CVPR 2026 oral: fine-grained negative queries make MLLMs hallucinate, DPO fix gains 24.2%
zeynepakata · x · 2026-08-22
FINER (Fine-grained NEgative queRies) shows multimodal LLMs readily hallucinate on fine-grained negative queries — questions whose details subtly mismatch the image — especially when the mismatch co-occurs with genuinely present elements.
Key contributions:
- Two benchmarks: FINER-CompreCap and FINER-DOCCI, covering multi-object, multi-attribute, multi-relation and "what" question settings
- FINER-Tuning: DPO on FINER-style data
- Finetuning four frontier MLLMs yields up to 24.2% gains (InternVL3.5-14B) on the new benchmarks, improving eight existing hallucination suites and general multimodal performance across six benchmarks
Code, benchmarks and models are open-sourced; the paper was an oral (award candidate) at CVPR 2026.
More from Multimodal
- Help Request: Methods for Removing Objects/Figures from Video — Glittering-Cold-2981 · 2026-08-22
- NVFP4 Version of LTX-2.5 Text Encoder Released for ComfyUI — superPussyman · 2026-08-22
- MiniMax-H3 Pruned Ref-Delta Model Converted to Single-File ComfyUI Format — marres · 2026-08-22
- AI Movies Leave the Demo Reel: A $2M Feature Film in 4 Weeks — lmoroney · 2026-08-22
- Phosphene 4.6.0: Local AI Video Editor with H3 and LTX 2.5 — Opening-Ad5541 · 2026-08-22
- H3 generates Tom and Jerry animation with smearing on fast action — Ok-Giraffe-8670 · 2026-08-22