Position Bias in VLMs

We describe our submission to Subtask~1b of ImageEval 2026. The task consists of selecting the single visually valid statement from three candidates paired with an image, and is in both English and Modern Standard Arabic (MSA). Our system uses a pre-trained 7B vision--language model in a zero-shot setting, without task-specific training or fine-tuning. We formulate the task as a constrained selection problem and choose the answer using next-token log-probabilities. To reduce answer-position bias, we apply a Latin-square rotation to the candidate statements. Our system achieves CI scores of 0.0510 on the English test set and 0.0960 on the MSA test set, reducing CI relative to the official 3B baselines reported on devtest by 83.7% and 80.4%, respectively.

Vision-language models (VLMs) increasingly act as judges that pick the best of several generated images, so their choices decide what users see. Such judges are usually validated by score agreement with human ratings, not by the images they return. We audit VLM judges as decision-makers: on 300 culturally situated prompts, we compare the returned image with human ratings the judge never sees and with random choice from the same candidates, and repeat every decision with the candidates reordered. A 4B-parameter judge barely beats random and falls short of a CLIP similarity baseline. It picks the first image shown in 49% of calls (chance: 28%), and reordering changes its choice on 60% of prompts. For this judge, agreement across orders is informative: decisions that survive reordering are much better than random, whereas agreement with a weaker second judge keeps the wrong ones. An 8B judge shows almost no position bias and outperforms CLIP, yet for it the same filter mostly discards good decisions. Agreement helps only when it targets the judge's failure mode, so filters must be re-audited whenever the judge changes. The 4B judge's slight rise in stereotype ratings is no longer detectable after aggregating across orders or with the larger judge.

Medical visual question answering (Med-VQA) is often assumed to require medical fine-tuning, large models, or complex multi-agent pipelines. We revisit this assumption with MedProb, a lightweight probing framework that predicts multiple-choice Med-VQA answers from frozen VLM representations without free-text generation. Across PATH-VQA, SLAKE, and VQA-RAD, MedProb recovers substantially more answer-relevant signal than prompting and performs stronger than medical VLMs and agentic systems. Probing also reduces the apparent gap between small and large models compared to prompting, suggesting that smaller VLMs contain more recoverable Med-VQA signal than generation-based evaluation reveals. Across 14 matched general-purpose and medical VLM pairs, medical adaptation does not consistently improve this linear decodability. Finally, free-text generation exhibits an answer-position bias of up to 10 percentage points, whereas MedProb also has positional bias, however, it is impacted differently than prompting. Our main results target the multiple-choice/multiclass Med-VQA setting; we additionally show the probe can be extended to open-ended generation via a rejection-sampling scoring procedure.

As generative multimedia evolves from static image synthesis to complex, interleaved visual narratives, a foundational bottleneck has emerged: the judgment crisis. While human perception naturally synthesizes the temporal and logical flow of a story, automated evaluation systems remain largely "blind" to sequential continuity, often failing to distinguish between a coherent narrative and a semantically shuffled or contradictory sequence. This work identifies a critical structural gap in current multimodal evaluation paradigms, arguing that the reliance on Large Vision-Language Models (LVLMs) as judges is fundamentally limited by architectural biases. Our analysis reveals a profound performance dichotomy: while models may appear competent in isolated pointwise scoring, they suffer a catastrophic collapse when required to perform pairwise discrimination of temporal order. We demonstrate that this is not merely a data-scarcity issue but a structural one. Through a series of diagnostic probes, we uncover systematic positional asymmetries, specifically primacy and recency effects, where a model's judgment of a story is significantly influenced by the placement of a frame, often more than by its semantic consistency. These biases, potentially rooted in causal masking and rotary embeddings, suggest that current transformer-based judges are inherently ill-equipped for long-form visual reasoning. By exposing these blind spots, we challenge the multimedia community to move beyond snapshot-centric metrics and instead pioneer Temporally-Aware Evaluation paradigms that treat visual sequences as unified logical structures rather than unordered collections of frames.

