From Next-Token to Mathematics: The Learning Dynamics of Mathematical Reasoning in Language Models

• URL: http://arxiv.org/abs/2407.00900v3
• Date: Tue, 14 Oct 2025 22:32:31 GMT
• Title: From Next-Token to Mathematics: The Learning Dynamics of Mathematical Reasoning in Language Models
• Authors: Shubhra Mishra, Gabriel Poesia, Noah D. Goodman,
• Abstract summary: Large Language Models (LLMs) solely trained on next-token prediction learn to solve a wide range of problems involving mathematical reasoning.<n>We show the first analysis of how mathematical reasoning abilities of several open-weight LLMs develop during pre-training and post-training.
• Score: 38.71041354422434
• License: http://creativecommons.org/licenses/by-sa/4.0/
• Abstract: Large Language Models (LLMs) solely trained on next-token prediction learn to solve a wide range of problems involving mathematical reasoning. But how does this ability evolve during training? We show the first analysis of how mathematical reasoning abilities of several open-weight LLMs develop during pre-training and post-training. To this end, we construct MathCAMPS, a synthetic dataset of novel mathematical reasoning problems grounded in 44 fine-grained skills taken from the Common Core curriculum from K to 8th grades. In one experiment, we show that mathematical skills are learned during pre-training in an order that measurably correlates with the human-designed curriculum, even though training data are randomly ordered. We also show a detailed analysis of which mathematical abilities benefit from instruction tuning, a widely used post-training method and, in contrast, which skills suffer. Our work paves the way for an empirical understanding of LLM training dynamics in relation to reasoning.

Related papers

• MathSmith: Towards Extremely Hard Mathematical Reasoning by Forging Synthetic Problems with a Reinforced Policy [43.86485569038631]
  MathSmith is a novel framework for challenging mathematical problems to enhance LLM reasoning.<n>Rather than modifying existing problems, MathSmith constructs new ones from scratch by randomly sampling concept-explanation pairs from PlanetMath.<n>To increase difficulty, we design nine predefined strategies as soft constraints during rationales.<n>Experiments show MathSmith consistently outperforms existing baselines under both short and long CoT settings.
  arXiv  Detail & Related papers  (2025-08-07T17:32:14Z)
• ClozeMath: Improving Mathematical Reasoning in Language Models by Learning to Fill Equations [29.51572057789961]
  We propose a new approach named ClozeMath to fine-tune large language models for mathematical reasoning.<n>Our ClozeMath involves a text-infilling task that predicts masked equations from a given solution, analogous to cloze exercises used in human learning.<n> Experiments on GSM8K, MATH, and GSM-Symbolic show that ClozeMath surpasses the strong baseline Masked Thought in performance and robustness.
  arXiv  Detail & Related papers  (2025-06-04T09:27:21Z)
• Challenging the Boundaries of Reasoning: An Olympiad-Level Math Benchmark for Large Language Models [86.45058529521258]
  OlymMATH is a novel Olympiad-level mathematical benchmark designed to rigorously test the complex reasoning capabilities of LLMs. OlymMATH features 200 meticulously curated problems, each manually verified and available in parallel English and Chinese versions.
  arXiv  Detail & Related papers  (2025-03-27T11:20:17Z)
• MetaLadder: Ascending Mathematical Solution Quality via Analogical-Problem Reasoning Transfer [37.81465564673498]
  Large Language Models (LLMs) have demonstrated promising capabilities in solving mathematical reasoning tasks. We propose textbfMetaLadder, a framework that explicitly prompts LLMs to recall and reflect on meta-problems. Our experiments on mathematical benchmarks demonstrate that our MetaLadder significantly boosts LLMs' problem-solving accuracy.
  arXiv  Detail & Related papers  (2025-03-19T04:36:35Z)
• PromptCoT: Synthesizing Olympiad-level Problems for Mathematical Reasoning in Large Language Models [59.920971312822736]
  We introduce PromptCoT, a novel approach for automatically generating high-quality Olympiad-level math problems. The proposed method synthesizes complex problems based on mathematical concepts and the rationale behind problem construction. Our method is evaluated on standard benchmarks including GSM8K, MATH-500, and AIME2024, where it consistently outperforms existing problem generation methods.
