Consistent Reasoning on GSM8K Verification
85.5AccuracyGibbs
Evaluation Results
| Method | Links | |
|---|---|---|
| GibbsLLM=Llama-3.1-8B-Instruct, n=42026.06 | 85.5 | |
| GibbsLLM=Llama-3.1-8B-Instruct, n=642026.06 | 84.7 | |
| GibbsLLM=Llama-3.1-8B-Instruct, n=162026.06 | 83.2 | |
| ICMLLM=OLMo-3-32B-Think, n=642026.06 | 82.5 | |
| ICMLLM=Llama-3.1-8B-Instruct, n=642026.06 | 81 | |
| GibbsLLM=OLMo-3-32B-Think, n=642026.06 | 80.1 | |
| 4-Pass (autoreg.)LLM=Llama-3.1-8B-Instruct2026.06 | 77.4 | |
| GibbsLLM=OLMo-3-32B-Think, n=42026.06 | 77 | |
| Zero-shot (greedy)LLM=OLMo-3-32B-Think2026.06 | 74.3 | |
| GibbsLLM=OLMo-3-32B-Think, n=162026.06 | 74.3 | |
| 1-Pass (autoreg.)LLM=Llama-3.1-8B-Instruct2026.06 | 73.7 | |
| 4-Pass (autoreg.)LLM=OLMo-3-32B-Think2026.06 | 68.1 | |
| Zero-shot (greedy)LLM=Llama-3.1-8B-Instruct2026.06 | 67 | |
| 1-Pass (autoreg.)LLM=OLMo-3-32B-Think2026.06 | 65.1 |