Pith. sign in

REVIEW 3 major objections 5 minor 22 references

Self-Error-Instruct: Generalizing from Errors for LLMs Mathematical Reasoning

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read This paper tries to establish that an LLM's mathematical reasoning can be improved by mining its own wrong answers, grouping the mistakes into error types, synthesizing new practice problems for each type, and keeping only the synthetic…

desk verdict A genuinely new data-synthesis pipeline for math reasoning, with large Qwen gains, but the out-of-domain results are threatened by an unchecked leakage channel. read the letter →

arxiv 2505.22591 v1 pith:EUA3INHJ submitted 2025-05-28 cs.CL cs.AIcs.LG

classification cs.CLcs.AIcs.LG
keywords mathematicalreasoningerror-typegeneralizationsynthetictrainingdataself-instructone-shotselectionfine-tuningbadcaseanalysis
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Self-Error-Instruct (SEI) is a training-data pipeline for improving an LLM's mathematical reasoning by mining its own mistakes. The paper claims that grouping a model's incorrect answers into error types, synthesizing new practice problems for each type, and then keeping only the synthetic examples that fix bad cases without hurting good ones yields large gains: average accuracy increases of 24.94% for Qwen2.5-Math-7B, 1.72% for Llama-3-8B-Instruct, and 0.98% for Mathstral-7B-v0.1 across six math benchmarks. The central bet is that generalizing from error types, rather than from isolated bad cases, produces training data that transfers to unseen problems. If true, this means small, targeted synthetic datasets can beat full training-set fine-tuning for math reasoning.

What carries the argument

The load-bearing object is the SEI pipeline: bad-case extraction by answer mismatch; error-keyphrase generation by an instructor model; clustering keyphrases into error types; error-type-specific data synthesis in a self-instruct style using a few sampled bad cases as prompts; one-shot learning selection on a 100-case dev set; and iterative LoRA fine-tuning. The one-shot selector is the component doing the paper's main work: it ranks each synthetic sample by the number of dev cases (50 bad plus 50 good) whose answer is correct when the sample is prepended as an exemplar, so selected data must both repair weaknesses and preserve existing competence. The pipeline also filters synthetic data against GSM8K and MATH train/test sets using Rouge-L similarity to avoid leakage.

What would settle it

Take the 30,000 generated samples, resample the 100-case dev set multiple times, and rerun the one-shot selection; if the top-5% subset chosen under different dev sets does not consistently beat random selection on the six test sets, the selection mechanism is not the source of the reported gains.

Watch

Extended reading notes

Core claim

The paper's central claim is that error-type generalization, not the individual bad cases themselves, drives the improvement. Starting from a target model's wrong answers on GSM8K and MATH training problems, SEI has an instructor model summarize each mistake as a keyphrase, clusters those keyphrases into named error types, and then asks the instructor model to generate new problems that exercise each type. The generated set is filtered and scored: each candidate synthetic example is used as a one-shot prompt over 100 validation cases so that a score counts how many previously wrong cases it fixes and how many originally correct cases it preserves. The top 5% of 30,000 generated examples is used to fine-tune the target model, and the loop repeats for three iterations. The paper reports that this curated training set outperforms training on the raw math corpus, on the bad cases alone, and on data from prior single-case error-generation baselines, while also beating the full synthesized dataset.

Load-bearing premise

The one-shot selection score computed on a fixed 100-case dev set, split evenly between good and bad cases, reliably predicts which synthetic examples will most improve the target model after fine-tuning.

