List your company
ResearchFormerly WizWand

Mathematical Reasoning on AMC23 (Pass@1 and Pass@8)

86.02Pass@1 ScoreGRPO w/ MAPR
16.527234.568652.6170.6514Sep 26, 2025Nov 10, 2025Dec 26, 2025Feb 10, 2026Mar 28, 2026May 13, 2026Jun 28, 2026
Updated 4mo ago

Evaluation results

MethodLinks
GRPO w/ MAPRBase Model=Qwen3-14B2025.0986.0295.12
GRPOBase Model=Qwen3-14B2025.0981.5696.2
GRPO w/ MAPRBase Model=Qwen3-8B2025.0979.5394.39
GRPOBase Model=Qwen3-8B2025.0973.6792.77
GRPO w/ MAPRBase Model=Qwen3-4B2025.0970.1693.18
oracle ZS vs MemBase Model=Qwen2.5-Math-7B, #Train Params=-2026.0667.5—
ELM offlineBase Model=Qwen2.5-Math-7B, #Train Params=0.07M2026.0665—
ELM kvBackbone=Qwen2.5-Math-7B, Setting=Online2026.0662.5—
GRPOBackbone=Qwen2.5-Math-7B, Setting=Offline2026.0662.5—
TTRLBackbone=Qwen2.5-Math-7B, Setting=Offline2026.0662.5—
Full GTBase Model=Qwen2.5-Math-7B, #Train Params=7B2026.0662.5—
Full MajBase Model=Qwen2.5-Math-7B, #Train Params=7B2026.0662.5—
ELM-LoRAofflineBase Model=Qwen2.5-Math-7B, #Train Params=40M2026.0662.5—
ELM-Full ModelofflineBase Model=Qwen2.5-Math-7B, #Train Params=7B2026.0660—
GRPOBase Model=Qwen3-4B2025.0959.384.93
ELM inputBackbone=Qwen2.5-Math-7B, Setting=Online2026.0656—
PASS2026.0653.883.7
Zero-shot CoTBackbone=Qwen2.5-Math-7B2026.0650—
Zero-shot CoTBase Model=Qwen2.5-Math-7B, #Train Params=-2026.0650—
GRPO(ORM)2026.064783.5
LAGBackbone=Qwen2.5-Math-7B, Setting=Online2026.0640—
ICLBackbone=Qwen2.5-Math-7B, Setting=Online2026.0638.3—
Reasoning BankBackbone=Qwen2.5-Math-7B, Setting=Online2026.0637.5—
oracle ZS vs MemBase Model=Llama-3.1-8B-Instr., #Train Params=-2026.0635—
ELM offlineBase Model=Llama-3.1-8B-Instr., #Train Params=0.08M2026.0632.5—
ELM-LoRAofflineBase Model=Llama-3.1-8B-Instr., #Train Params=42M2026.0630—
GRPOBackbone=Llama-3.1-8B-Instr., Setting=Offline2026.0627.5—
TTRLBackbone=Llama-3.1-8B-Instr., Setting=Offline2026.0627.5—
GRPOBase Model=Llama-3.1-8B-Instr., #Train Params=8B2026.0627.5—
TTRLBase Model=Llama-3.1-8B-Instr., #Train Params=8B2026.0627.5—
ELM-Full ModelofflineBase Model=Llama-3.1-8B-Instr., #Train Params=8B2026.0627.5—
ELM inputBackbone=Llama-3.1-8B-Instr., Setting=Online2026.0626—
Reasoning BankBackbone=Llama-3.1-8B-Instr., Setting=Online2026.0624.2—
ELM kvBackbone=Llama-3.1-8B-Instr., Setting=Online2026.0623.3—
Base Model (Qwen2.5-Math-7B)Backbone=Qwen2.5-Math-7B2026.0622.870.4
ICLBackbone=Llama-3.1-8B-Instr., Setting=Online2026.0621.7—
Zero-shot CoTBackbone=Llama-3.1-8B-Instr.2026.0620—
Zero-shot CoTBase Model=Llama-3.1-8B-Instr., #Train Params=-2026.0620—
LAGBackbone=Llama-3.1-8B-Instr., Setting=Online2026.0619.2—
AsymPOBase Model=Qwen3-1.7B-Base, Training Data=randomly sampled 4k subset of DAPO-17K, Evaluation Rollouts=82026.06—67.5
GRPOBase Model=Qwen3-1.7B-Base, Training Data=randomly sampled 4k subset of DAPO-17K, Evaluation Rollouts=82026.06—65
SPOBase Model=Qwen3-1.7B-Base, Training Data=randomly sampled 4k subset of DAPO-17K, Evaluation Rollouts=82026.06—67.5