Qwen3-8B · Natural Questions
Selective answering (Probing)TAPS (Probing)Abstention training
Risk versus coverage on Qwen3-8B / NQTAPS and selective answering curves with five fixed abstention-training points. At 50.45 percent coverage, selective answering has 41.3 percent risk, TruthRL 37.3 percent, and TAPS 33.3 percent.0204060020406080100Risk (%)Coverage (%)r-tuning: 47.23% coverage, 46.29% risktruthrl-0.5: 46.37% coverage, 36.63% risktruthrl-0: 50.45% coverage, 37.27% risktruthrl--0.5: 50.13% coverage, 37.76% riskucpo: 17.43% coverage, 34.73% riskUCPOR-TuningTruthRL

Selective answering41.3%

Abstention training37.3%

TAPS33.3%

Original interactive results by Junwei Zhang & collaborators ↗. Drag the slider to compare methods at the same answer rate.