Pith. sign in

REVIEW 3 major objections 4 minor 53 references

Can Large Models Teach Student Models to Solve Mathematical Problems Like Human Beings? A Reasoning Distillation Method via Multi-LoRA Interaction

T0 review · 3 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

Pith's one-line read The paper claims that small language models can beat prior distillation methods on GSM8K by splitting reasoning into three interacting LoRA adapters and checking agreement between two of them.

desk verdict A plausible multi-LoRA distillation recipe with a consistency loop; the posted full text is corrupted, so the strong GSM8K claims are uncheckable, and the authors need to control for inference compute and shared-teacher bias before I'd trust them. read the letter →

arxiv 2508.13037 v1 pith:N5EMFOU6 submitted 2025-08-18 cs.CL cs.AI

classification cs.CLcs.AI
keywords mathematicalreasoningdistillationsmalllanguagemodelsLoRAadaptersSystem1/2thinkingchain-of-thoughtknowledgegenerationconsistencycheckingGSM8K
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

The paper claims that small language models can be taught to solve math problems better than existing distillation methods without relying on massive teacher-generated training sets. It proposes LoRID, which trains three lightweight LoRA adapters on the student: an Intuitive Reasoner that writes chain-of-thought directly, a Knowledge Generator that extracts knowledge from a problem, and a Deep Reasoner that reasons from that knowledge. At inference, the outputs of the intuitive and deep reasoners are compared; disagreement triggers another round of generation, creating a self-checking loop inspired by System 1 and System 2 thinking. On GSM8K, the paper reports the best accuracy among the compared methods on five base models, beating the runner-up by 1.8% to 16.1%.

What carries the argument

The central mechanism is a multi-LoRA interaction. LoRA (low-rank adaptation) is a finetuning technique that trains small adapter matrices instead of the whole network. LoRID trains three such adapters on the same student backbone: IR for direct chain-of-thought, KG for knowledge extraction, and DR for knowledge-conditioned reasoning. The consistency check at inference—comparing IR's and DR's answers and iterating on disagreement—is the part that turns random generation variance into a corrective feedback loop.

What would settle it

Take a held-out set of math problems where the teacher LLM's reasoning is systematically biased or where teacher answers are deliberately wrong on a subset; if the two adapters still agree on the wrong answers, the consistency check is confirming shared bias rather than correcting error. Also ablate the consistency iteration and measure GSM8K accuracy; if accuracy barely moves, the mutual-feedback mechanism is not carrying the reported gains.

Watch

Extended reading notes

Core claim

On its own terms, the discovery is a training-and-inference recipe: instead of cramming a small model with large quantities of teacher-generated chain-of-thought data, LoRID decomposes reasoning into three cooperating low-rank adapters. The Intuitive Reasoner (IR) is trained on question-reasoning pairs to generate chain-of-thought directly; the Knowledge Generator (KG) outputs only knowledge for a given question; the Deep Reasoner (DR) consumes that knowledge to produce reasoning. At inference, the model checks whether IR and DR agree on the final answer, and if they do not, it iterates. The paper claims that this consistency loop, rather than data volume alone, is what drives the gains, yie

Load-bearing premise

The method assumes that when the Intuitive Reasoner and Deep Reasoner agree, the answer is correct, even though both were trained from the same teacher model's data and could share its blind spots.

Editorial extensions

If this is right

  • A student model can carry three specialized reasoning roles in parallel LoRA adapters, so adding a reasoning style does not require retraining the full model.
  • Knowledge extraction and reasoning are trained separately and then composed, so each component can be improved or replaced without disturbing the other.
  • Consistency checking gives a stopping rule: the model can keep generating until two reasoning paths agree, reducing the influence of a single unlucky sample.
  • The reported gains on GSM8K hold across five different base models, suggesting the recipe is not tied to one architecture.

