math-reasoning

Implement a reasoning LLM in PyTorch from scratch, step by step

(ArXiv25) Vision Matters: Simple Visual Perturbations Can Boost Multimodal Math Reasoning

Solving Inequality Proofs with Large Language Models.

An official implementation of "SPARK: Synergistic Policy And Reward Co-Evolving Framework"

Capable, auditable coding that runs fully offline on a 16 GB machine. A verification-first layer (hard test execution, symbolic checking, agentic repair) that takes a local 7B to parity with its 671B teacher on verifiable tasks. MIT, pre-registered, reproducible.

Vietnamese math reasoning distillation from Phi-3 to TinyLlama-1.1B using QLoRA and judge-filtered reasoning traces.

Reproducible math reasoning eval with resumable generation and separate scoring. Swap parsers without re-calling models.

Prompt-engineering study on LLM math reasoning (GSM8K) and code generation (HumanEval): zero/few-shot, self-consistency, self-verification, and experiments on prompt quality, complexity, demonstrations, and diversity.

ArmLLM 2025 solutions covering ViT from scratch, SigLIP–Qwen LaTeX OCR, GRPO reasoning post-training, inference-time reasoning strategies, and adversarial vision attacks.

Companion page for "Artificial Intelligence for Mathematical Reasoning: An Integrated Survey of Language Models, Neuro-symbolic Systems, and Verified Discovery." A curated four-axis index covering MWPs, LLMs and reasoning models, multimodal geometry, Lean theorem proving, and verified discovery (FunSearch, AlphaEvolve, Erdős problems).

A minimal JEPA-based language model demonstrating latent-space reasoning on GSM8K using a single decoder-only Transformer.

Vision-language pipeline for visual math QA: custom vision-to-text adapter, MPS adapter training, MathVista evaluation (Qwen2.5-VL 0.717, +20.7pp visual ablation)

Data cleaning and structuring pipeline for math reasoning tasks using Qwen3-0.6B for LLM post-training.

STaR × S1 math pipeline on Qwen2.5-1.5B. LoRA, strict Final: format, ~20–30% acc (OpenR1-Math split).

CPU-trained reasoning model pipeline. LoRA SFT + DPO on SmolLM2-360M, GSM8K math reasoning, single-laptop deployment.

A controlled LoRA finetuning study on process supervision for mathematical reasoning with Qwen2.5-Math-7B-Instruct.

NLP course final project (2026), Nanjing Normal University, supervised by 孔力: GSM8K math QA with Seq2Seq, Transformer and LLMs.

Paired-framing pilot on observable self-doubt in math reasoning models

Comprehensive framework for mathematical reasoning research with dual research capabilities

GRPO (Group Relative Policy Optimization) implemented from scratch in PyTorch. 10 ablation experiments.