Fact Generation (Targeted) on WD50K
50.8AccuracyKREPE
Evaluation results
| Method | Links | |
|---|---|---|
| KREPELLM=N/A2026.05 | 50.8 | |
| Random FactsLLM=Gemini 3 Pro2026.05 | 48.2 | |
| Random FactsLLM=GPT 5.22026.05 | 46.3 | |
| Neighbor SetsLLM=Gemini 3 Pro2026.05 | 44.5 | |
| Neighbor SetsLLM=GPT 5.22026.05 | 39.7 | |
| Few-shot FactsLLM=Gemini 3 Pro2026.05 | 28.1 | |
| Iterative PredictionLLM=N/A2026.05 | 27 | |
| Re-rankingLLM=Gemini 3 Pro2026.05 | 25.6 | |
| AutoregressiveLLM=GPT 5.22026.05 | 25 | |
| Few-shot FactsLLM=GPT 5.22026.05 | 22.2 | |
| Gibbs SamplingLLM=N/A2026.05 | 19.9 | |
| AutoregressiveLLM=Gemini 3 Pro2026.05 | 19 | |
| Re-rankingLLM=GPT 5.22026.05 | 11.2 |