List your company
ResearchFormerly WizWand

Mathematical Reasoning on AMC23 (Pass@k)

98.6Pass@kQwen2.5-7B + PPO
44.31258.40672.586.594Sep 28, 2025Nov 3, 2025Dec 10, 2025Jan 16, 2026Feb 22, 2026Mar 31, 2026May 7, 2026
Updated 5mo ago

Evaluation results

MethodLinks
Qwen2.5-7B + PPOk=128, Base Model=Qwen2.5-7B, RL Method=PPO2025.0998.6
Qwen2.5-7Bk=128, Base Model=Qwen2.5-7B2025.0998.4
Qwen2.5-7B + GRPO w/ VERL.k=128, Base Model=Qwen2.5-7B, RL Method=GRPO, Framework=VERL2025.0998.3
Qwen2.5-7B + PPO w/ VERL.k=128, Base Model=Qwen2.5-7B, RL Method=PPO, Framework=VERL2025.0998
Qwen2.5-7B + GRPOk=128, Base Model=Qwen2.5-7B, RL Method=GRPO2025.0997.8
Mathstral-7B-v0.1 + GRPO w/ VERL.k=128, Base Model=Mathstral-7B-v0.1, RL Method=GRPO, Framework=VERL2025.0997
Qwen2.5-3B + PPO w/ VERL.k=128, Base Model=Qwen2.5-3B, RL Method=PPO, Framework=VERL2025.0996.8
Qwen2.5-3Bk=128, Base Model=Qwen2.5-3B2025.0996.7
Qwen2.5-3B + PPOk=128, Base Model=Qwen2.5-3B, RL Method=PPO2025.0996.5
Qwen2.5-3B + GRPO w/ VERL.k=128, Base Model=Qwen2.5-3B, RL Method=GRPO, Framework=VERL2025.0995.9
Llama-3.2-3B-Instruct + GRPO w/ VERL.k=128, Base Model=Llama-3.2-3B-Instruct, RL Method=GRPO, Framework=VERL2025.0995.7
Llama-3.2-3B-Instruct + GRPOk=128, Base Model=Llama-3.2-3B-Instruct, RL Method=GRPO2025.0995.4
Llama-3.1-8B-Instruct + GRPO w/ VERL.k=128, Base Model=Llama-3.1-8B-Instruct, RL Method=GRPO, Framework=VERL2025.0995.1
Llama-3.1-8B-Instruct + GRPOk=128, Base Model=Llama-3.1-8B-Instruct, RL Method=GRPO2025.0994.9
Llama-3.2-3B-Instruct + PPO w/ VERL.k=128, Base Model=Llama-3.2-3B-Instruct, RL Method=PPO, Framework=VERL2025.0994.7
Llama-3.1-8B-Instructk=128, Base Model=Llama-3.1-8B-Instruct2025.0994.6
Llama-3.2-3B-Instruct + PPOk=128, Base Model=Llama-3.2-3B-Instruct, RL Method=PPO2025.0994.5
Mathstral-7B-v0.1 + PPO w/ VERL.k=128, Base Model=Mathstral-7B-v0.1, RL Method=PPO, Framework=VERL2025.0993.8
Llama-3.2-3B-Instructk=128, Base Model=Llama-3.2-3B-Instruct2025.0993.5
Llama-3.1-8B-Instruct + PPOk=128, Base Model=Llama-3.1-8B-Instruct, RL Method=PPO2025.0992.4
Qwen2.5-3B + GRPOk=128, Base Model=Qwen2.5-3B, RL Method=GRPO2025.0992.2
Llama-3.1-8B-Instruct + PPO w/ VERL.k=128, Base Model=Llama-3.1-8B-Instruct, RL Method=PPO, Framework=VERL2025.0991.9
Mathstral-7B-v0.1 + PPOk=128, Base Model=Mathstral-7B-v0.1, RL Method=PPO2025.0991.7
Mathstral-7B-v0.1k=128, Base Model=Mathstral-7B-v0.12025.0988.5
Mathstral-7B-v0.1 + GRPOk=128, Base Model=Mathstral-7B-v0.1, RL Method=GRPO2025.0987.3
Mistral-7B-v0.3k=128, Base Model=Mistral-7B-v0.32025.0973.5
Mistral-7B-v0.3 + GRPO w/ VERL.k=128, Base Model=Mistral-7B-v0.3, RL Method=GRPO, Framework=VERL2025.0964.5
Mistral-7B-v0.3 + GRPOk=128, Base Model=Mistral-7B-v0.3, RL Method=GRPO2025.0963.2
GRPO(λ)-0.99Backbone=Qwen3-1.7B2026.0560
GRPOBackbone=Qwen3-1.7B2026.0557.5
GRPO(λ)-0.9Backbone=Qwen3-1.7B2026.0557.5
S-trace-0.9Backbone=Qwen3-1.7B2026.0557.5
OPOBackbone=Qwen3-1.7B2026.0555
Mistral-7B-v0.3 + PPO w/ VERL.k=128, Base Model=Mistral-7B-v0.3, RL Method=PPO, Framework=VERL2025.0946.5
Mistral-7B-v0.3 + PPOk=128, Base Model=Mistral-7B-v0.3, RL Method=PPO2025.0946.4