List your company
ResearchFormerly WizWand

Mathematical Reasoning on AIME 2025 (Pass@1, Pass@32, Avg)

43.33Pass@1FG-ExPO
-1.73329.965921.66533.3641Sep 28, 2025Nov 9, 2025Dec 21, 2025Feb 1, 2026Mar 15, 2026Apr 26, 2026Jun 7, 2026
Updated 4mo ago

Evaluation results

MethodLinks
FG-ExPOBackbone=Qwen3-8B-Base2026.0543.3376.67—
GRPOBackbone=Qwen3-8B-Base2026.054063.33—
TTRL-GuardBackbone=Qwen3-4B2026.0536.2——
TTRL (w/ FRS)Backbone=Qwen3-4B2026.0536.1——
TTRL (w/ RCSU)Backbone=Qwen3-4B2026.0535.9——
SCOPEBackbone=Qwen3-4B2026.0535.6——
TTRL (w/ MPS)Backbone=Qwen3-4B2026.0535.6——
TTRLBackbone=Qwen3-4B2026.0533.3——
CoVerRLBackbone=Qwen3-4B2026.0531.8——
Qwen3-8B-BaseTraining=RLAD (ours), Context=30K2026.0231.766.466.5
Qwen3-8B-BaseTraining=KDRL, Context=30K2026.0227.865.264.9
ISPOBase Model=Qwen2.5-Math-7B, Decoding Strategy=Greedy (T=0)2026.0624.4——
TTRL-GuardBackbone=Qwen2.5-7B-Instruct2026.0524.1——
TTRL (w/ FRS)Backbone=Qwen2.5-7B-Instruct2026.0523.3——
ISPOBase Model=Qwen2.5-7B, Decoding Strategy=Greedy (T=0)2026.0623.3——
FG-ExPOBackbone=DeepSeek-R1-Distill-Qwen-1.5B2026.0523.1260—
Qwen3-8B-BaseTraining=GRPO, Context=30K2026.0222.848.561
GRPOBackbone=DeepSeek-R1-Distill-Qwen-1.5B2026.0522.656.67—
ISPOBase Model=Qwen2.5-Math-1.5B, Decoding Strategy=Greedy (T=0)2026.0622.2——
Qwen3-8B-BaseTraining=SFT, Context=30K2026.0220.841.556
TTRL (w/ RCSU)Backbone=Qwen2.5-7B-Instruct2026.0520.1——
TTRL (w/ MPS)Backbone=Qwen2.5-7B-Instruct2026.0520——
PREPOBase Model=Qwen2.5-Math-1.5B, Decoding Strategy=Greedy (T=0)2026.0620——
Scaf-GRPOBase Model=Qwen2.5-Math-7B, Decoding Strategy=Greedy (T=0)2026.0620——
Scaf-GRPOBase Model=Qwen2.5-7B, Decoding Strategy=Greedy (T=0)2026.0620——
BaseBackbone=DeepSeek-R1-Distill-Qwen-1.5B2026.0518.5436.67—
GRPO-SBase Model=Qwen2.5-7B, Decoding Strategy=Greedy (T=0)2026.0618.3——
CoVerRLBackbone=Qwen2.5-7B-Instruct2026.0516.7——
GTPOBase Model=Qwen2.5-7B, Decoding Strategy=Greedy (T=0)2026.0616.7——
ProGRPOBase Model=Qwen2.5-7B, Decoding Strategy=Greedy (T=0)2026.0615.9——
TTRLBackbone=Qwen2.5-7B-Instruct2026.0515.6——
SCOPEBackbone=Qwen2.5-7B-Instruct2026.0514.6——
Scaf-GRPOBase Model=Qwen2.5-Math-1.5B, Decoding Strategy=Greedy (T=0)2026.0613.3——
Base modelBase Model=Qwen2.5-Math-7B, Decoding Strategy=Greedy (T=0)2026.0613.3——
Vanilla GRPOBase Model=Qwen2.5-Math-7B, Decoding Strategy=Greedy (T=0)2026.0613.3——
PREPOBase Model=Qwen2.5-Math-7B, Decoding Strategy=Greedy (T=0)2026.0612.8——
PREPOBase Model=Qwen2.5-7B, Decoding Strategy=Greedy (T=0)2026.0610.2——
Qwen2.5-7B + GRPO w/ VERL.Backbone=Qwen2.5-7B, RL Method=GRPO, VERL=Enabled2025.0910——
Vanilla GRPOBase Model=Qwen2.5-Math-1.5B, Decoding Strategy=Greedy (T=0)2026.0610——
Qwen3-1.7B-BaseTraining=RLAD (ours), Context=30K2026.028.324.839
Llama-3.2-3B-Instruct + GRPO w/ VERL.Backbone=Llama-3.2-3B-Instruct, RL Method=GRPO, VERL=Enabled2025.096.7——
Qwen2.5-7B + GRPOBackbone=Qwen2.5-7B, RL Method=GRPO2025.096.7——
Qwen2.5-7B + PPO w/ VERL.Backbone=Qwen2.5-7B, RL Method=PPO, VERL=Enabled2025.096.7——
Base ModelBackbone=Qwen2.5-7B-Instruct2026.056.4——