Reading between the lines

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

  • Because both adapters are trained from the same teacher-generated data, agreement between them may be a shared-bias signal rather than independent verification; using a second teacher or a verifier for one path would test this.
  • Some of the gain may come from spending extra inference compute when outputs disagree; a baseline that samples multiple chain-of-thoughts and takes a majority vote would isolate how much the multi-LoRA structure adds beyond that compute.
  • The same knowledge-then-reason decomposition might transfer to other multi-step tasks like code debugging or proof checking, but the paper only demonstrates math word problems.
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 / 4 minor

Summary. The paper proposes LoRID, a reasoning-distillation method that trains three LoRA modules on a small student model: an Intuitive Reasoner (IR) generating direct chain-of-thought, a Knowledge Generator (KG) producing problem-specific knowledge, and a Deep Reasoner (DR) performing reasoning conditioned on that knowledge. At inference, IR and DR outputs are compared; if inconsistent, the inference process is iterated. The abstract claims state-of-the-art accuracy on GSM8K and other benchmarks, outperforming the second-best method by 2.3%, 16.1%, 2.4%, 12.3%, and 1.8% across five base models.

Significance. The dual-process (System 1/System 2) motivation is conceptually appealing, and the multi-LoRA design is computationally lightweight compared with full-model distillation. If the reported gains are real and attributable to the interaction mechanism rather than to extra inference compute, the method would be a useful contribution to low-cost mathematical reasoning. However, the supplied full text is heavily corrupted, and the visible abstract omits the experimental setup. The central SOTA claim is therefore not currently verifiable. I see no circular derivation in the method, but the consistency signal between two adapters trained on the same teacher data is a genuine correctness-risk concern that needs empirical support.

major comments (3)
  1. [Abstract / Inference loop] The abstract says that when IR and DR outputs are inconsistent, the inference process needs to be iterated. No iteration cap, termination criterion, or inference-cost matching is reported. If LoRID performs multiple paired generations per problem while the baselines are single-pass, the large GSM8K margins (2.3% to 16.1%) could reflect additional sampling / self-consistency rather than the multi-LoRA interaction. Please report the iteration distribution, maximum number of iterations, and compare with baselines under matched sampling budgets (e.g., self-consistency at k=1, 5, 10).
  2. [Intuitive Reasoner and Deep Reasoner training] Both IR and DR are trained on knowledge-augmented data generated by the same teacher LLM. Agreement between their outputs is therefore not independent evidence of correctness; shared teacher errors could produce false confirmation. The paper should report agreement rates on correct versus incorrect examples and show that consistency selects the correct answer more often than a random or confidence-based selection rule. Without this, the consistency loop is not justified as a reliable error-correction mechanism.
  3. [Full text / Experimental setup] The submitted full text is corrupted mojibake and cannot be read. I cannot verify any table, algorithm, equation, dataset split, or base-model list. No code or data are provided. This prevents technical review. A resubmission must be readable and include: base model identities, dataset split sizes, LoRA hyperparameters (rank, alpha), number of training steps, decoding temperature, number of independent runs, standard deviations, and ablations isolating the contribution of KG, DR, and the consistency loop.
minor comments (4)
  1. [Abstract] The claim 'outperforms the second-best method by 2.3%, 16.1%, ...' does not name the second-best method or state which of the five base models corresponds to each margin. Please clarify.
  2. [Introduction / Motivation] The System 1 / System 2 analogy is used as motivation. The paper should explicitly state that this is a loose analogy, not a cognitive claim, or provide a precise mapping to the proposed modules.
  3. [Knowledge Generator] The phrase 'outputs only knowledge' is ambiguous. What is the exact output format of KG, and how is that knowledge presented to DR? An example input-output pair would help.
  4. [Notation] The manuscript uses 'LoRA', 'LoRA block', and 'multi-LoRA Interaction' with inconsistent capitalization. Please standardize the notation and define all acronyms at first use.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity identified; the benchmark claims are externally evaluated and not forced by construction.

full rationale

The available evidence from the abstract describes a training pipeline (LLM-generated knowledge-enhanced data, three LoRA adapters IR/KG/DR, and a consistency-based inference loop) followed by evaluation on external benchmarks such as GSM8K. No quantity reported as a result is defined in terms of the target metric, and no fitted parameter is renamed as a prediction. The consistency check between IR and DR is an inference-time selection heuristic, not a derivation: the fact that both adapters are trained from the same teacher data raises a substantive correctness/robustness concern about false confirmation, but it does not make the benchmark outcome equivalent to the method's inputs by construction. No load-bearing self-citation, imported uniqueness theorem, ansatz-smuggling citation, or renaming of a known result appears in the available text. The full text is corrupted, so equations and baselines cannot be inspected in detail, but on the evidence available the central claim is evaluated against external benchmarks rather than against the method's own training data or definitions. Therefore no significant circularity is identified.