Editorial extensions

If this is right

  • Training on error-type-generalized data improves both in-domain (GSM8K, MATH) and out-of-domain (TAL, GaoKao, SAT, College) math tests, so the effect is not just memorizing benchmark formats.
  • Raw training data and bad-case-only training can degrade already instruction-tuned Llama3 and Mathstral, so the synthesized data, not the original corpus, carries the gain.
  • One-shot learning selection outperforms random selection, gradient-based LESS, and even the full synthesized dataset, implying active selection matters more than data volume.
  • From-scratch training on all selected data across three iterations beats iteratively training round by round, consistent with small per-round data causing overfitting.
  • The bad-case fix rate rises with iterations, indicating the method keeps finding addressable weaknesses rather than exhausting them after one pass.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The same recipe likely transfers beyond math: any setting where a target model emits checkable outputs and errors can be clustered, such as code generation, logic puzzles, or structured extraction, could use error-type generalization.
  • The one-shot selection score is a cheap proxy for fine-tuning gain; a testable extension is to replace the fixed 100-case dev set with a model-adaptive or online dev set that updates as errors are fixed.
  • Because the instructor model does all keyphrase extraction, clustering, and synthesis, the approach inherits the instructor's ceiling; a promising experiment is to measure how performance degrades with weaker or open-weight instructors.
  • The reported advantage over gradient-based selection could be sensitive to dev-set size and sampling; replicating selection with multiple random 100-case dev sets and multiple fine-tuning seeds would tell whether the selection signal is stable.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper presents Self-Error-Instruct (SEI), a four-stage framework for improving LLM mathematical reasoning by synthesizing training data that generalizes from a target model's error types rather than from individual bad cases. Bad cases are extracted from the GSM8K/MATH training sets; GPT-4o generates error keyphrases, they are clustered into error types, and targeted synthetic data are generated via a self-instruct approach. A one-shot learning score computed on a 100-example dev set selects the top 5% of generated data, and the target model is continued with LoRA over three iterations. Experiments on Llama-3-8B-Instruct, Qwen2.5-Math-7B, and Mathstral-7B-v0.1 report in-domain and out-of-domain gains, with Qwen improving substantially and Llama3/Mathstral gaining about 1.7% and 1.0% on average. The paper also compares selection strategies, training schemes, and data volumes.

Significance. If the results are robust, the proposed error-type generalization and one-shot selection are useful and practical contributions. Strengths include a clearly specified pipeline, public code, several baselines and ablations, out-of-domain evaluation, and analyses of fix rates and training strategies. The one-shot selection criterion that rewards both fixing bad cases and preserving good cases is a reasonable and potentially transferable idea. However, the strength of the empirical claims is weakened by the absence of uncertainty quantification and by an unmeasured risk of overlap between GPT-4o-generated data and the out-of-domain test sets.

major comments (3)
  1. [Section 4.1, Table 2, Figure 3] The out-of-domain generalization claim is exposed to a test-set contamination risk that is not measured. Section 4.1 states that generated data are filtered with Rouge-L > 0.7 only against the GSM8K and MATH training and test sets; no overlap check is reported between the 1,500 selected training samples and the TAL-SCQ, GaoKaoBench-Math, SAT-MATH, or CollegeMath test sets. Because the instructor model GPT-4o is likely trained on these public benchmarks, and because the out-of-domain gains for Qwen2.5-Math-7B are large (TAL +26.14, GaoKao +42.13, SAT +29.40, College +11.46 in Figure 3), the results could reflect memorized near-duplicates rather than genuine generalization. The authors should report max Rouge-L or embedding similarity between each selected synthetic sample and each out-of-domain test item, and ideally show that the gains persist on the least-similar subset of out-of-domain examples.
  2. [Section 5.1, Table 2] No error bars, confidence intervals, or significance tests are reported, and the claimed improvements for two of the three models are small: Llama-3-8B-Instruct improves by 1.72% and Mathstral by 0.98% on average. These margins are likely within run-to-run noise for greedy-decoding evaluations of fine-tuned models, so the statement that SEI-ICL 'outperforms others by substantial margins in all math datasets' is not supported for all models. The authors should provide multiple seeds or bootstrap estimates for the main table, and at least clarify which comparisons are stable.
  3. [Section 3.3, Section 5.2] The one-shot selection method is evaluated on a single fixed dev set of 100 examples (50 bad and 50 good cases), and this same dev set is used to select 1,500 training samples. The claim that one-shot ICL selection outperforms random selection and LESS could be driven by overfitting to this small dev set, and no stability analysis is given. The authors should repeat the selection with multiple dev-set resamples, or report the variance of the selected data and of the downstream performance.
minor comments (5)
  1. [Section 2.1, References] The text attributes LLM2LLM to 'Tong et al., 2024a' in two places (Sections 2.1 and 4.4), but the reference list gives Tong et al. 2024a as a paper on backdoor attacks; the correct citation for LLM2LLM is Lee et al. 2024.
  2. [Table 2] The table contains formatting errors, including missing spaces and merged numbers such as '15.7935.59', '38.2415.79', and '25.19 TAL', which should be corrected.
  3. [Equations (4)-(5)] The notation in Equations (4) and (5) is ambiguous: the superscript and subscript on r_i^j are not clearly indexed to the synthetic sample and the dev-set query, and the symbol for the model-generated reasoning path is not introduced before use.
  4. [Section 4.4, Table 2] The phrase 'For fair comparison, the generalized data sizes for the baselines are kept consistent with SEI-ICL' is unclear for the Training Data and Bad Cases baselines, which appear to use the full training set rather than 1,500 samples; the intended data sizes should be stated explicitly.
  5. [Figure 4 and Section 5.5] The x-axis label 'Synthetic Data Size' is ambiguous: it should be stated whether the plotted performance corresponds to the full generated set or to a subset selected by the one-shot criterion, and whether the same selection pipeline was applied at each size.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the method is an empirical pipeline whose key claims are tested on held-out benchmarks, and its selection criterion is fit to a training-derived dev set, not to the evaluation labels.

