Sudoku solving on Sudoku-Extreme (test)
100AccuracyLDT
Evaluation results
| Method | Links | ||
|---|---|---|---|
| LDTTraining steps=4K, Parameters=800K, Train Set=1K, Inference cost=0.028s/ex, Training Cost=15m (1× B200)2026.05 | 100 | 100 | |
| EqRModel Type=Iterative (weight-tied)2026.05 | 99.8 | — | |
| LDTTraining steps=4K, Parameters=800K, Train Set=1K (no-aug), Inference cost=0.051s/ex, Training Cost=13m (1× B200), Data augmentation=false2026.05 | 99.7 | 100 | |
| LDTTraining steps=2K, Parameters=800K, Train Set=1K, Inference cost=0.18s/ex, Training Cost=7m (1× B200)2026.05 | 99.3 | 100 | |
| SotakuParameters=800K, Train Set=2.7M, Training Cost=2h 40m (1× H100)2026.05 | 98.9 | 98.9 | |
| GRAM#Params=10M, Supervision steps=162026.05 | 97 | — | |
| TRMNumber of Parameters=5M, Features=MLP-mixer, Hidden size=5122026.03 | 87.4 | — | |
| TRMParameters=5M, Train Set=1K, Training Cost=36h (4× L40S)2026.05 | 87.4 | 87.4 | |
| TRM#Params=7M, Supervision steps=162026.05 | 87.4 | — | |
| LDTTraining steps=1K, Parameters=800K, Train Set=1K, Inference cost=0.78s/ex, Training Cost=4m (1× B200)2026.05 | 85.6 | 100 | |
| TRMModel Type=Iterative (weight-tied)2026.05 | 84.8 | — | |
| TRMNumber of Parameters=5M, Features=AMP, MLP-mixer, Hidden size=5122026.03 | 79.1 | — | |
| URMModel Type=Iterative (weight-tied)2026.05 | 77.6 | — | |
| TRMNumber of Parameters=7M, Features=Attention, Hidden size=5122026.03 | 74.7 | — | |
| Looped TF#Params=7M, Supervision steps=162026.05 | 61.3 | — | |
| HRMNumber of Parameters=27M, Hidden size=5122026.03 | 55 | — | |
| HRMParameters=27M, Train Set=1K2026.05 | 55 | 55 | |
| HRM#Params=27M, Supervision steps=162026.05 | 55 | — | |
| HRMModel Type=Iterative (weight-tied)2026.05 | 55 | — | |
| HRMNumber of Parameters=27M, Features=AMP, Hidden size=5122026.03 | 45.9 | — | |
| 64-Layer ModelModel Type=Feedforward, Number of Layers=642026.05 | 2.6 | — | |
| 16-Layer ModelModel Type=Feedforward, Number of Layers=162026.05 | 2.1 | — | |
| 4-Layer ModelModel Type=Feedforward, Number of Layers=42026.05 | 1.8 | — | |
| DeepSeek R1Number of Parameters=671B2026.03 | 0 | — | |
| Claude Opus 4.6Parameters=?, Train Set=—2026.05 | 0 | 0 | |
| DeepSeek V4-ProParameters=1.6T, Train Set=—2026.05 | 0 | 0 | |
| GPT-5.4Parameters=?, Train Set=—2026.05 | 0 | 0 | |
| Deepseek-R1#Params=671B2026.05 | 0 | — | |
| Claude 3.7 16k#Params=N/A2026.05 | 0 | — | |
| o3-mini-high#Params=N/A2026.05 | 0 | — | |
| Direct Pred#Params=27M, Supervision steps=162026.05 | 0 | — |