Agent action prediction on Switchboard 138-session (test)
63.3wF1DualTurn (A – LoRA (main))
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| DualTurn (A – LoRA (main))variant=A, finetuning=LoRA2026.03 | 63.3 | 34.9 | 87.4 | |
| DualTurn (E – Full FT)variant=E, finetuning=Full2026.03 | 62.6 | 33.7 | 87.9 | |
| DualTurn (B – LoRA + CB)variant=B, finetuning=LoRA + CB2026.03 | 61.3 | 7.7 | 87.6 | |
| DualTurn (G – Text-Aware PT)variant=G, pretraining=Text-Aware2026.03 | 60.5 | 8.5 | 87.3 | |
| DualTurn (C – No Pretrain)variant=C, pretraining=none2026.03 | 60.4 | 7.9 | 86.3 | |
| DualTurn (D – LSTM)variant=D, architecture=LSTM2026.03 | 60.2 | 7.7 | 86.8 | |
| DualTurn (F – Discrete)variant=F, type=Discrete2026.03 | 60.2 | 7.2 | 86.7 | |
| VAP (LR-6)aggregator=logistic regression2026.03 | 38.9 | 0 | 78 | |
| VAP (native)type=baseline2026.03 | 27.6 | — | 78.5 |