Assumptions & free parameters 2 free parameters · 3 assumptions · 3 invented entities

The central claim depends on two unverified domain assumptions: that teacher-generated knowledge data is the right training signal, and that agreement between two student adapters indicates correctness. The three LoRA modules are novel components but carry no independent evidence outside the paper's own experiments.

free parameters (2)
  • Multi-LoRA configuration (rank, alpha, number of training steps)
    The abstract gives no values for these hyperparameters; the method's performance could depend on them.
  • Consistency iteration cap and sampling temperature
    The iterative consistency loop requires an iteration budget and generation temperature; neither is reported in the abstract.
assumptions (3)
  • domain assumption LLM-generated knowledge-enhanced data improves SLM reasoning more than plain chain-of-thought data.
    The method's first step builds knowledge-enhanced datasets; if this premise fails, the entire pipeline loses its basis.
  • domain assumption Distilling via separate knowledge generation and deep reasoning modules is a valid proxy for System 2 human thinking.
    The paper motivates IR, KG, and DR by dual-process psychology; this is a modeling assumption, not a demonstrated fact.
  • domain assumption Consistency between two student-generated outputs is a correct answer signal.
    The iterative inference relies on agreement; if both adapters share errors, the check is uninformative.
invented entities (3)
  • Intuitive Reasoner (IR) LoRA
    purpose: Directly generate chain-of-thought solutions for problems.
    It is a trained component with no external falsifiable handle beyond the reported benchmark accuracy.
  • Knowledge Generator (KG) LoRA
    purpose: Produce knowledge statements from questions before reasoning.
    It is a trained component with no external falsifiable handle beyond the reported benchmark accuracy.
  • Deep Reasoner (DR) LoRA
    purpose: Perform reasoning conditioned on the generated knowledge.
    It is a trained component with no external falsifiable handle beyond the reported benchmark accuracy.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Can Large Models Teach Student Models to Solve Mathematical Problems Like Human Beings? A Reasoning Distillation Method via Multi-LoRA Interaction." pith.science (2026). https://pith.science/paper/N5EMFOU6

@misc{pith2026250813037,
  author       = {Pith},
  title        = {Pith review of: Can Large Models Teach Student Models to Solve Mathematical Problems Like Human Beings? A Reasoning Distillation Method via Multi-LoRA Interaction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/N5EMFOU6}},
  note         = {Machine review of arXiv:2508.13037}
}
read the original abstract

Recent studies have demonstrated that Large Language Models (LLMs) have strong mathematical reasoning abilities but rely on hundreds of billions of parameters. To tackle the challenge of poor reasoning in Small Language Models (SLMs), existing methods typically leverage LLMs to generate massive amounts of data for cramming training. In psychology, they are akin to System 1 thinking, which resolves reasoning problems rapidly based on experience and intuition. However, human learning also requires System 2 thinking, where knowledge is first acquired and then reinforced through practice. Inspired by such two distinct modes of thinking, we propose a novel method based on the multi-LoRA Interaction for mathematical reasoning Distillation (LoRID). First, we input the question and reasoning of each sample into an LLM to create knowledge-enhanced datasets. Subsequently, we train a LoRA block on the student model as an Intuitive Reasoner (IR), which directly generates Chain-of-Thoughts for problem-solving. Then, to imitate System 2 thinking, we train the Knowledge Generator (KG) and Deep Reasoner (DR), respectively. The former outputs only knowledge after receiving problems, while the latter uses that knowledge to perform reasoning. Finally, to address the randomness in the generation of IR and DR, we evaluate whether their outputs are consistent, and the inference process needs to be iterated if not. This step can enhance the mathematical reasoning ability of SLMs through mutual feedback. Experimental results show that LoRID achieves state-of-the-art performance, especially on the GSM8K dataset, where it outperforms the second-best method by 2.3%, 16.1%, 2.4%, 12.3%, and 1.8% accuracy across the five base models, respectively.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

