mathematical-reasoning

ToRA is a series of Tool-integrated Reasoning LLM Agents designed to solve challenging mathematical reasoning problems by interacting with tools [ICLR'24].

Resources of deep learning for mathematical reasoning (DL4MATH).

[NeurIPS 2024] Code for the paper "Diffusion of Thoughts: Chain-of-Thought Reasoning in Diffusion Language Models"

[ICLR 2026]🚀ReVisual-R1 is a 7B open-source multimodal language model that follows a three-stage curriculum—cold-start pre-training, multimodal reinforcement learning, and text-only reinforcement learning—to achieve faithful, concise, and self-reflective state-of-the-art performance in visual and textual reasoning.

[ICLR 2026] Harder Is Better: Boosting Mathematical Reasoning via Difficulty-Aware GRPO and Multi-Aspect Question Reformulation

Small and Efficient Mathematical Reasoning LLMs

[ACL'24] Code and data of paper "When is Tree Search Useful for LLM Planning? It Depends on the Discriminator"

[EMNLP '23] Discriminator-Guided Chain-of-Thought Reasoning

[NeurIPS 2025 D&B Track] Evaluation Code Repo for Paper "PolyMath: Evaluating Mathematical Reasoning in Multilingual Contexts"

[NeurIPS 2024] Code and Data Repo for Paper "Embedding Trajectory for Out-of-Distribution Detection in Mathematical Reasoning"

[ACL 2024 Findings] The official repo for "ConceptMath: A Bilingual Concept-wise Benchmark for Measuring Mathematical Reasoning of Large Language Models".

The lecture notes for my discrete mathematics classes.

MathPrompter Implementation: This repository hosts an implementation based on the 'MathPrompter: Mathematical Reasoning Using Large Language Models' paper by Microsoft Research. The code replicates the methods discussed in the paper.

[NeurIPS'25] Beyond Accuracy: Dissecting Mathematical Reasoning for LLMs Under Reinforcement Learning

[AAAI 2025] Assessing the Creativity of LLMs in Proposing Novel Solutions to Mathematical Problems

[AAAI 2026] The official implementation of the paper "BayesAgent: Bayesian Agentic Reasoning Under Uncertainty via Verbalized Probabilistic Graphical Modeling".

Fuzzy reasoning of Generalized Quantifiers (EMNLP 2024)

1.12B-parameter math reasoning model trained from scratch. Full end-to-end pipeline: tokenizer → TPU pretrain → SFT → GRPO → eval — built and documented by hand.

Official implementation for "Enhancing Mathematical Reasoning in LLMs by Stepwise Correction" (ACL 2025 oral)

Learning at the Right Pace: Adaptive Data Scheduling improves LLM reinforcement learning with semantic-cluster and policy-boundary sampling.