Embodied Question Answering on ExploreEQA
47.7Success Rate (SR)ExploreEQA
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| ExploreEQAOptimizer=AFlow, Surviving change (axis)=VLM single-letter prompt rewrite (prompt)2026.06 | 47.7 | 4.7 | |
| ExploreEQAOptimizer=KDLoop, Surviving change (axis)=Bayesian prior + sentinel MCQ filter + “A/B/C/D-only” (state-mem, obs, prompt)2026.06 | 46 | 3 | |
| ExploreEQAOptimizer=Baseline2026.06 | 43 | — |