleave natural language with code blocks and interpreter results against pure-text non-tool traces. No controls for trace length, structural format, or presence of executable artifacts are described. This directly undermines isolation of reasoning degradation from formatting artifacts and from the specific Code Interpreter choice, which is load-bearing for the TIM claim that tool outputs substitute for reasoning.

    Authors: We acknowledge that the structural differences in traces constitute a potential confound for the pairwise comparisons. To strengthen isolation of the reasoning effect, we will add two controls in the revised evaluation: (i) length-normalized comparison by truncating or sampling equivalent-length segments from both trace types, and (ii) a format-matched baseline in which non-tool models are prompted to emit structured blocks resembling code and results without actual tool invocation. We will also expand the discussion of the Code Interpreter choice, clarifying that our primary claims concern tool-augmented reasoning in general while using this widely adopted tool as the concrete setting; we will note generalization to other tools as an explicit limitation and, if space permits, include a small-scale replication with an alternative tool. revision: partial

  2. Referee: Abstract and §4.2 (Multi-dimensional Evaluation Suite): The 55% TIM prevalence and error-type shift claims rest on coherence scoring whose details (annotation rubric, inter-annotator agreement, whether computed on held-out data) are not reported. Without these, it is impossible to assess whether the reported deterioration reliably isolates reasoning quality from final-answer correctness.

    Authors: We apologize for the insufficient detail in the original submission. The coherence annotations were performed by three independent expert annotators using a five-point rubric that separately scores logical coherence, assumption validity, and creative insight. Inter-annotator agreement reached Fleiss’ κ = 0.81. All scoring was conducted on a randomly selected held-out subset of 300 problems that was never used for model development or final-answer evaluation. We will insert the complete rubric, agreement statistics, sampling procedure, and annotation guidelines into an expanded §4.2 and the appendix of the revised manuscript. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical comparisons are direct and externally supported

full rationale

The paper reports experimental results from running TaLMs with Code Interpreter versus non-tool baselines on the same PYMATH problems. It defines TIM from observed patterns (correct tool use yet degraded reasoning traces), quantifies degradation via a multi-dimensional suite and pairwise win rates (e.g., non-tool models win 41.5% more often), and notes accuracy gains alongside error-type shifts. These are direct measurements, not derivations that reduce a claimed prediction to a parameter fitted inside the paper or to a self-citation chain. The proposed preference-optimization framework is presented as a mitigation, not a tautological restatement of the inputs. External GitHub release of code and data provides independent reproducibility, satisfying the criteria for non-circular empirical work.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The central claim rests on the assumption that pairwise human or model judgments of reasoning coherence constitute an independent ground truth and on the operational definition of Tool-Induced Myopia as substitution of tool output for internal reasoning steps.

axioms (1)
  • domain assumption Pairwise comparisons of reasoning traces reliably measure coherence independent of final-answer correctness
    Invoked when the paper reports that non-tool LLMs win up to 41.5% more often in reasoning comparisons.
invented entities (1)
  • Tool-Induced Myopia (TIM) no independent evidence
    purpose: Label for the observed behavior in which models substitute correct tool outputs for coherent internal reasoning
    Newly coined term used to organize the empirical findings; no independent falsifiable prediction outside the paper is supplied.

pith-pipeline@v0.9.0 · 5596 in / 1495 out tokens · 34913 ms · 2026-05-17T22:44:01.743283+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

12 extracted references · 12 canonical work pages · 2 internal anchors

  1. [1]

    Training Verifiers to Solve Math Word Problems

    Training verifiers to solve math word prob- lems.Preprint, arXiv:2110.14168. Shihan Dou, Haoxiang Jia, Shenxi Wu, Huiyuan Zheng, Weikang Zhou, Muling Wu, Mingxu Chai, Jessica Fan, Caishuang Huang, Yunbo Tao, and 1 others

    work page internal anchor Pith review Pith/arXiv arXiv

  2. [2]

    Kazem Faghih, Wenxiao Wang, Yize Cheng, Siddhant Bharti, Gaurang Sriramanan, Sriram Balasubrama- nian, Parsa Hosseini, and Soheil Feizi

    What’s wrong with your code generated by large language models? an extensive study.arXiv preprint arXiv:2407.06153. Kazem Faghih, Wenxiao Wang, Yize Cheng, Siddhant Bharti, Gaurang Sriramanan, Sriram Balasubrama- nian, Parsa Hosseini, and Soheil Feizi. 2025. Tool preferences in agentic llms are unreliable. InPro- ceedings of the 2025 Conference on Empiric...

