Sequence-level uncertainty estimation on SPANUQ-BENCH 2,000 sequences (test)
0.858Spearman CorrelationMLP Probe
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| MLP ProbeGranularity=Token, Oracle setting=ground-truth span boundaries2026.07 | 0.858 | 0.867 | 95.6 | 0.071 | |
| SPANUQGranularity=Span2026.07 | 0.851 | 0.839 | 94.8 | 0.077 | |
| SelfCheckGPT-NLIGranularity=Sequence2026.07 | 0.805 | 0.75 | 90.8 | 0.097 | |
| SE (10 samples)Granularity=Sequence2026.07 | 0.606 | 0.521 | 80.5 | 0.427 | |
| FActScoreGranularity=Claim2026.07 | 0.489 | 0.505 | 77.9 | 0.138 | |
| INSIDE (Jaccard)Granularity=Sequence2026.07 | 0.331 | 0.314 | 31.5 | 0.246 | |
| Verbalized Conf.Granularity=Sequence2026.07 | 0.322 | 0.229 | 64.1 | 0.166 | |
| KLEGranularity=Sequence2026.07 | 0.145 | 0.041 | 42.5 | 0.784 | |
| P(True)Granularity=Sequence2026.07 | 0.14 | 0.185 | 55.2 | 0.172 |