Qwen3-1.7B-BaseTraining=KDRL, Context=30K2026.026.224.437.2
Vanilla GRPOBase Model=Qwen2.5-7B, Decoding Strategy=Greedy (T=0)2026.066.1——
Qwen3-1.7B-BaseTraining=GRPO, Context=30K2026.025.724.236.5
Qwen3-8B-BaseTraining=-, Context=30K2026.024.724.136.8
Llama-3.2-3B-Instruct + PPOBackbone=Llama-3.2-3B-Instruct, RL Method=PPO2025.093.3——
Llama-3.2-3B-Instruct + PPO w/ VERL.Backbone=Llama-3.2-3B-Instruct, RL Method=PPO, VERL=Enabled2025.093.3——
Llama-3.1-8B-InstructBackbone=Llama-3.1-8B-Instruct, RL Method=None2025.093.3——
Llama-3.1-8B-Instruct + GRPO w/ VERL.Backbone=Llama-3.1-8B-Instruct, RL Method=GRPO, VERL=Enabled2025.093.3——
Qwen2.5-7B + PPOBackbone=Qwen2.5-7B, RL Method=PPO2025.093.3——
Mathstral-7B-v0.1 + PPOBackbone=Mathstral-7B-v0.1, RL Method=PPO2025.093.3——
TTRL (w/ FRS)Backbone=Llama-3.2-3B-Instruct2026.053.3——
TTRL-GuardBackbone=Llama-3.2-3B-Instruct2026.053.3——
Base modelBase Model=Qwen2.5-Math-1.5B, Decoding Strategy=Greedy (T=0)2026.063.3——
Base modelBase Model=Qwen2.5-7B, Decoding Strategy=Greedy (T=0)2026.062.5——
CoVerRLBackbone=Llama-3.2-3B-Instruct2026.051.8——
Base ModelBackbone=Qwen3-4B2026.051.8——
Qwen3-1.7B-BaseTraining=-, Context=30K2026.020.96.222.2
TTRL (w/ RCSU)Backbone=Llama-3.2-3B-Instruct2026.050.9——
SCOPEBackbone=Llama-3.2-3B-Instruct2026.050.8——
TTRL (w/ MPS)Backbone=Llama-3.2-3B-Instruct2026.050.8——
Llama-3.2-3B-InstructBackbone=Llama-3.2-3B-Instruct, RL Method=None2025.090——
Llama-3.2-3B-Instruct + GRPOBackbone=Llama-3.2-3B-Instruct, RL Method=GRPO2025.090——
Llama-3.1-8B-Instruct + GRPOBackbone=Llama-3.1-8B-Instruct, RL Method=GRPO2025.090——
Llama-3.1-8B-Instruct + PPOBackbone=Llama-3.1-8B-Instruct, RL Method=PPO2025.090——
Llama-3.1-8B-Instruct + PPO w/ VERL.Backbone=Llama-3.1-8B-Instruct, RL Method=PPO, VERL=Enabled2025.090——
Qwen2.5-3BBackbone=Qwen2.5-3B, RL Method=None2025.090——
Qwen2.5-3B + GRPOBackbone=Qwen2.5-3B, RL Method=GRPO2025.090——
Qwen2.5-3B + GRPO w/ VERL.Backbone=Qwen2.5-3B, RL Method=GRPO, VERL=Enabled2025.090——
Qwen2.5-3B + PPOBackbone=Qwen2.5-3B, RL Method=PPO2025.090——
Qwen2.5-3B + PPO w/ VERL.Backbone=Qwen2.5-3B, RL Method=PPO, VERL=Enabled2025.090——
Qwen2.5-7BBackbone=Qwen2.5-7B, RL Method=None2025.090——
Mathstral-7B-v0.1Backbone=Mathstral-7B-v0.1, RL Method=None2025.090——
Mathstral-7B-v0.1 + GRPOBackbone=Mathstral-7B-v0.1, RL Method=GRPO2025.090——
Mathstral-7B-v0.1 + GRPO w/ VERL.Backbone=Mathstral-7B-v0.1, RL Method=GRPO, VERL=Enabled2025.090——
Mathstral-7B-v0.1 + PPO w/ VERL.Backbone=Mathstral-7B-v0.1, RL Method=PPO, VERL=Enabled2025.090——
Mistral-7B-v0.3Backbone=Mistral-7B-v0.3, RL Method=None2025.090——
Mistral-7B-v0.3 + GRPOBackbone=Mistral-7B-v0.3, RL Method=GRPO2025.090——
Mistral-7B-v0.3 + GRPO w/ VERL.Backbone=Mistral-7B-v0.3, RL Method=GRPO, VERL=Enabled2025.090——
Mistral-7B-v0.3 + PPOBackbone=Mistral-7B-v0.3, RL Method=PPO2025.090——
Mistral-7B-v0.3 + PPO w/ VERL.Backbone=Mistral-7B-v0.3, RL Method=PPO, VERL=Enabled2025.090——
Base ModelBackbone=Llama-3.2-3B-Instruct2026.050——
TTRLBackbone=Llama-3.2-3B-Instruct2026.050——