full rationale

The paper's central claim is that data synthesized and selected through error-type generalization improves mathematical reasoning on in-domain and out-of-domain test sets. Nothing in the claimed derivation chain makes a test-set result equivalent to an input by construction. The bad cases are defined by answer mismatch on the GSM8K and MATH training sets (Eq. 1); error keyphrases and clusters are generated by GPT-4o from those cases (Eqs. 2-3); synthetic data are filtered with Rouge-L against GSM8K and MATH; and the one-shot selection score (Eqs. 4-5) is computed on a 100-sample dev set drawn from the training data. The final evaluation uses held-out GSM8K/MATH test splits and four separate OOD benchmarks, none of which enter the selection score or the Rouge-L filter. The dev-set-based selection is self-referential in the weak sense that it optimizes a proxy on training data, but the reported gains are measured on independent test labels, so no prediction reduces to the fitted input. The paper's self-citations in related work and appendix are descriptive rather than load-bearing, and the Limitations section candidly reports cost, dataset-scope, and time constraints rather than hiding a circular step. The possible OOD test-set leakage raised by the reader is a fairness/robustness concern, not a circularity concern: no quoted equation makes the OOD improvements equal to the filtering or selection criterion. Thus the appropriate finding is no significant circularity.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The framework has no new physical or mathematical entities. The free parameters are experimental choices for data synthesis and selection. The key domain assumptions concern the reliability of GPT-4o's error analysis, the predictive power of the one-shot selection score, and the sufficiency of the Rouge-L filter for preventing test leakage.

free parameters (5)
  • top_pct_selection = 5%
    Main results use the top 5% of synthetic data; Table 3 shows 5% is best among 5%, 10%, and 20%, but this choice appears to be made after inspecting results.
  • rouge_l_threshold = 0.7
    Filter threshold to remove data similar to GSM8K and MATH train/test sets; chosen by hand and may not prevent semantic test-set leakage.
  • samples_per_error_type = 5 bad cases + 3 generated examples
    Number of in-context examples used during self-instruct synthesis, chosen by hand.
  • iterations = 3
    Number of iterative rounds for data synthesis and model training, chosen by hand.
  • dev_set_size = 100 (50 good, 50 bad)
    Validation set for one-shot selection, chosen by hand.
assumptions (4)
  • domain assumption GPT-4o can accurately identify error keyphrases and cluster them into meaningful, non-overlapping error types.
    The entire self-error-instruct step relies on GPT-4o's error analysis; if keyphrases are wrong, the synthesized data is mis-targeted. Entered in Section 3.2.
  • domain assumption One-shot score on a 100-case dev set is a reliable proxy for data quality in fine-tuning.
    Used for selection in Section 3.3; the rank correlation with actual fine-tuning utility is not established beyond the paper's own experiments.
  • domain assumption Rouge-L filtering with threshold 0.7 prevents test-set contamination.
    Section 4.1; GPT-4o may generate problems semantically similar to public test sets without exceeding this threshold.
  • domain assumption Fine-tuning with LoRA on 1,500 synthetic examples will improve math reasoning without harming prior capabilities.
    Underlies the iterative training and from-scratch training in Sections 3.4 and 5.4.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Self-Error-Instruct: Generalizing from Errors for LLMs Mathematical Reasoning." pith.science (2026). https://pith.science/paper/EUA3INHJ

@misc{pith2026250522591,
  author       = {Pith},
  title        = {Pith review of: Self-Error-Instruct: Generalizing from Errors for LLMs Mathematical Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EUA3INHJ}},
  note         = {Machine review of arXiv:2505.22591}
}
read the original abstract

Although large language models demonstrate strong performance across various domains, they still struggle with numerous bad cases in mathematical reasoning. Previous approaches to learning from errors synthesize training data by solely extrapolating from isolated bad cases, thereby failing to generalize the extensive patterns inherent within these cases. This paper presents Self-Error-Instruct (SEI), a framework that addresses these model weaknesses and synthesizes more generalized targeted training data. Specifically, we explore a target model on two mathematical datasets, GSM8K and MATH, to pinpoint bad cases. Then, we generate error keyphrases for these cases based on the instructor model's (GPT-4o) analysis and identify error types by clustering these keyphrases. Next, we sample a few bad cases during each generation for each identified error type and input them into the instructor model, which synthesizes additional training data using a self-instruct approach. This new data is refined through a one-shot learning process to ensure that only the most effective examples are kept. Finally, we use these curated data to fine-tune the target model, iteratively repeating the process to enhance performance. We apply our framework to various models and observe improvements in their reasoning abilities across both in-domain and out-of-domain mathematics datasets. These results demonstrate the effectiveness of self-error instruction in improving LLMs' mathematical reasoning through error generalization.

