Completion Sensing on Completion-sense mini benchmark (18 forward transitions across 6 episodes)
0.7222FTACLIP-direct
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| CLIP-directBackbone=ViT-B/32, Alpha (α)=0.7, Baseline penalty=true2026.03 | 0.7222 | 1 | 0.3333 | 5 | |
| VLM-caption pipelineBackbone=Qwen2.5-VL-3B + MiniLM, Pipeline configuration=+ progress-description2026.03 | 0.6667 | 0.8333 | 0.3333 | 2 | |
| VLM-caption pipelineBackbone=Qwen2.5-VL-3B + MiniLM, Pipeline configuration=+ observe-only2026.03 | 0.6667 | 1 | 0.1667 | 2 | |
| VLM-caption pipelineBackbone=Qwen2.5-VL-3B + MiniLM, Pipeline configuration=+ evidence-gating2026.03 | 0 | 0 | 0 | 2 |