  arXiv  Detail & Related papers  (2025-03-04T06:32:30Z)
• MathConstruct: Challenging LLM Reasoning with Constructive Proofs [0.9320657506524149]
  mc is a new benchmark of 126 challenging problems sourced from various math competitions. mc is suitable for Large Language Models evaluation, as solution correctness can be easily verified.
  arXiv  Detail & Related papers  (2025-02-14T14:44:22Z)
• MATH-Perturb: Benchmarking LLMs' Math Reasoning Abilities against Hard Perturbations [90.07275414500154]
  We observe significant performance drops on MATH-P-Hard across various models. We also raise concerns about a novel form of memorization where models blindly apply learned problem-solving skills.
  arXiv  Detail & Related papers  (2025-02-10T13:31:46Z)
• Mathfish: Evaluating Language Model Math Reasoning via Grounding in Educational Curricula [25.549869705051606]
  We investigate whether language models' (LMs) mathematical abilities can discern skills and concepts enabled by math content. We develop two tasks for evaluating LMs' abilities to assess math problems. We find that LMs struggle to tag and verify standards linked to problems, and instead predict labels that are close to ground truth, but differ in subtle ways.
  arXiv  Detail & Related papers  (2024-08-08T05:28:34Z)
• AI-Assisted Generation of Difficult Math Questions [78.7547836422727]
  Current training positions mathematical reasoning as a core capability. There is unmet demand for diverse and challenging math questions. We present a design framework that combines the strengths of LLMs with a human-in-the-loop approach.
  arXiv  Detail & Related papers  (2024-07-30T17:55:36Z)
• MathOdyssey: Benchmarking Mathematical Problem-Solving Skills in Large Language Models Using Odyssey Math Data [20.31528845718877]
  Large language models (LLMs) have significantly advanced natural language understanding and demonstrated strong problem-solving abilities. This paper investigates the mathematical problem-solving capabilities of LLMs using the newly developed "MathOdyssey" dataset.
  arXiv  Detail & Related papers  (2024-06-26T13:02:35Z)
• Lean Workbook: A large-scale Lean problem set formalized from natural language math problems [50.22847430754973]
  Large language models are not good at math theorem proving using formal languages like Lean. A significant challenge in this area is the scarcity of training data available in these formal languages. We propose a novel pipeline that iteratively generates and filters synthetic data to translate natural language mathematical problems into Lean 4 statements.
  arXiv  Detail & Related papers  (2024-06-06T08:25:43Z)
• Learning Beyond Pattern Matching? Assaying Mathematical Understanding in LLMs [58.09253149867228]
  This paper assesses the domain knowledge of LLMs through its understanding of different mathematical skills required to solve problems. Motivated by the use of LLMs as a general scientific assistant, we propose textitNTKEval to assess changes in LLM's probability distribution. Our systematic analysis finds evidence of domain understanding during in-context learning. Certain instruction-tuning leads to similar performance changes irrespective of training on different data, suggesting a lack of domain understanding across different skills.
  arXiv  Detail & Related papers  (2024-05-24T12:04:54Z)
• MathBench: Evaluating the Theory and Application Proficiency of LLMs with a Hierarchical Mathematics Benchmark [82.64129627675123]
  MathBench is a new benchmark that rigorously assesses the mathematical capabilities of large language models. MathBench spans a wide range of mathematical disciplines, offering a detailed evaluation of both theoretical understanding and practical problem-solving skills.
  arXiv  Detail & Related papers  (2024-05-20T17:52:29Z)
• Mathify: Evaluating Large Language Models on Mathematical Problem Solving Tasks [34.09857430966818]
  We introduce an extensive mathematics dataset called "MathQuest" sourced from the 11th and 12th standard Mathematics NCERT textbooks. We conduct fine-tuning experiments with three prominent large language models: LLaMA-2, WizardMath, and MAmmoTH. Our experiments reveal that among the three models, MAmmoTH-13B emerges as the most proficient, achieving the highest level of competence in solving the presented mathematical problems.
