Reproducibility Evaluation on MCR-Bench
0MSEFP32 (+LayerCast)
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| FP32 (+LayerCast)Model=DeepSeek-14B2026.06 | 0 | 0 | |
| FP32 (+LayerCast)Model=Qwen3-14B2026.06 | 0 | 0 | |
| FP32 (+LayerCast)Model=Qwen3-8B2026.06 | 0 | 0 | |
| FP32 (+LayerCast)Model=Llama-8B2026.06 | 0 | 0 | |
| FP32 (+LayerCast)Model=Qwen3-32B2026.06 | 0 | 0 | |
| HEAL-FullModel=DeepSeek-14B2026.06 | 0 | 0.3 | |
| HEAL-FullModel=Llama-8B2026.06 | 0 | 0 | |
| HEAL-FullModel=Qwen3-14B2026.06 | 0 | 0.2 | |
| HEAL-FullModel=Qwen3-32B2026.06 | 0 | 0.3 | |
| HEAL+Model=Qwen3-14B2026.06 | 0 | 0.2 | |
| HEAL+Model=DeepSeek-14B2026.06 | 0 | 0.2 | |
| HEAL-FullModel=Qwen3-8B2026.06 | 0 | 0.1 | |
| HEAL+Model=Llama-8B2026.06 | 0 | 0 | |
| HEAL++Model=DeepSeek-14B2026.06 | 0 | 0.3 | |
| HEAL+Model=Qwen3-8B2026.06 | 0 | 0.4 | |
| HEAL+Model=Qwen3-32B2026.06 | 0 | 0.3 | |
| HEAL++Model=Qwen3-8B2026.06 | 0 | 0.7 | |
| HEAL-BaseModel=DeepSeek-14B2026.06 | 0 | 2 | |
| HEAL++Model=Qwen3-32B2026.06 | 0 | 0.6 | |
| HEAL++Model=Qwen3-14B2026.06 | 0 | 0.1 | |
| HEAL-BaseModel=Qwen3-14B2026.06 | 0 | 0.8 | |
| HEAL++Model=Llama-8B2026.06 | 0 | 0 | |
| HEAL-BaseModel=Qwen3-8B2026.06 | 0 | 0.7 | |
| HEAL-BaseModel=Llama-8B2026.06 | 0 | 0.7 | |
| TF32 (+LayerCast)Model=Qwen3-14B2026.06 | 0 | 1.3 | |
| HEAL-BaseModel=Qwen3-32B2026.06 | 0 | 1.4 | |
| TF32 (+LayerCast)Model=DeepSeek-14B2026.06 | 0 | 2.7 | |
| TF32 (+LayerCast)Model=Qwen3-8B2026.06 | 0 | 1.8 | |
| FP16Model=Qwen3-14B2026.06 | 0 | 3.9 | |
| TF32 (+LayerCast)Model=Llama-8B2026.06 | 0 | 3.2 | |
| FP16Model=DeepSeek-14B2026.06 | 0 | 6.9 | |
| FP16Model=Llama-8B2026.06 | 0 | 3.1 | |
| FP16Model=Qwen3-8B2026.06 | 0 | 5.1 | |
| TF32 (+LayerCast)Model=Qwen3-32B2026.06 | 0 | 1.9 | |
| FP16Model=Qwen3-32B2026.06 | 0.0001 | 1.1 | |
| BF16Model=Qwen3-14B2026.06 | 0.0013 | 5.9 | |
| BF16Model=DeepSeek-14B2026.06 | 0.0014 | 12.8 | |
| BF16Model=Qwen3-8B2026.06 | 0.0016 | 6.2 | |
| BF16Model=Llama-8B2026.06 | 0.0019 | 11.7 | |
| BF16Model=Qwen3-32B2026.06 | 0.0046 | 3.9 |