List your company
ResearchFormerly WizWand

Mathematical Reasoning on AIME 2024

94AccuracyGPT-5-Mini-R
49.2860.8972.584.11Jan 31, 2025Apr 9, 2025Jun 16, 2025Aug 24, 2025Oct 31, 2025Jan 7, 2026Mar 17, 2026
Updated 7mo ago

Evaluation results

MethodLinks
GPT-5-Mini-R2026.0394——
Agentic Proposingtraining_budget=11,000 mixed trajectories, recipe=GRPO, evaluation=Mean@64, proposer=Agentic-Proposer-30B2026.0293.5——
WSC-GMM*Model=DeepSeek-R1-8B, SelfStepConf=true2026.0393.33——
GPT-5-Mini2026.0393——
PromptCoT 2.0training_budget=11,000 mixed trajectories, recipe=GRPO, evaluation=Mean@642026.0292.1——
WSC-GMM*Model=Qwen3-32B, Thinking mode=false, SelfStepConf=true2026.0390.16——
PG@16Model=Deepseek-R1-Distill-Qwen32B2026.0390——
Gemini 2.5 proEvaluation Protocol=pass@12025.0988.7——
OpenMathReasoningtraining_budget=11,000 mixed trajectories, recipe=GRPO, evaluation=Mean@642026.0287.9——
Qwen3-30B-A3B-Thinking-2507mode=zero-shot, evaluation=Mean@642026.0287.7——
SC@16Model=Deepseek-R1-Distill-Qwen32B2026.0386.67——
SCModel=DeepSeek-R1-8B2026.0386.67——
OpenR1training_budget=11,000 mixed trajectories, recipe=GRPO, evaluation=Mean@642026.0286.3——
OpenThoughts-S3training_budget=11,000 mixed trajectories, recipe=GRPO, evaluation=Mean@642026.0285.7——
OpenCodeReasoningtraining_budget=11,000 mixed trajectories, recipe=GRPO, evaluation=Mean@642026.0285——
o3-mini-highaccess=API only, examples_for_finetuning=N/A2025.0183.8——
Qwen3-235B-A22BModel Category=Large Language Models, Model Scale=235B-A22B, Reasoning Strategy=Thinking2025.1283.8——
SwiRBackbone=Qwen3-32B2025.1082.92——
Qwen3-32BModel Category=Large Language Models, Model Scale=32B, Reasoning Strategy=Thinking2025.1281.4——
CoTBackbone=Qwen3-32B2025.1080.42——
MultiRole-R1Model Backbone=Qwen3-8B2025.0780.1——
CoT (Greedy)Backbone=Qwen3-32B2025.1080——
r1access=Open Weights, examples_for_finetuning=>800K2025.0179.8——
R1# Examples=>>800K2025.1279.8——
SFT+GRPOModel Backbone=Qwen3-8B2025.0779.6——
FIGR2025.1279.58——
SwiRBackbone=Qwen3-8B2025.1079.17——
More thinkModel Backbone=Qwen3-8B2025.0778.8——
SFTModel Backbone=Qwen3-8B2025.0778.3——
Self-RefineModel Backbone=Qwen3-8B2025.0777.5——
Role-PlayModel Backbone=Qwen3-8B2025.0776.6——
BLASSTBackbone=Qwen3-8B, Attention Method=BLASST, Sparsity=~50%, Phase=Decode2025.1276.5——
BLASSTModel=Qwen3-8B, Sparsity=~50%, Deployment Phase=Decode Phase2025.1276.5——
GLM-4.5VModel Category=Large Vision-Language Models, Model Scale=108B2025.1276.2——
Zero-ShotModel Backbone=Qwen3-8B2025.0776——
CoTBackbone=Qwen3-8B2025.1075.83——
o3-mini-mediumaccess=API only, examples_for_finetuning=N/A2025.0175.8——
BLASSTModel=Qwen3-8B, Sparsity=~75%, Deployment Phase=Decode Phase2025.1275.33——
Dense AttentionBackbone=Qwen3-8B, Attention Method=Dense, Phase=Decode2025.1275——
Dense AttentionModel=Qwen3-8B, Sparsity=Dense, Deployment Phase=Decode Phase2025.1275——
Soft ThinkingBackbone=Qwen3-32B2025.1074.58——
RocketKVBackbone=Qwen3-8B, Attention Method=RocketKV, Phase=Decode2025.1273.54——
Qwen3-VL-32B-InstructModel Category=Large Vision-Language Models, Model Scale=32B, Reasoning Strategy=Instruct2025.1273.33——
Text-only RLMode=Text-only, Reasoning Strategy=RL2025.1273.33——
MultiRole-R1Model Backbone=R-D-Qwen-14B2025.0773.3——
SFT+GRPOModel Backbone=R-D-Qwen-14B2025.0772.8——
Avg@16Model=Deepseek-R1-Distill-Qwen32B2026.0372.6——
QuestBackbone=Qwen3-8B, Attention Method=Quest, Phase=Decode2025.1271.5——
More thinkModel Backbone=R-D-Qwen-14B2025.0771.4——
Self-RefineModel Backbone=R-D-Qwen-14B2025.0770.6——