Figures

Figures reproduced from arXiv: 2505.22591 by the authors.

Figure 1
Figure 1. The left table shows some error types of Qwen2.5-Math-7B on Math and GSM8K training set, while the [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. An overview of our Self-Error-Instruct framework. It consists of four key steps: (1) [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. The effects of two one-shot ICL strategies on [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Comparison of GSM8K and MATH performance under different synthetic data sizes. [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Prompt for Generating Error Keyphrases. Error Keyphrases Clustering Prompt: You are an expert in error analysis and categorization. You will be given a list of error keyphrases. Your task is to: 1. Analyze the given error keyphrases and identify common themes or patter…
Figure 6
Figure 6. Figure 6: Prompt for Clustering Error Keyphrases [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]
Figure 7
Figure 7. Figure 7: One-Shot Learning Prompt for Selecting Syn [PITH_FULL_IMAGE:figures/full_fig_p015_7.png]
Figure 9
Figure 9. Figure 9: Prompt for MATH Error Type-Specific Data [PITH_FULL_IMAGE:figures/full_fig_p015_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

22 extracted references · 4 canonical work pages

  1. [3]

    Wapiti: A water- mark for finetuned open-source llms.arXiv preprint arXiv:2410.06467. Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman

  2. [5]

    Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt

    The llama 3 herd of models.Preprint, arXiv:2407.21783. Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt

  3. [6]

    Edward J

    Measuring mathematical problem solving with the math dataset.Preprint, arXiv:2103.03874. Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen

  4. [7]

    Albert Q

    Lora: Low-rank adaptation of large language models.Preprint, arXiv:2106.09685. Albert Q. Jiang, Alexandre Sablayrolles, Arthur Men- sch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guil- laume Lample, Lucile Saulnier, Lélio Renard Lavaud, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thom...

  5. [9]

    InFindings of the Association for Computational Linguistics: ACL 2024, pages 6498–6526, Bangkok, Thailand

    LLM2LLM: Boosting LLMs with novel iterative data enhancement. InFindings of the Association for Computational Linguistics: ACL 2024, pages 6498–6526, Bangkok, Thailand. Associ- ation for Computational Linguistics. Ming Li, Yong Zhang, Zhitao Li, Jiuhai Chen, Lichang Chen, Ning Cheng, Jianzong Wang, Tianyi Zhou, and Jing Xiao. 2024a. From quantity to quali...

  6. [10]

    On llms- driven synthetic data generation, curation, and evalu- ation: A survey.Preprint, arXiv:2406.15126. OpenAI. 2024a. Gpt-4 technical report.Preprint, arXiv:2303.08774. OpenAI. 2024b. Gpt-4o. OpenAI

  7. [11]

    https://openai.com/o1/

    O1 Model. https://openai.com/o1/. Accessed: 2024-12-11. Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul F Christiano, Jan Leike, and Ryan Lowe

  8. [12]

    Deepseekmath: Pushing the limits of mathemati- cal reasoning in open language models.Preprint, arXiv:2402.03300. TAL

Show all 22 references
  1. [13]

    Zhengyang Tang, Xingxing Zhang, Benyou Wang, and Furu Wei

    Does synthetic data generation of llms help clinical text mining?Preprint, arXiv:2303.04360. Zhengyang Tang, Xingxing Zhang, Benyou Wang, and Furu Wei

  2. [14]

    Gemini Team

    Mathscale: Scaling instruc- tion tuning for mathematical reasoning.Preprint, arXiv:2403.02884. Gemini Team

  3. [15]

    Terry Tong, Qin Liu, Jiashu Xu, and Muhao Chen

    Gemini: A family of highly capa- ble multimodal models.Preprint, arXiv:2312.11805. Terry Tong, Qin Liu, Jiashu Xu, and Muhao Chen. 2024a. Securing multi-turn conversational language models from distributed backdoor attacks. InFind- ings of the Association for Computational Lin...

