FINER: MLLMs Hallucinate under Fine-grained Negative Queries

1. Problem Statement

Multimodal Large Language Models (MLLMs) have made significant strides in visual perception but suffer from hallucinations—generating content not supported by the image. While existing benchmarks (e.g., POPE, DASH, AMBER) effectively detect coarse hallucinations (e.g., the presence/absence of a single object), they fail to capture fine-grained errors.

Real-world queries often involve complex combinations of multiple objects, attributes, and spatial relations. Current models struggle when a query contains a subtle contradiction (e.g., "a cat with brown fur sitting under the chair" when the cat is actually on the chair). The paper identifies that MLLMs exhibit a "brittleness" to fine-grained negative queries, where the model fails to reject a query that contains a single incorrect detail amidst otherwise correct information, leading to high false-positive rates.

2. Methodology

The authors propose a comprehensive framework consisting of new benchmarks, a data construction pipeline, and a novel training strategy.

A. FINER Benchmarks

The authors introduce FINER (FIne-grained NEgative queRies), comprising two benchmarks:

1. FINER-CompreCap: Based on human-annotated scene graphs from the CompreCap dataset (COCO images).
2. FINER-DOCCI: Based on dense, human-annotated captions from the DOCCI dataset, allowing for larger-scale evaluation beyond COCO object classes.

Construction Pipeline:

• Scene Graph Extraction: For DOCCI, a multi-stage pipeline using Gemini-2.0-Flash and Qwen2.5-VL-72B extracts objects, attributes, and relations from long captions, verified by human sampling to ensure accuracy.
• Negative Generation: For every positive element (Object, Attribute, Relation), the system generates four semantically plausible but incorrect negative counterparts (e.g., replacing "door frame" with "pillar").
• Query Construction: The benchmarks create four distinct settings to test specific hallucination types:
  • Multi-obj: Multiple objects, one is incorrect.
  • Multi-attr: Multiple attributes for an object, one is incorrect.
  • Multi-rel: Multiple relations, one is incorrect.
  • Wh-questions: "What" questions where the premise contains a false attribute (e.g., "What is the red dog doing?" when the dog is black).
• Evaluation Metric: To prevent models from simply memorizing "No" patterns, the benchmarks use Multiple Choice Questions (MCQs). The primary metric is Paired Accuracy ($Acc_{paired}$), which requires the model to answer both the positive query (Yes) and the corresponding negative query (No) correctly.

B. FINER-Tuning

To address these vulnerabilities, the authors propose FINER-Tuning, a training strategy based on Direct Preference Optimization (DPO).

• Data Generation: Using a scalable pipeline with the open-source Phi-4-14B model, they generate training data from long captions (Pixmo-caption).
  • Positive Phrases ($\Psi^+$): Extracted from captions (Objects, Attributes, Relations, Wh-contexts).
  • Negative Phrases ($\Psi^-$): Generated by modifying $\Psi^+$ to introduce fine-grained contradictions.
• Preference Tuples: The training data consists of pairs $(x, q, a^+, a^-)$ where:
  • $q^+$ (Positive Query) is paired with an accepted answer $a^+$ (confirming the image) and a rejected answer $a^-$ (denying the image).
  • $q^-$ (Negative Query) is paired with an accepted answer $a^+$ (denying the image) and a rejected answer $a^-$ (confirming the image).
• Training Objective: The model is optimized to rank the correct response ($a^+$) higher than the hallucinated response ($a^-$) for both positive and negative inputs, forcing the model to learn to detect subtle mismatches.

3. Key Contributions

1. FINER Benchmarks: The first benchmarks specifically designed to evaluate MLLMs on fine-grained negative queries involving multiple objects, attributes, and relations simultaneously.
2. FINER-Tuning: A model-agnostic, DPO-based training recipe that significantly reduces hallucinations without requiring costly closed-source models for data annotation (using Phi-4-14B instead).
3. Systematic Analysis: A comprehensive study revealing that MLLM performance degrades linearly as query granularity increases, and that current "hallucination-aware" fine-tuning methods (like RLAIF-V) do not generalize well to fine-grained errors.

4. Experimental Results

The authors evaluated frontier models (LLaVA-1.6, Qwen2.5-VL, InternVL-3.5) on FINER and other standard benchmarks.

• Performance on FINER:

  • Baseline models show a steep drop in accuracy as query granularity increases (e.g., InternVL-3.5-14B drops from ~75% on single-object queries to ~20% on complex multi-relation queries).
  • FINER-Tuning yields massive improvements. For InternVL-3.5-14B, it achieved up to 24.2% gain on the Multi-rel setting, outperforming its own 38B parameter version.
  • On FINER-DOCCI, FINER-Tuned InternVL-3.5-14B matched or approached the performance of Gemini-2.5-Flash in 3 out of 4 settings.

• Generalization to Other Benchmarks:

  • FINER-Tuning consistently improved performance on existing hallucination benchmarks (DASH, POPE, AMBER, HallusionBench, MMHal-Bench).
  • Notably, it reduced hallucination rates on generative benchmarks (e.g., lowering MMHal-Bench hallucination rate to 10% for InternVL-3.5-14B).
  • Unlike some prior methods that suffer from an "alignment tax" (improving safety at the cost of capability), FINER-Tuning improved general capabilities (MMStar, TextVQA, ChartQA) alongside hallucination reduction.

• Ablation Studies:

  • Training on both positive and negative queries is crucial; training only on negative queries or using Supervised Fine-Tuning (SFT) resulted in inferior performance or catastrophic forgetting.
  • Training on all four subsets (Obj, Attr, Rel, Wh) yielded the most balanced and robust results.

5. Significance and Conclusion

The paper demonstrates that current MLLMs are highly susceptible to fine-grained hallucinations, particularly when queries involve complex semantic combinations. The introduction of FINER provides a rigorous standard for evaluating these subtle failures, which were previously overlooked by coarse benchmarks.

The FINER-Tuning method proves that aligning models to reject fine-grained negative inputs via DPO is a highly effective strategy. It not only fixes specific hallucination types but also enhances the model's overall visual understanding and reasoning capabilities. This work suggests that future MLLM development must prioritize granular negation awareness to achieve true trustworthiness in real-world applications like medical imaging or autonomous driving, where a single attribute error can be critical.

Resources: Code, benchmarks, and models are publicly available at the project website.