Mathematical Reasoning on GSM8K (Accuracy, Throughput, Speedup)
85.9AccuracyLatent Thinking Optimization
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Latent Thinking OptimizationModel=Llama-3-8B2025.09 | 85.9 | — | — | |
| Majority VotingModel=Llama-3-8B2025.09 | 80.1 | — | — | |
| Prefix-CacheGen Length=256, Prompting Shot=5-shot2026.01 | 79.5 | 26 | 3.9 | |
| LLaDAGen Length=256, Prompting Shot=5-shot2026.01 | 79.3 | 6.6 | 1 | |
| Streaming-dLLMGen Length=512, Prompting Shot=5-shot2026.01 | 78.7 | 65.9 | 20.6 | |
| Fast-dLLMGen Length=256, Prompting Shot=5-shot2026.01 | 78.5 | 50.9 | 7.7 | |
| Streaming-dLLMGen Length=256, Prompting Shot=5-shot2026.01 | 78.5 | 66.2 | 10 | |
| Base ModelModel=Llama-3-8B2025.09 | 78.4 | — | — | |
| dKV-CacheGen Length=512, Prompting Shot=5-shot2026.01 | 77.6 | 5.5 | 1.7 | |
| LLaDAGen Length=512, Prompting Shot=5-shot2026.01 | 77.5 | 3.2 | 1 | |
| Fast-dLLMGen Length=512, Prompting Shot=5-shot2026.01 | 77.2 | 32.1 | 10 | |
| Prefix-CacheGen Length=512, Prompting Shot=5-shot2026.01 | 77 | 11 | 3.4 | |
| dKV-CacheGen Length=256, Prompting Shot=5-shot2026.01 | 76.5 | 11.6 | 1.8 | |
| Latent Thinking OptimizationModel=Llama-2-13B2025.09 | 53.4 | — | — | |
| Majority VotingModel=Llama-2-13B2025.09 | 41.7 | — | — | |
| Latent Thinking OptimizationModel=Llama-2-7B2025.09 | 38.9 | — | — | |
| Base ModelModel=Llama-2-13B2025.09 | 30.6 | — | — | |
| Majority VotingModel=Llama-2-7B2025.09 | 27.5 | — | — | |
| Base ModelModel=Llama-2-7B2025.09 | 22.3 | — | — |