    work page arXiv 2025

  3. [3]

    Let's Verify Step by Step

    Let’s verify step by step.Preprint, arXiv:2305.20050. MingShan Liu and Jialing Fang. 2025. Enhancing mathematical reasoning in large language models with self-consistency-based hallucination detection. Preprint, arXiv:2504.09440. Nelson Liu, Tianyi Zhang, and Percy Liang. 2023. Eval- uating verifiability in generative search engines. In Findings of the As...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  4. [4]

    Hallucination-free? assessing the reliabil- ity of leading ai legal research tools.Preprint, arXiv:2405.20362. T.J. McCabe. 1976. A complexity measure.IEEE Transactions on Software Engineering, SE-2(4):308– 320. MetaAI. 2025. The llama 4 herd: The beginning of a new era of natively multimodal ai innova- tion. https://ai.meta.com/blog/llama-4-multimodal- i...

    work page arXiv 1976

  5. [5]

    Purvis, B., Mao, Y., and Robinson, D

    Proof or bluff? evaluating llms on 2025 usa math olympiad.Preprint, arXiv:2503.21934. Long Phan, Alice Gatti, Ziwen Han, Nathaniel Li, Josephina Hu, Hugh Zhang, Chen Bo Calvin Zhang, and Mohamed Shaaban et al. 2025. Humanity’s last exam.Preprint, arXiv:2501.14249. Cheng Qian, Emre Can Acikgoz, Hongru Wang, Xiusi Chen, Avirup Sil, Dilek Hakkani-Tür, Gokhan...

    work page arXiv 2025

  6. [6]

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Niko- lay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al

    Direct preference optimization: your language model is secretly a reward model. InProceedings of the 37th International Conference on Neural In- formation Processing Systems, NIPS ’23, Red Hook, NY , USA. Curran Associates Inc. Hayley Ross, Ameya Sunil Mahabaleshwarkar, and Yoshi Suhara. 2025. When2Call: When (not) to call tools. InProceedings of the 2025...

    work page arXiv 2025

  7. [7]

    Pure Python

    Long-form factuality in large language models. InProceedings of the 38th International Conference on Neural Information Processing Systems, NeurIPS ’24, Red Hook, NY , USA. Curran Associates Inc. Shijie Xia, Xuefeng Li, Yixin Liu, Tongshuang Wu, and Pengfei Liu. 2025. Evaluating mathematical reason- ing beyond accuracy. InProceedings of the Thirty- Ninth ...

    work page arXiv 2025

  8. [8]

    Parse the gold solution into an ordered list of atomic logical steps (Step 1, Step 2, . . . ). A step is the smallest self-contained claim or transformation needed to progress the proof

    work page

  9. [9]

    Mark a step as present if the solution makes the identical deduction or provides an equivalent justification

    For each gold step, decide whether the same reasoning (possibly re-worded) appears in the given solution. Mark a step as present if the solution makes the identical deduction or provides an equivalent justification

    work page

  10. [10]

    gold_steps

    Collect all steps that are absent from the given solution. Output format(strict JSON): { "gold_steps": [ { "step": <integer>, "summary": "<one-line summary of gold step>" } ], "missing_steps": [ { "step": <integer>, "summary": "<one-line summary of gold step that is absent>" } ] } A.4 Prevalent Error Types Annotation Prompt Reasoning Error Detection Promp...

    work page 1966

  11. [11]

    cut–counting

    Upper bound.Let M be the total number of moves in the path, so it visits M+ 1 cells. Each move changes the column by ±1, so the sequence of columns forms a walk on the path graph 1–2–· · ·–7. To visit Vj dis- tinct cells in column j, one must cross the edge j−1↔j at least Vj times, while the total number of crossings of each edge is at most M. Summing ove...

    work page

  12. [12]

    cut–counting

    Construction of a 43–cell tour. One checks by explicit construction (for instance, by a backtracking computer search or by an easy hand–drawn “zig–zag”)that there exists a path of length 42, thus visiting 43 distinct cells. One such path, written as a sequence of coordinates, is (1,1)→(2,2)→(1,3)→ · · · →(7,1) →(6,2)→(5,1)→(4,2)→. . . →(6,6)→(7,7). Each s...

    work page 2023