Referring Expression Comprehension on Mixed 3-way (RefCOCOg + Ref-Adv-S + Ref-L4) (test)
82Acc@0.5SFT + IVT
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| SFT + IVTIVT=true, Refinement steps=T=22026.06 | 82 | 74.1 | 48.3 | 74.3 | |
| SFT + GRPO + IVTIVT=true, Refinement steps=T=22026.06 | 80.6 | 72.5 | 47.1 | 72.9 | |
| SFT + GRPO (single-step)IVT=false2026.06 | 80 | 70.9 | 45.7 | 71.6 | |
| Base (single-shot)IVT=false2026.06 | 79.6 | 70.9 | 45.5 | 71.9 | |
| Base + IVT (no training)IVT=true, Refinement steps=T=22026.06 | 48.7 | 44.2 | 28.3 | 44.2 |