  arXiv  Detail & Related papers  (2024-04-19T08:45:42Z)
• GSM-Plus: A Comprehensive Benchmark for Evaluating the Robustness of LLMs as Mathematical Problem Solvers [68.77382332826167]
  Large language models (LLMs) have achieved impressive performance across various mathematical reasoning benchmarks. One essential and frequently occurring evidence is that when the math questions are slightly changed, LLMs can behave incorrectly. This motivates us to evaluate the robustness of LLMs' math reasoning capability by testing a wide range of question variations.
  arXiv  Detail & Related papers  (2024-02-29T15:26:14Z)
• InternLM-Math: Open Math Large Language Models Toward Verifiable Reasoning [98.53491178426492]
  We open-source our math reasoning LLMs InternLM-Math which is continue pre-trained from InternLM2. We unify chain-of-thought reasoning, reward modeling, formal reasoning, data augmentation, and code interpreter in a unified seq2seq format. Our pre-trained model achieves 30.3 on the MiniF2F test set without fine-tuning.
  arXiv  Detail & Related papers  (2024-02-09T11:22:08Z)
• CHAMP: A Competition-level Dataset for Fine-Grained Analyses of LLMs' Mathematical Reasoning Capabilities [25.857946070979576]
  Concept and Hint-Annotated Math Problems (CHAMP) consists of high school math competition problems annotated with concepts. This benchmark is difficult, with the best model only scoring 58.1% in standard settings. We find that models often arrive at the correct final answer through wrong reasoning steps.
  arXiv  Detail & Related papers  (2024-01-13T03:18:16Z)
• Learning Multi-Step Reasoning by Solving Arithmetic Tasks [6.398022050054328]
  This work investigates how to incorporate relatively small Language Models with the capabilities of multi-step reasoning. We propose to inject such abilities by continually pre-training LMs on a synthetic dataset MsAT. Our experiments on four math word problem datasets show the effectiveness of the proposed method.
  arXiv  Detail & Related papers  (2023-06-02T17:29:22Z)
• Towards a Holistic Understanding of Mathematical Questions with Contrastive Pre-training [65.10741459705739]
  We propose a novel contrastive pre-training approach for mathematical question representations, namely QuesCo. We first design two-level question augmentations, including content-level and structure-level, which generate literally diverse question pairs with similar purposes. Then, to fully exploit hierarchical information of knowledge concepts, we propose a knowledge hierarchy-aware rank strategy.
  arXiv  Detail & Related papers  (2023-01-18T14:23:29Z)
• A Survey of Deep Learning for Mathematical Reasoning [71.88150173381153]
  We review the key tasks, datasets, and methods at the intersection of mathematical reasoning and deep learning over the past decade. Recent advances in large-scale neural language models have opened up new benchmarks and opportunities to use deep learning for mathematical reasoning.
  arXiv  Detail & Related papers  (2022-12-20T18:46:16Z)
• JiuZhang: A Chinese Pre-trained Language Model for Mathematical Problem Understanding [74.12405417718054]
  This paper aims to advance the mathematical intelligence of machines by presenting the first Chinese mathematical pre-trained language model(PLM) Unlike other standard NLP tasks, mathematical texts are difficult to understand, since they involve mathematical terminology, symbols and formulas in the problem statement. We design a novel curriculum pre-training approach for improving the learning of mathematical PLMs, consisting of both basic and advanced courses.
  arXiv  Detail & Related papers  (2022-06-13T17:03:52Z)
• SMART: A Situation Model for Algebra Story Problems via Attributed Grammar [74.1315776256292]
  We introduce the concept of a emphsituation model, which originates from psychology studies to represent the mental states of humans in problem-solving. We show that the proposed model outperforms all previous neural solvers by a large margin while preserving much better interpretability.
  arXiv  Detail & Related papers  (2020-12-27T21:03:40Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.