Technical Summary: ReactBench

Problem Statement

Multimodal Large Language Models (MLLMs) have demonstrated significant progress in visual-language understanding but remain prone to multimodal hallucinations, where models generate semantically coherent responses inconsistent with visual input. While existing benchmarks (e.g., POPE, AMBER) have played a foundational role in diagnosing these issues, they exhibit three critical limitations as MLLMs advance:

1. Simplistic Scenarios: Many rely on basic scenes and limited query formats, leading to saturation by state-of-the-art models and inflated performance metrics that fail to probe true capability boundaries.
2. Isolated Cause Analysis: Recent efforts often investigate specific causes (e.g., language priors or co-occurrence bias) in isolation, lacking a unified, cause-driven framework. Furthermore, some benchmarks rely on text-to-image generation for adversarial samples, potentially compromising real-world representativeness.
3. Lack of Fine-Grained Diagnosis: Existing evaluations typically focus on aggregate performance or coarse-grained summaries of hallucination behaviors, rarely utilizing Chain-of-Thought (CoT) reasoning traces to trace hallucinated outputs back to specific sub-causes within the reasoning process.

Methodology

To address these gaps, the authors introduce ReactBench, a systematic, cause-driven benchmark designed to shift the evaluation paradigm from outcome-oriented detection to cause-oriented diagnosis. The methodology encompasses four targeted tasks, each mapping to a specific failure mechanism, and a semi-automated pipeline for dataset construction.

1. Task Definitions

ReactBench evaluates MLLMs across four tasks corresponding to four representative causes of hallucination:

• Relational Erasure (Co-occurrence Bias): Targets hallucinations induced by statistical co-occurrence. High-frequency object triplets (e.g., skateboard-wheels) are selected, and one entity is visually erased. Models are queried about the existence of the erased object using four dimensions: direct questioning, commonsense-guided induction, premise-based induction, and absurd-contrast induction.
• Counterfactual Attribute (Language Priors): Targets reliance on prior knowledge over visual evidence. High-frequency object-attribute tuples (e.g., squirrel-soft grey fur) are edited to counterfactual states (e.g., spiky quills). Models are evaluated on their ability to identify the edited attribute against the same four query dimensions.
• Alteration Tracing (Cross-Image Comparative Perception): Targets deficiencies in comparing image pairs. Objects in an image undergo micro-alterations (movement, removal, addition, replacement). Models must identify changes across three settings: location-hinted, operation-hinted, and no-hinted queries.
• Dense Counting (Fine-Grained Perceptual Bottlenecks): Targets difficulties in counting multiple objects of the same class in dense scenes. Models are queried for exact quantities or interval-based ranges.

2. Dataset Construction Pipeline

The benchmark construction involves a semi-automated pipeline ensuring high fidelity:

• Source Data: Images are sourced from Visual Genome (for the first three tasks) and FSC147 (for Dense Counting).
• Statistical Filtering: Candidate images are selected based on co-occurrence probabilities and object frequencies to ensure strong bias or prior conditions.
• MLLM-Based Editing: Instance-level information is fed to an MLLM to generate precise editing instructions, which are executed by an image editing model (Qwen-Image-Edit-2511) to create adversarial samples.
• Exam-Style QA Generation: MLLMs generate large-scale question-answer pairs (MC, YN, SA) based on the edited images and metadata, incorporating specific inducing dimensions (e.g., absurd contrasts).
• Human Verification: A two-stage human review process validates image editing quality and QA correctness, ensuring visual fidelity and answer accuracy.

3. Evaluation Framework

• Metrics: Beyond standard accuracy, the authors propose React-Score, a hierarchical weighted metric that accounts for question difficulty (Short-Answer weighted at 50%, MC/YN at 25%) and task difficulty (Relational Erasure and Dense Counting weighted at 30% each).
• Fine-Grained Sub-Cause Analysis: The authors utilize CoT reasoning traces to perform a diagnostic analysis. They define a taxonomy of sub-causes for each task (e.g., Activity-Entity Completion for Relational Erasure, Blind Prior Assertion for Counterfactual Attribute) and use an LLM to attribute hallucinated responses to these specific sub-causes.

Key Results

Extensive evaluations were conducted on 13 advanced open-source MLLMs (including Qwen3.5, InternVL3, LLaVA series, and MiMO-VL).

• Overall Performance: Current MLLMs remain notably vulnerable, with average accuracies ranging roughly between 40% and 60%. The top-performing model, Qwen3VL-32B-Thinking, achieved a React-Score of 65.3, while earlier architectures like LLaVA-v1.6-7B scored only 34.5.
• Effect of Chain-of-Thought (CoT): The introduction of CoT reasoning yielded diverging effects. While it helped mitigate language-prior-driven errors in the Counterfactual Attribute task (e.g., +8.0% improvement for Qwen3VL-32B-Instruct), it often degraded performance on perception-intensive tasks like Alteration Tracing and Dense Counting. This suggests that lengthy reasoning chains may amplify visual uncertainty and accumulate intermediate misinterpretations.
• Task-Specific Insights:
  • Perception vs. Priors: Models showed clearer gains on perception-oriented tasks (Alteration Tracing, Dense Counting) compared to tasks rooted in pre-training data distributions (Relational Erasure, Counterfactual Attribute), where performance stalled slightly above 60%.
  • Question Formats: Open-ended Short-Answer (SA) questions proved significantly harder (19.1%–28.4% accuracy) than Multiple-Choice (50.4%–64.9%) or Yes/No formats, indicating a lack of visual grounding for free-form generation.
  • Inducing Dimensions: "Absurd-contrast" queries were highly effective at inducing hallucinations, causing the most severe performance drops.
• Sub-Cause Distribution: The analysis revealed dominant failure patterns:
  • Relational Erasure: 68% of errors stemmed from Activity-Entity Completion and Scene-Object Association.
  • Counterfactual Attribute: 83.1% of errors were due to Blind Prior Assertion and Common Attribute Projection.
  • Alteration Tracing: 89.7% of errors resulted from Cross-Image State Assimilation and Fine-Grained Blindness.
  • Dense Counting: Regular Shortcut Error (49.1%) and Sequential Enumeration Deficiency (30.9%) were the primary bottlenecks.

Significance and Claims

The paper claims that ReactBench provides a systematic and interpretable testbed for diagnosing multimodal hallucinations. Its significance lies in:

1. Shifting the Paradigm: Moving from simple outcome detection to a cause-driven framework that systematically exposes specific failure mechanisms (co-occurrence bias, language priors, comparative perception deficits, and fine-grained bottlenecks).
2. Diagnostic Granularity: Offering a fine-grained sub-cause taxonomy that traces hallucinations back to specific reasoning breakdowns, providing actionable insights for model improvement beyond generic "larger scale" approaches.
3. Real-World Fidelity: By using real images with targeted edits rather than synthetic text-to-image generation, the benchmark maintains higher visual fidelity and real-world representativeness.
4. Guidance for Future MLLMs: The results suggest that future models require targeted mitigation strategies, such as suppressing plausible but visually unsupported priors and strengthening visual verification mechanisms (e.g., localized comparison and object-level tracking), rather than relying solely on scaling.

The authors conclude that ReactBench not only serves as an evaluation benchmark but also offers actionable guidance for improving the robustness of future MLLMs.