List your company

Mathematical Reasoning on HMMT (Accuracy, Variance, and Efficiency Metrics)

88.9AccuracyInsightReplay (IR3)
-1.933621.648245.2368.8118May 14, 2026May 18, 2026May 22, 2026May 26, 2026May 30, 2026Jun 3, 2026Jun 8, 2026

Evaluation results

MethodLinks
InsightReplay (IR3)Model=Gemma4-31B, Inference Mode=IR32026.0588.9——0.813,730——1.62
Verify-Only (VO)Model=Gemma4-31B, Inference Mode=VO2026.0588.70.6——10,4301.23——
InsightReplay (IR3)Model=Qwen3.5-35B, Inference Mode=IR32026.0588.6——0.151,840——1.29
Base (CoT)Model=Qwen3.5-35B, Inference Mode=Base2026.0588.5———40,124———
Verify-Only (VO)Model=Qwen3.5-35B, Inference Mode=VO2026.0588.50——44,0291.1——
InsightReplay (IR1)Model=Qwen3.5-35B, Inference Mode=IR12026.0588.4—0.1—45,161—1.13—
Base (CoT)Model=Gemma4-31B, Inference Mode=Base2026.0588.1———8,499———
InsightReplay (IR1)Model=Gemma4-31B, Inference Mode=IR12026.0588.1—0—10,541—1.24—
InsightReplay (IR3)Model=Qwen3.5-9B, Inference Mode=IR32026.0584.5——158,490——1.29
InsightReplay (IR1)Model=Qwen3.5-9B, Inference Mode=IR12026.0584.2—0.7—51,088—1.12—
Verify-Only (VO)Model=Qwen3.5-9B, Inference Mode=VO2026.0583.90.4——49,7961.09——
Base (CoT)Model=Qwen3.5-9B, Inference Mode=Base2026.0583.5———45,477———
InsightReplay (IR3)Model=R1-Distill-32B, Inference Mode=IR32026.0545.5——1.717,278——1.34
InsightReplay (IR1)Model=R1-Distill-32B, Inference Mode=IR12026.0544.5—0.7—14,271—1.11—
Verify-Only (VO)Model=R1-Distill-32B, Inference Mode=VO2026.0544.10.3——13,8671.07——
Base (CoT)Model=R1-Distill-32B, Inference Mode=Base2026.0543.8———12,912———
Verify-Only (VO)Model=Gemma4-E4B, Inference Mode=VO2026.0537.71.3——9,2031.23——
InsightReplay (IR1)Model=Gemma4-E4B, Inference Mode=IR12026.0536.7—0.3—9,064—1.21—
InsightReplay (IR3)Model=Gemma4-E4B, Inference Mode=IR32026.0536.6——0.211,906——1.59
Base (CoT)Model=Gemma4-E4B, Inference Mode=Base2026.0536.4———7,481———
Verify-Only (VO)Model=R1-Distill-7B, Inference Mode=VO2026.0529.10.8——15,4491.07——
InsightReplay (IR1)Model=R1-Distill-7B, Inference Mode=IR12026.0529—0.7—15,818—1.1—
InsightReplay (IR3)Model=R1-Distill-7B, Inference Mode=IR32026.0529——0.719,853——1.38
Base (CoT)Model=R1-Distill-7B, Inference Mode=Base2026.0528.3———14,417———
DAPO w/ EDASBackbone=Qwen3-4B2026.0523.12———————
DAPOBackbone=Qwen3-4B2026.0522.4———————
DAPO w/ EDASBackbone=Qwen3-8B2026.0520.1———————
DAPOBackbone=Qwen3-8B2026.0515.94———————
DAPO w/ En AdvBackbone=Qwen3-8B2026.0515.21———————
DAPO w/ En AdvBackbone=Qwen3-4B2026.0513.65———————
GRPO w/ EDASBackbone=Qwen3-8B2026.0512.5———————
PKPOBackbone=Qwen3-8B2026.0512.08———————
GRPOBackbone=Qwen3-8B2026.0511.25———————
DAPO w/ EDASBackbone=Qwen3-4B-Base2026.0510.94———————
BaselineBackbone=Qwen3-4B2026.0510.83———————
BaselineBackbone=Qwen3-8B2026.0510.42———————
DAPOBackbone=Qwen3-4B-Base2026.058.33———————
INFUSERModel Backbone=Qwen3-8B-Base2026.067.04———————
DAPO w/ En AdvBackbone=Qwen3-4B-Base2026.056.67———————
AZRModel Backbone=Qwen3-8B-Base2026.066.01———————
AZRModel Backbone=Qwen3-4B-Base2026.064.5———————
R-ZeroModel Backbone=Qwen3-8B-Base2026.063.86———————
BaseModel Backbone=Qwen3-8B-Base2026.062.96———————
INFUSERModel Backbone=Qwen3-4B-Base2026.062.94———————
R-ZeroModel Backbone=Qwen3-4B-Base2026.062.65———————
BaseModel Backbone=Qwen3-4B-Base2026.062.49———————
BaselineBackbone=Qwen3-4B-Base2026.051.56———————