SATE-Bench: 316 scenes show MLLMs stop exploring too early with high confidence
Tianchen Guo · hf · 2026-10-27
SATE-Bench is a new interactive benchmark (316 real scenes, 2,525 multi-turn QA pairs) testing whether MLLMs can actively explore occluded, human-centric 3D scenes by querying adjacent viewpoints, scored with step-penalized efficiency. Findings: SOTA models terminate exploration prematurely while reporting high confidence and struggle with fine-grained 3D geometry; matched-view controls attribute the Active–Passive gap to insufficient evidence acquisition. GRPO fine-tuning rewarding view-seeking boosts exploration but lowers accuracy via over-exploration and poor multi-view fusion. Scaling passive perception doesn't yield active spatial reasoning. Code and data are open.
More from Multimodal
- From SDXL to Krea2: Can Stacked LoRAs Be Replaced by Full Model Finetuning? — Bitter_Bag_3429 · 2026-10-28
- Gnani AI Launches Prisma v3.0 ASR for Code-Switched Indian Calls — CurieuxExplorer · 2026-10-28
- Designer shares UI concepts for an AI film studio — hewarsaber · 2026-10-28
- Open-source VoiceStudio offers local voice cloning and dubbing in 646 languages, free ElevenLabs rival — thisdudelikesAI · 2026-10-28
- A2O LITE shows off NEON infinite clip powered by 4D Gaussian Splatting — basedjensen · 2026-10-28
- LayerShop: open-source Photoshop-style UI wrapping ComfyUI generative AI, first preview on Windows — HonestTicket6013 · 2026-10-28