FINER discovers that MLLMs are highly prone to hallucination when images contain fine-grained mismatches co-occurring with real elements.
March 19, 2026
Original Paper
FINER: MLLMs Hallucinate under Fine-grained Negative Queries
arXiv · 2603.17662
The Takeaway
It exposes a specific, previously overlooked failure mode in vision-language models. By proposing 'FINER-Tuning' with DPO, it shows how to significantly boost hallucination resistance (up to 24%) across multiple frontier models and benchmarks.
From the abstract
Multimodal large language models (MLLMs) struggle with hallucinations, particularly with fine-grained queries, a challenge underrepresented by existing benchmarks that focus on coarse image-related questions. We introduce FIne-grained NEgative queRies (FINER), alongside two benchmarks: FINER-CompreCap and FINER-DOCCI. Using FINER, we analyze hallucinations across four settings: multi-object, multi-attribute, multi-relation, and ``what'' questions. Our benchmarks reveal that MLLMs hallucinate whe