  4. [16]

    InFind- ings of the Association for Computational Linguistics: ACL 2024, pages 15496–15523, Bangkok, Thailand

    Knowledge-infused prompting: Assessing and advancing clinical text data generation with large language models. InFind- ings of the Association for Computational Linguistics: ACL 2024, pages 15496–15523, Bangkok, Thailand. Association for Computational Linguistics. Boyang Xue, ...

  5. [17]

    Preprint, arXiv:2310.08372

    Improving fac- tual consistency for knowledge-grounded dialogue systems via knowledge enhancement and alignment. Preprint, arXiv:2310.08372. An Yang, Beichen Zhang, Binyuan Hui, Bofei Gao, Bowen Yu, Chengpeng Li, Dayiheng Liu, Jian- hong Tu, Jingren Zhou, Junyang Lin, Keming L...

  6. [18]

    Jiahao Ying, Mingbao Lin, Yixin Cao, Wei Tang, Bo Wang, Qianru Sun, Xuanjing Huang, and Shuicheng Yan

    Qwen2.5-math tech- nical report: Toward mathematical expert model via self-improvement.Preprint, arXiv:2409.12122. Jiahao Ying, Mingbao Lin, Yixin Cao, Wei Tang, Bo Wang, Qianru Sun, Xuanjing Huang, and Shuicheng Yan

  7. [19]

    InFindings of the Association for Compu- tational Linguistics: EMNLP 2024, pages 11185– 11208, Miami, Florida, USA

    LLMs-as-instructors: Learn- ing from errors toward automating model improve- ment. InFindings of the Association for Compu- tational Linguistics: EMNLP 2024, pages 11185– 11208, Miami, Florida, USA. Association for Com- putational Linguistics. Erxin Yu, Jing Li, Ming Liao, Siq...

  8. [20]

    Wanjun Zhong, Ruixiang Cui, Yiduo Guo, Yaobo Liang, Shuai Lu, Yanlin Wang, Amin Saied, Weizhu Chen, and Nan Duan

    Evaluating the performance of large language models on gaokao benchmark.Preprint, arXiv:2305.12474. Wanjun Zhong, Ruixiang Cui, Yiduo Guo, Yaobo Liang, Shuai Lu, Yanlin Wang, Amin Saied, Weizhu Chen, and Nan Duan

  9. [21]

    In Findings of the Association for Computational Lin- guistics: NAACL 2024, pages 2299–2314, Mexico City, Mexico

    AGIEval: A human-centric benchmark for evaluating foundation models. In Findings of the Association for Computational Lin- guistics: NAACL 2024, pages 2299–2314, Mexico City, Mexico. Association for Computational Lin- guistics. Chunting Zhou, Pengfei Liu, Puxin Xu, Srini Iyer,...

  10. [22]

    Cluster name

    Lima: Less is more for alignment.Preprint, arXiv:2305.11206. A Overview of Prompts Used A.1 Prompt for Training and Inference For all the models, we use the built-in chat tem- plates for training and inference. Figure 7 illus- trates the one-shot learning prompt for the Qwen2....

  11. [2020]

    Lingjie Chen, Ruizhong Qiu, Siyu Yuan, Zhining Liu, Tianxin Wei, Hyunsik Yoo, Zhichen Zeng, Deqing Yang, and Hanghang Tong

    Language models are few-shot learn- ers.Preprint, arXiv:2005.14165. Lingjie Chen, Ruizhong Qiu, Siyu Yuan, Zhining Liu, Tianxin Wei, Hyunsik Yoo, Zhichen Zeng, Deqing Yang, and Hanghang Tong

  12. [2021]

    Training verifiers to solve math word prob- lems.Preprint, arXiv:2110.14168. Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al- Dahle, Aiesha Letman, Akhil Mathur, Alan Schel- ten, Alex Vaughan, Amy Yang, Angela Fan, Anirudh Goyal, A...

  13. [2023]

    Nicholas Lee, Thanakul Wattanawong, Sehoon Kim, Karttikeya Mangalam, Sheng Shen, Gopala Anu- manchipalli, Michael Mahoney, Kurt Keutzer, and Amir Gholami

    Mistral 7b.Preprint, arXiv:2310.06825. Nicholas Lee, Thanakul Wattanawong, Sehoon Kim, Karttikeya Mangalam, Sheng Shen, Gopala Anu- manchipalli, Michael Mahoney, Kurt Keutzer, and Amir Gholami

  14. [2024]

    Anthropic

    Learning from mistakes makes llm better reasoner.Preprint, arXiv:2310.20689. Anthropic

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.