53 extracted references · 38 canonical work pages

  1. [1]

    Gpt-4 technical report

    Josh Achiam, Steven Adler, Sandhini Agarwal, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774 , 2023

  2. [2]

    Brown, Benjamin Mann, Nick Ryder, et al

    Tom B. Brown, Benjamin Mann, Nick Ryder, et al. Language models are few-shot learners. In NeurIPS , 2020

  3. [3]

    Monte-carlo tree search: A new framework for game ai

    Guillaume Chaslot, Sander Bakkes, Istvan Szita, et al. Monte-carlo tree search: A new framework for game ai. In AAAI , 2008

  4. [4]

    Evaluating large language models trained on code

    Mark Chen, Jerry Tworek, Heewoo Jun, et al. Evaluating large language models trained on code. arXiv preprint arXiv:2107.03374 , 2021

  5. [5]

    Masked thought: Simply masking partial reasoning steps can improve mathematical reasoning learning of language models

    Changyu Chen, Xiting Wang, Ting-En Lin, et al. Masked thought: Simply masking partial reasoning steps can improve mathematical reasoning learning of language models. In ACL , 2024

  6. [6]

    Alphamath almost zero: process supervision without process

    Guoxin Chen, Minpeng Liao, Chengxi Li, et al. Alphamath almost zero: process supervision without process. In NeurIPS , 2024

  7. [7]

    Autoprm: Automating procedural supervision for multi-step reasoning via controllable question decomposition

    Zhaorun Chen, Zhuokai Zhao, Zhihong Zhu, et al. Autoprm: Automating procedural supervision for multi-step reasoning via controllable question decomposition. In NAACL , 2024

  8. [8]

    Training verifiers to solve math word problems

    Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, et al. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168 , 2021

Show all 53 references
  1. [9]

    To RA : A tool-integrated reasoning agent for mathematical problem solving

    Zhibin Gou, Zhihong Shao, Yeyun Gong, et al. To RA : A tool-integrated reasoning agent for mathematical problem solving. In ICLR , 2024

  2. [10]

    The llama 3 herd of models

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, et al. The llama 3 herd of models. arXiv e-prints , pages arXiv--2407, 2024

  3. [11]

    Measuring mathematical problem solving with the math dataset

    Dan Hendrycks, Collin Burns, Saurav Kadavath, et al. Measuring mathematical problem solving with the math dataset. In NeurIPS , 2021

  4. [12]

    Lora: Low-rank adaptation of large language models

    Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, et al. Lora: Low-rank adaptation of large language models. In ICLR , 2022

  5. [13]

    Enhancing sequential recommendation via llm-based semantic embedding learning

    Jun Hu, Wenwen Xia, Xiaolu Zhang, et al. Enhancing sequential recommendation via llm-based semantic embedding learning. In WWW , 2024

  6. [14]

    Qdmr-based planning-and-solving prompting for complex reasoning tasks

    Jinfeng Huang, Qiaoqiao She, Wenbin Jiang, et al. Qdmr-based planning-and-solving prompting for complex reasoning tasks. In COLING , 2024

  7. [15]

    Openai o1 system card

    Aaron Jaech, Adam Kalai, Adam Lerer, et al. Openai o1 system card. arXiv preprint arXiv:2412.16720 , 2024

  8. [16]

    Mistral 7b

    Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, et al. Mistral 7b. arXiv preprint arXiv:2310.06825 , 2023

  9. [17]

    Thinking, fast and slow

    Daniel Kahneman. Thinking, fast and slow. Farrar, Straus and Giroux , 2011

  10. [18]

    Large language models are zero-shot reasoners

    Takeshi Kojima, Shixiang Shane Gu, Machel Reid, et al. Large language models are zero-shot reasoners. In NeurIPS , 2022

  11. [19]

    Mugglemath: Assessing the impact of query and response augmentation on math reasoning

    Chengpeng Li, Zheng Yuan, Hongyi Yuan, et al. Mugglemath: Assessing the impact of query and response augmentation on math reasoning. In ACL , 2024

  12. [20]

