AI Paper Atlas

Research map / Benchmarks / MATH

MATH leaderboard

227 papers at ICML, NeurIPS, ICLR, CVPR, AAAI, ICCV and ECCV, and the journals TPAMI, IJCV and JMLR report results on MATH. Best accuracy: 96.67 (HieraMAS (Ours), ICML 2026).

Numbers are as each paper reports them, read by AI from its results tables and not checked by hand. Papers can differ in setting (training data, model size, extra data); the topic column shows what each paper is about. Open the paper before you compare or quote a number.

MATH · accuracy

186 papers, each with its best reported number for its own method; higher is better.

#accuracy ↑MethodPaperTopic
196.67HieraMAS (Ours)
GPT-5-Mini
HieraMAS: Optimizing Intra-Node LLM Mixtures and Inter-Node Topology for Multi-Agent Systems
ICML 2026 · Table 1
LLM agents
295.33C-Evolve
GPT-4.1 mini
C-Evolve: Consensus-based Evolution for Prompt Groups
ICLR 2026 · Table 1
AI and society
394.3RRM-32B
DeepSeek-R1-Distill-Qwen-32B
Reward Reasoning Models
NeurIPS 2025 · Binary preference classificati
Reasoning and chain of thought
493.5Archon all-source task-specificAn Architecture Search Framework for Inference-Time Techniques
ICML 2025 · Main results table
Reasoning and chain of thought
593ThoughtComm
Qwen 3-1.7B
Thought Communication in Multiagent Collaboration
NeurIPS 2025 · Table 1
LLM agents
692.86AgentNet
DeepSeek-V3
AgentNet: Decentralized Evolutionary Coordination for LLM-based Multi-Agent Systems
NeurIPS 2025 · Main Results table
LLM agents
791.66CARD
deepseek-v3
CARD: Towards Conditional Design of Multi-agent Topological Structures
ICLR 2026 · Table 1
LLM agents
891TypedThinker + SC @5
GPT-4o
TypedThinker: Diversify Large Language Model Reasoning with Typed Thinking
ICLR 2025 · Table: GPT-4o
Agents, code and math
990.72R-KV (budget 1,024)
Qwen-14B
R-KV: Redundancy-aware KV Cache Compression for Reasoning Models
NeurIPS 2025 · Appendix, Main Results table
Reasoning and chain of thought
1090.54XRPO
Qwen3-1.7B
XRPO: Pushing the Limits of GRPO with Targeted Exploration and Exploitation
ICML 2026 · Table 1b
RL from feedback
176 more papers on this leaderboard, the baselines reported on it and a CSV download are free with a scime account.
Sign in free to see all

MATH · score

27 papers, each with its best reported number for its own method; higher is better.

#score ↑MethodPaperTopic
196.5LoGos(32B)
Qwen2.5-32B-Base
Mixing Expert Knowledge: Bring Human Thoughts Back To the Game of Go
NeurIPS 2025 · Main results table
Agents, code and math
296.5AoT
DeepSeek-V3
Atom of Thoughts for Markov LLM Test-Time Scaling
NeurIPS 2025 · Performance Comparison table
Reasoning and chain of thought
387.9Qwen2.5-math-7B-Instruct + Countermath-SFT
Qwen2.5-math-7B-Instruct
One Example Shown, Many Concepts Known! Counterexample-Driven Conceptual Reasoning in Mathematical LLMs
ICML 2025 · Out-of-distribution evaluation
Agents, code and math
487.62AgentGrad
GPT-5-mini
AgentGrad: Intervention-guided Prompt Optimization for Multi Agent Systems
NeurIPS 2026 · Table 1
LLM agents
581.12CFAFairness and Explainability: Bridging the Gap towards Fair Model Explanations
AAAI 2023 · Table 1
Fairness
680.1InfiFPO
Phi-4
InfiFPO: Implicit Model Fusion via Preference Optimization in Large Language Models
NeurIPS 2025 · Table Main Results
Alignment and preferences
774.74GLD + ULD
Phi4
InfiGFusion: Graph-on-Logits Distillation via Efficient Gromov-Wasserstein for Model Fusion
NeurIPS 2025 · Ablation study on loss compone
Reasoning and chain of thought
872.71FlowBot
GPT-4.1 mini
FlowBot: Inducing LLM Workflows with Bilevel Optimization and Textual Gradients
ICML 2026 · Table 3
Agents, code and math
968.75RAIN-Merging
DeepSeek-R1-Distill-Qwen-7B + Qwen2.5-7B-Instruct
RAIN-Merging: A Gradient-Free Method to Enhance Instruction Following in Large Reasoning Models with Preserved Thinking Format
ICLR 2026 · Table 1
Reasoning and chain of thought
1066.8InfoSteer interventionSteering Information Utility in Key-Value Memory for Language Model Post-Training
NeurIPS 2025 · Task type table
Instruction and fine-tuning
17 more papers on this leaderboard, the baselines reported on it and a CSV download are free with a scime account.
Sign in free to see all

MATH · pass@1

15 papers, each with its best reported number for its own method; higher is better.

#pass@1 ↑MethodPaperTopic
190rStar-Math (7B SLM+7B PPM)⁶⁴
Qwen2.5-Math-7B
rStar-Math: Small LLMs Can Master Math Reasoning with Self-Evolved Deep Thinking
ICML 2025 · Table: results on challenging
Reasoning and chain of thought
288.2TAPO-1.5B
DeepSeek-R1-Distill-Qwen-1.5B
TAPO: Dynamic Teacher and Perturbed Answer Injection for Policy Optimization
AAAI 2026 · Table 1
RL from feedback
376.6W-REINFORCE
Qwen2.5-Math-7B
The Surprising Effectiveness of Negative Reinforcement in LLM Reasoning
NeurIPS 2025 · Table 1
Reasoning and chain of thought
464.55Discounted GRPO
Qwen2.5-7B-Instruct
Learning to Reason Efficiently with Discounted Reinforcement Learning
ICLR 2026 · Table 1
Reasoning and chain of thought
561.08MARGE-RL
Qwen2-7B-Instruct
MARGE: Improving Math Reasoning with Guided Exploration
ICML 2025 · Table 1
Reasoning and chain of thought
658.3KPMath-Plus
Qwen1.5-72B
Key-Point-Driven Data Synthesis with Its Enhancement on Mathematical Reasoning
AAAI 2025 · Main results table
Agents, code and math
756.2RiskPO
Qwen2.5-Math-1.5B-Instruct
RiskPO: Risk-based Policy Optimization with Verifiable Reward for LLM Post-Training
ICLR 2026 · Table 2
RL from feedback
855.4WizardMath-SFT + PRM + IRM
Mistral-7B
WizardMath: Empowering Mathematical Reasoning for Large Language Models via Reinforced Evol-Instruct
ICLR 2025 · PRM and IRM effects table
Agents, code and math
955.06ProSeCo SFT + ProSeCo Max Sampling
LLaDA-Base 8B
Learn from Your Mistakes: Self-Correcting Masked Diffusion Models
NeurIPS 2026 · Table 1
Fast diffusion sampling
1051.15PKPO k_opt=1
Llama3.1-8B
Pass@K Policy Optimization: Solving Harder Reinforcement Learning Problems
NeurIPS 2025 · Table 2
Policy gradient and actor-critic
5 more papers on this leaderboard, the baselines reported on it and a CSV download are free with a scime account.
Sign in free to see all

Not affiliated with any conference, journal or the benchmark's authors. Corrections: contact@latenticlabs.com.