Causal-HalBench: Uncovering LVLMs Object Hallucinations Through Causal Intervention

AI summary

Overview

Research area: Object hallucination in Large Vision-Language Models (LVLMs), studied through the lens of causal inference and spurious correlations.

Technical level: Intermediate. The paper builds on accessibility concepts (yes/no questions about whether an object is in an image) but introduces formal machinery from causal inference — Structural Causal Models, intervention, Average Causal Effect — that benefits from some background.

Scope: The paper formalizes co-occurrence-driven object hallucination in LVLMs with a Structural Causal Model, builds a counterfactual benchmark (Causal-HalBench) to measure it, and evaluates nine mainstream LVLMs on it.

What This Paper Is About

LVLMs are asked whether an object is present in an image, and they frequently answer incorrectly — a failure called object hallucination. The authors argue this happens largely because training data contain objects that co-occur very often (for example, skis and humans), so models learn a shortcut association instead of actually looking at the image. Existing benchmarks can detect hallucination but do not formally define or quantify these spurious correlations. This paper supplies a causal definition and a benchmark designed to measure how much each model leans on co-occurrence bias.

Key Contributions

1. First causal framing of LVLM object hallucination. The authors introduce a Structural Causal Model (SCM) over the input image (X), the question (Q), the model's output (Y), and a "Co-occurrence Bias" variable (C). C acts as a confounder, creating a backdoor path X ← C → Y through which spurious correlations form.

2. Visual Content Intervention (VCI) and causal metrics. VCI replaces a high-co-occurrence object with a low-co-occurrence one in an image to break the non-causal path. The paper defines Average Causal Effect (ACE) and Direct Causal Strength (DCS), then operationalizes them into three measurable metrics: Contextual object Accuracy Change (CAC), Absent object Accuracy Change (AAC), and Counterfactual object Hallucination Rate (CHR).

3. An automated, scalable counterfactual data pipeline. A three-stage pipeline — Intervention Objects Selection, Counterfactual Description Generation, and Counterfactual Inpainting — uses the proprietary LVLM Gemini for object selection and description, the Segment Anything Model (SAM) for masks (subsequently dilated), and FLUX-controlnet for inpainting.

4. Causal-HalBench. A benchmark built from MSCOCO validation images containing 757 original images, 1387 generated counterfactual images (2144 images in total), and 9709 image-question pairs (4161 with a "yes" answer and 5548 with "no").

Main Findings

• Nearly all evaluated models show high CAC. This means contextual objects become noticeably harder for models to recognize once the counterfactual edit removes the co-occurring cue, confirming broad vulnerability to spurious correlations.

• Qwen2.5-VL-7B has the lowest CAC (1.8%), but this comes alongside a lower Acc(Q_c) than other models — so the low score is not a clean win.

• Kimi-VL-A3B has by far the highest AAC (10.2%), well above InternVL2.5-8B (7.1%) and MiniCPM-o-2_6 (2.8%), with the remaining models lower still. High AAC means the model hallucinates related non-existent objects more when spurious cues are disrupted.

• CHR results split the field. LLaVA-NEXT-8B performs best at 6.8% and Kimi-VL-A3B is close at 7.4%; the worst are InternVL2.5-8B (29.3%) and Qwen2.5-VL-7B (27.3%).

• No single metric tells the whole story. Kimi-VL-A3B does well on CHR but worst on AAC; Qwen2.5-VL-7B does well on AAC but poorly on CHR. The authors argue this shows why single-accuracy evaluations are insufficient.

• Newer models can be more susceptible, not less. The paper reports that models such as Qwen2.5-VL, which perform well on general benchmarks, can show markedly stronger susceptibility to spurious correlations than earlier models like LLaVA-NEXT, which the authors attribute to larger training data scale.

• Closed-source models do not clearly beat open-source ones on Causal-HalBench.

• Causal-HalBench is harder than POPE. Accuracy is consistently lower on Causal-HalBench than on POPE across the seven models compared (for example, InternVL2.5-8B: 78.8 vs 88.7; Kimi-VL-A3B: 84.4 vs 88.5), while the relative ordering of models is broadly consistent.

• CHAIR hallucination worsens on counterfactual images. On the counterfactual-only subset, CHAIR-S and CHAIR-I scores rise for every model (for example, LLaVA-NEXT-8B: 29.0 → 38.8 for CHAIR-S and 13.7 → 18.1 for CHAIR-I), and models with higher CHAIR scores tend to have higher CHR.

• The intervention balances co-occurrence distributions. Heatmaps of object co-occurrence show the original dataset has concentrated high-frequency pairs, while the modified dataset increases previously low-frequency or absent pairs and dilutes existing strong correlations.

• Synthetic data quality is verified with CLIP Score. For the target object, the score drops from 26.5 (original) to 22.4 (synthetic); for the counterfactual object it rises from 20.6 to 27.5, using the prompt "a photo of {object}".