    Neuro-symbolic data generation for math reasoning

    Zenan Li, Zhi Zhou, Yuan Yao, et al. Neuro-symbolic data generation for math reasoning. In NeurIPS , 2024

  13. [21]

    The flan collection: Designing data and methods for effective instruction tuning

    Shayne Longpre, Le Hou, Tu Vu, et al. The flan collection: Designing data and methods for effective instruction tuning. In ICML , 2023

  14. [22]

    Decoupled weight decay regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 , 2017

  15. [23]

    Mathgenie: Generating synthetic data with question back-translation for enhancing mathematical reasoning of llms

    Zimu Lu, Aojun Zhou, Houxing Ren, et al. Mathgenie: Generating synthetic data with question back-translation for enhancing mathematical reasoning of llms. In ACL , 2024

  16. [24]

    Wizardmath: Empowering mathematical reasoning for large language models via reinforced evol-instruct

    Haipeng Luo, Qingfeng Sun, Can Xu, et al. Wizardmath: Empowering mathematical reasoning for large language models via reinforced evol-instruct. arXiv preprint arXiv:2308.09583 , 2023

  17. [25]

    Teaching small language models to reason

    Lucie Charlotte Magister, Jonathan Mallinson, Jakub Adamek, et al. Teaching small language models to reason. In ACL , 2023

  18. [26]

    Direct preference optimization: Your language model is secretly a reward model

    Rafael Rafailov, Archit Sharma, Eric Mitchell, et al. Direct preference optimization: Your language model is secretly a reward model. In NeurIPS , 2023

  19. [27]

    Self-refine instruction-tuning for aligning reasoning in language models

    Leonardo Ranaldi and Andr \`e Freitas. Self-refine instruction-tuning for aligning reasoning in language models. In EMNLP , 2024

  20. [28]

    Deepseekmath: Pushing the limits of mathematical reasoning in open language models

    Zhihong Shao, Peiyi Wang, Qihao Zhu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300 , 2024

  21. [29]

    Large language models are in-context semantic reasoners rather than symbolic reasoners

    Xiaojuan Tang, Zilong Zheng, Jiaqi Li, et al. Large language models are in-context semantic reasoners rather than symbolic reasoners. arXiv preprint arXiv:2305.14825 , 2023

  22. [30]

    Mathscale: Scaling instruction tuning for mathematical reasoning

    Zhengyang Tang, Xingxing Zhang, Benyou Wang, et al. Mathscale: Scaling instruction tuning for mathematical reasoning. In ICML , 2024

  23. [31]

    Gemini: a family of highly capable multimodal models

    Gemini Team, Rohan Anil, Sebastian Borgeaud, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 , 2023

  24. [32]

    Dart-math: Difficulty-aware rejection tuning for mathematical problem-solving

    Yuxuan Tong, Xiwen Zhang, Rui Wang, et al. Dart-math: Difficulty-aware rejection tuning for mathematical problem-solving. In NeurIPS , 2024

  25. [33]

    Openmathinstruct-1: A 1.8 million math instruction tuning dataset

    Shubham Toshniwal, Ivan Moshkov, Sean Narenthiran, et al. Openmathinstruct-1: A 1.8 million math instruction tuning dataset. In NeurIPS , 2024

  26. [34]

    Llama 2: Open foundation and fine-tuned chat models

    Hugo Touvron, Louis Martin, Kevin Stone, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 , 2023

  27. [35]

    Zheng, Han Yu, et al

    Wei Wang, Vincent W. Zheng, Han Yu, et al. A survey of zero-shot learning: Settings, methods, and applications. ACM Trans. Intell. Syst. Technol. , 10(2):1--37, 2019

  28. [36]

    Self-consistency improves chain of thought reasoning in language models

    Xuezhi Wang, Jason Wei, Dale Schuurmans, et al. Self-consistency improves chain of thought reasoning in language models. In ICLR , 2023

  29. [37]

    Mathcoder: Seamless code integration in LLM s for enhanced mathematical reasoning

    Ke Wang, Houxing Ren, Aojun Zhou, et al. Mathcoder: Seamless code integration in LLM s for enhanced mathematical reasoning. In ICLR , 2024

  30. [38]