SFTModel Backbone=R-D-Qwen-14B2025.0770.2——
CoT (Greedy)Backbone=Qwen3-8B2025.1070——
Zero-ShotModel Backbone=R-D-Qwen-14B2025.0769.7——
Role-PlayModel Backbone=R-D-Qwen-14B2025.0769.1——
Soft ThinkingBackbone=Qwen3-8B2025.1067.92——
s1K-1.1 + AIR-Step# Examples=1K, Distillation Source=Deepseek R12025.1266.67——
MultiRole-R1Model Backbone=R-D-Qwen-7B2025.0763.2——
TTRL-GTModel=Qwen2.5-7B-Base2026.0362.71——
SFT+GRPOModel Backbone=R-D-Qwen-7B2025.0762.7——
DeepSeek-R1-Distill-Qwen-7B + ERC-DAPOModel Backbone=Qwen-7B, Training Algorithm=ERC-DAPO2025.1262.1——
DeepSeek-R1-Distill-Qwen-7B + DAPOModel Backbone=Qwen-7B, Training Algorithm=DAPO2025.1262——
r1-distill-Qwen-14Baccess=Open Weights, examples_for_finetuning=800K2025.0161.7——
R1-distill-Qwen-14B# Examples=800K2025.1261.7——
Qwen3-VL-8B-InstructModel Category=Large Vision-Language Models, Model Scale=8B, Reasoning Strategy=Instruct2025.1261.46——
SFTModel Backbone=R-D-Qwen-7B2025.0758.9——
More thinkModel Backbone=R-D-Qwen-7B2025.0758.6——
r1-distill-Qwen-32Baccess=Open Weights, examples_for_finetuning=800K2025.0158.3——
R1-distill-Qwen-32B# Examples=800K2025.1258.3——
MultiRole-R1Model Backbone=R-D-Llama-8B2025.0758.1——
PACS-8BEvaluation Protocol=pass@1, Backbone=Qwen3-8B2025.0957.58——
r1-distill-Llama-70Baccess=Open Weights, examples_for_finetuning=800K2025.0157.1——
R1-distill-Llama-70B# Examples=800K2025.1257.1——
SFTModel Backbone=R-D-Llama-8B2025.0756.8——
s1access=Open Weights and Open Data, examples_for_finetuning=1K, budget_forcing=Wait 4x2025.0156.7——
s1.1access=Open Weights and Open Data, examples_for_finetuning=1K, budget_forcing=None2025.0156.7——
s1.1access=Open Weights and Open Data, examples_for_finetuning=1K, budget_forcing=Wait 1x2025.0156.7——
s1.1access=Open Weights and Open Data, examples_for_finetuning=1K, budget_forcing=Wait 2x2025.0156.7——
s1K-1.1# Examples=1K, Distillation Source=Deepseek R12025.1256.7——
AIR-Sample# Examples=1K, Distillation Source=Deepseek R12025.1256.7——
SFT+GRPOModel Backbone=R-D-Llama-8B2025.0756.7——
DCPOBackbone=Qwen3-4B, Optimization Method=DCPO, Sampling Strategy=x322026.0156.6——
o3-mini-lowaccess=API only, examples_for_finetuning=N/A2025.0156.3——
LIMOaccess=Open Weights and Open Data, examples_for_finetuning=8172025.0156.3——
Role-PlayModel Backbone=R-D-Qwen-7B2025.0756.3——
Self-RefineModel Backbone=R-D-Qwen-7B2025.0755.9——
Zero-ShotModel Backbone=R-D-Qwen-7B2025.0755.5——
DeepSeek-R1-Distill-Qwen-7B + GRPOModel Backbone=Qwen-7B, Training Algorithm=GRPO2025.1255.3——
PACS-4BEvaluation Protocol=pass@1, Backbone=Qwen3-4B2025.0955.1——
DeepSeek-R1-Distill-Qwen-7BModel Backbone=Qwen-7B, Training Algorithm=Baseline2025.1254.5——
AEPOBackbone=Qwen3-4B, Optimization Method=AEPO, Sampling Strategy=x322026.0154.5——
More thinkModel Backbone=R-D-Llama-8B2025.0754——
Role-PlayModel Backbone=R-D-Llama-8B2025.0753.4——
s1access=Open Weights and Open Data, examples_for_finetuning=1K, budget_forcing=Wait 1x2025.0153.3——
s1access=Open Weights and Open Data, examples_for_finetuning=1K, budget_forcing=Wait 2x2025.0153.3——
RLOOBackbone=Qwen-3-32B, Context length=4K2026.0353—55
GRPOBackbone=Qwen3-4B, Optimization Method=GRPO, Sampling Strategy=x322026.0152.9——
Entropy-RegBackbone=Qwen3-4B, Optimization Method=Entropy-Reg, Sampling Strategy=x322026.0152.4——
Entropy-AdvBackbone=Qwen3-4B, Optimization Method=Entropy-Adv, Sampling Strategy=x322026.0151.6——
DIVERBase Model=DeepSeek-R1-Distill-Qwen-7B2025.0951.1——
GRPOBackbone=Qwen-3-32B, Context length=4K2026.0351—52