Methodology in Plain English

The authors start from the observation that image datasets are unbalanced: some object pairs show up together far more often than chance would suggest. A model trained on such data can learn "if I see a human, skis are likely nearby" rather than checking the pixels.

To describe this, they draw a causal diagram. The image, the question, and the model's answer are the visible pieces; hidden behind them is a "co-occurrence bias" variable that influences both which objects appear in images and what the model tends to say. Because that hidden variable touches both sides, the apparent relationship between what is in the image and what the model says is partly spurious.

Their fix is to intervene on the image directly. If you surgically replace an object with one that almost never co-occurs with the scene, you cut the shortcut and can see how much the model's answer depended on it. Building such images at scale requires a pipeline: an LVLM (Gemini) picks a replacement object that will look plausible, then writes a description of the edited scene; SAM produces a mask around the original object, the mask is dilated to give the inpainting model room, and FLUX-controlnet paints the replacement in. Human filtering removes bad results.

Because the original and edited images share a question and differ only in the object, the change in accuracy isolates the causal effect. Three metrics capture this: CAC (how much accuracy on the contextual object drops after the edit), AAC (how much accuracy on a plausible absent object rises after the edit), and CHR (how often the model reports seeing the newly painted counterfactual object when it is actually there — a measure of true visual perception). All results are zero-shot, run on a single A40 GPU.

Why This Matters

Impact on research. The paper shifts the hallucination conversation from "does the model hallucinate?" to "how much of the error is caused by a specific spurious shortcut?" That reframing gives researchers a way to compare models along a dimension that standard accuracy hides, and it shows that benchmark gains do not automatically translate into causal faithfulness.

Real-world applications:

• Accessibility tools that describe images for blind and low-vision users, where a hallucinated object can be actively misleading rather than merely unhelpful.
• Medical or scientific image analysis, where models trained on imbalanced data may assert the presence of findings that merely co-occur with what is actually visible.
• Autonomous systems and robotics, where a model that trusts co-occurrence priors over pixels may report obstacles or objects that are not there.
• Content moderation and cataloging pipelines, where co-occurrence shortcuts can produce systematic false labels on the long tail of rare object combinations.

Industry relevance. Because the pipeline is automated and built on off-the-shelf components (a proprietary LVLM, SAM, and an inpainting model), the same recipe can be reapplied to a company's own domain data to audit deployed multimodal models before release. The finding that larger, newer models can be more susceptible is a caution against assuming scale solves faithfulness. The code is released at https://github.com/zhexu-ustc/Causal-HalBench.

Future Directions

• Mitigation, not just measurement. The paper evaluates susceptibility but does not propose a training or inference-time method that reduces CAC, AAC, or CHR. Turning the causal analysis into a debiasing objective is the obvious next step.

• Closing the loop between the metrics. No model scored well on all three metrics at once. Understanding whether CAC, AAC, and CHR can be optimized jointly, or whether they trade off structurally, remains open.

• Extending beyond object presence. The SCM is defined for the "Is there {object} in this image?" question format. Generalizing to attributes, spatial relations, counting, and open-ended captioning would broaden the framework's reach.

• Better approximation of the theoretical intervention. The authors acknowledge that inpainting is only an approximation of do(X = x_cf), and that visual artifacts could act as interfering factors. Improving the fidelity of counterfactual generation, or quantifying the residual artifact effect, would strengthen the causal claims.

• Explaining why newer models are worse. The paper attributes greater susceptibility to larger training data scale, but this is a hypothesis rather than a tested mechanism.

Target Audience

Researchers and engineers working on multimodal model evaluation, vision-language hallucination, and trustworthy or faithful AI. It is also relevant to practitioners who deploy LVLMs in settings where a false claim about an object's presence carries real cost. Readers without a causal inference background will still follow the pipeline and the benchmark tables, but will need to work through the SCM and ACE/DCS definitions in Section 3 to fully appreciate why the metrics are defined the way they are.

Authors’ abstract

Large Vision-Language Models (LVLMs) often suffer from object hallucination, making erroneous judgments about the presence of objects in images. We propose this primar- ily stems from spurious correlations arising when models strongly associate highly co-occurring objects during train- ing, leading to hallucinated objects influenced by visual con- text. Current benchmarks mainly focus on hallucination de- tection but lack a formal characterization and quantitative evaluation of spurious correlations in LVLMs. To address this, we introduce causal analysis into the object recognition scenario of LVLMs, establishing a Structural Causal Model (SCM). Utilizing the language of causality, we formally de- fine spurious correlations arising from co-occurrence bias. To quantify the influence induced by these spurious correla- tions, we develop Causal-HalBench, a benchmark specifically constructed with counterfactual samples and integrated with comprehensive causal metrics designed to assess model ro- bustness against spurious correlations. Concurrently, we pro- pose an extensible pipeline for the construction of these coun- terfactual samples, leveraging the capabilities of proprietary LVLMs and Text-to-Image (T2I) models for their genera- tion. Our evaluations on mainstream LVLMs using Causal- HalBench demonstrate these models exhibit susceptibility to spurious correlations, albeit to varying extents.