    Math-shepherd: Verify and reinforce llms step-by-step without human annotations

    Peiyi Wang, Lei Li, Zhihong Shao, et al. Math-shepherd: Verify and reinforce llms step-by-step without human annotations. In ACL , 2024

  31. [39]

    Self-training with direct preference optimization improves chain-of-thought reasoning

    Tianduo Wang, Shichen Li, and Wei Lu. Self-training with direct preference optimization improves chain-of-thought reasoning. In ACL , 2024

  32. [40]

    Chain-of-thought prompting elicits reasoning in large language models

    Jason Wei, Xuezhi Wang, Dale Schuurmans, et al. Chain-of-thought prompting elicits reasoning in large language models. In NeurIPS , 2022

  33. [41]

    From language modeling to instruction following: Understanding the behavior shift in LLM s after instruction tuning

    Xuansheng Wu, Wenlin Yao, Jianshu Chen, et al. From language modeling to instruction following: Understanding the behavior shift in LLM s after instruction tuning. In NAACL , 2024

  34. [42]

    Training large language models for reasoning through reverse curriculum reinforcement learning

    Zhiheng Xi, Wenxiang Chen, Boyang Hong, et al. Training large language models for reasoning through reverse curriculum reinforcement learning. In ICML , 2024

  35. [43]

    An Yang, Baosong Yang, Beichen Zhang, et al. Qwen2. 5 technical report. arXiv preprint arXiv:2412.15115 , 2024

  36. [44]

    Mumath-code: Combining tool-use large language models with multi-perspective data augmentation for mathematical reasoning

    Shuo Yin, Weihao You, Zhilong Ji, et al. Mumath-code: Combining tool-use large language models with multi-perspective data augmentation for mathematical reasoning. arXiv preprint arXiv:2405.07551 , 2024

  37. [45]

    M u M ath: Multi-perspective data augmentation for mathematical reasoning in large language models

    Weihao You, Shuo Yin, Xudong Zhao, et al. M u M ath: Multi-perspective data augmentation for mathematical reasoning in large language models. In NAACL , 2024

  38. [46]

    OVM , outcome-supervised value models for planning in mathematical reasoning

    Fei Yu, Anningzhe Gao, and Benyou Wang. OVM , outcome-supervised value models for planning in mathematical reasoning. In NAACL , 2024

  39. [47]

    Metamath: Bootstrap your own mathematical questions for large language models

    Longhui Yu, Weisen Jiang, Han Shi, et al. Metamath: Bootstrap your own mathematical questions for large language models. In ICLR , 2024

  40. [48]

    Scaling relationship on learning mathematical reasoning with large language models

    Zheng Yuan, Hongyi Yuan, Chengpeng Li, et al. Scaling relationship on learning mathematical reasoning with large language models. arXiv preprint arXiv:2308.01825 , 2023

  41. [49]

    Mammoth: Building math generalist models through hybrid instruction tuning

    Xiang Yue, Xingwei Qu, Ge Zhang, et al. Mammoth: Building math generalist models through hybrid instruction tuning. In ICLR , 2024

  42. [50]

    Accessing gpt-4 level mathematical olympiad solutions via monte carlo tree self-refine with llama-3 8b

    Di Zhang, Xiaoshui Huang, Dongzhan Zhou, et al. Accessing gpt-4 level mathematical olympiad solutions via monte carlo tree self-refine with llama-3 8b. arXiv preprint arXiv:2406.07394 , 2024

  43. [51]

    Learn beyond the answer: Training language models with reflection for mathematical reasoning

    Zhihan Zhang, Tao Ge, Zhenwen Liang, et al. Learn beyond the answer: Training language models with reflection for mathematical reasoning. In EMNLP , 2024

  44. [52]

    Solving challenging math word problems using gpt-4 code interpreter with code-based self-verification

    Aojun Zhou, Ke Wang, Zimu Lu, et al. Solving challenging math word problems using gpt-4 code interpreter with code-based self-verification. In ICLR , 2024

  45. [53]

    write newline

    " write newline "" before.all 'output.state := FUNCTION fin.entry add.period write newline FUNCTION new.block output.state before.all = 'skip after.block 'output.state := if FUNCTION new.sentence output.state after.block = 'skip output.state before.all = 'skip after.sentence '...

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.