EviScene brings agentic factuality assessment to AI images with 79% accuracy
ZiAn Wang · hf · 2026-10-14
A new paper formulates Image Factuality Assessment: beyond asking "is it AI-generated?", a system verifies whether an image's depicted situation is factually supported, returning a supported/refuted verdict with evidence-grounded explanation.
- EviLens benchmark: 10,174 expert-reviewed supported/refuted images across 12 factual error types, scoring both verdict accuracy and evidence sufficiency
- EviScene is a multimodal ReAct agent interleaving visual inspection, external retrieval, and evidence analysis, progressively refining factual questions during investigation
- Training combines SFT on reviewed teacher trajectories with privileged on-policy self-distillation on verified decisions
- A Qwen3.5-9B agent reaches 79.05% balanced accuracy and a 52.26% Strict Evidence Sufficiency Rate on a 1,527-image test set — the highest SESR among evaluated agents, beating the base agent by 6.71 and 9.37 points
Code, models, and data will be open-sourced.
More from Research
- FlashVSR+ runs untiled 4K video super-resolution on one RTX 4090 with 2x speedup — MMLab-CUHK · 2026-10-14
- Sber's GreenWAM world foundation model spans 20 robot families as both predictor and policy — SberRoboticsCenter · 2026-10-14
- Tencent's LifeSearch benchmark: best deep research agent scores just 63.4% on real-life tasks — Tencent-Hunyuan · 2026-10-14
- Science: Menopause may explain why women face higher dementia risk than men — AnnaCiaunica · 2026-10-14
- When will AI start delivering breakthroughs in physics like it has in math? — burny_tech · 2026-10-14
- Photonic Training of a 1B GPT-2 Without Backprop Lands in PNAS — IgorCarron · 2026-10-14