Standard accuracy metrics for VLMs often mask significant reliability failures in sensitive domains. In this work, we utilize a histopathology-validated brain MRI dataset to systematically assess the diagnostic robustness of four VLM families under evidence-preserving perturbations. By reordering anatomical slices and swapping target label positions, we evaluate whether models maintain consistent predictions when clinical evidence remains invariant. Our results reveal significant vulnerabilities in presentation-order stability, with models exhibiting prediction flips in up to 48.9% of cases under simple sequence reversals. We further identify a textual selection bias, where label reordering triggers inconsistent diagnoses in up to 67.8% of cases despite identical visual inputs. Negative-control tests further reveal diagnostic overcommitment: models generate categorical diagnoses in up to 76.1% of cases after expert-annotated lesion slices are removed. These results demonstrate that high accuracy can overestimate clinical reliability, masking sensitivity to sequential presentation and textual framing that is not captured by aggregate accuracy. Our findings highlight the necessity of stability-based metrics for the deployment of VLMs in safety-critical clinical applications. Our evaluation data and code will be made public upon acceptance.

Vision language models (VLMs) have achieved strong performance on visual document understanding benchmarks such as DocVQA, ChartQA, and MMLongBench-Doc. However, real-world documents combine multiple factors such as length, layout complexity, modality, and question difficulty, which makes it difficult to attribute model failures to specific causes. We introduce SynthDocBench, a fully synthetic benchmark for long-context visual document understanding that systematically controls factors including document length, layout structure, modality composition, and question type. The benchmark is constructed using a combinatorial design, each factor is varied independently across generated documents, enabling controlled analysis of model behavior. Documents are generated end to end using an LLM pipeline across six layout archetypes, with a 40 percent random override to prevent models from exploiting spurious correlations. Additionally, SynthDocBench spans long-context documents with substantially greater length and structural diversity than existing benchmarks. Evaluating seven frontier VLMs, we uncover three failure modes that existing benchmarks cannot surface: sharp degradation with document length, a systematic positional sensitivity in which the middle third of a document is hardest for five of six models and five of six models show a negative Early-to-Late trend (steepest decline: 8.3 percentage points), and breakdown of chart comprehension in long-document settings. These results suggest that current models may be overfitting to benchmark artifacts rather than achieving robust long-context visual document understanding.

Vision-language models (VLMs) achieve strong performance on spatial reasoning benchmarks, yet it remains unclear whether this reflects structured 3D understanding or reliance on statistical shortcuts in natural images. We introduce a representation-level analysis framework that constructs minimal contrastive pairs to measure how spatial axes are organized and disentangled within VLM embeddings. Our analysis across multiple model families reveals a consistent vertical-distance entanglement: models conflate vertical image position with distance, mirroring the perspective bias of natural photographs. This bias produces a significant accuracy gap between perspective-consistent and counter-heuristic examples, and intensifies under data scaling even as overall benchmark accuracy improves. We further show that models with similar benchmark scores can exhibit different internal representations, and that these differences predict accuracy and robustness across diverse spatial reasoning benchmarks. To isolate this bias from evaluation-set skew, we introduce SpatialTunnel, a synthetic benchmark designed to expose spatial shortcut biases by removing common correlations present in natural images. Experiments confirm that the entanglement is model-intrinsic, and that models with well-separated spatial axes exhibit greater robustness, suggesting that well-structured spatial representations lead to more reliable spatial reasoning across diverse benchmarks. Code and benchmark are available on the project page: https://cheolhong0916.github.io/whyfarlooksup.github.io/.

Multimodal judges struggle to ground decisions in visual evidence. We present MJ1, a multimodal judge trained with reinforcement learning that enforces visual grounding through a structured grounded verification chain (observations $\rightarrow$ claims $\rightarrow$ verification $\rightarrow$ evaluation $\rightarrow$ scoring) and a counterfactual consistency reward that penalizes position bias. Even without training, our mechanism improves base-model accuracy on MMRB2 by +3.8 points on Image Editing and +1.7 on Multimodal Reasoning. After training, MJ1, with only 3B active parameters, achieves 77.0% accuracy on MMRB2 and surpasses orders-of-magnitude larger models like Gemini-3-Pro. These results show that grounded verification and consistency-based training substantially improve multimodal judgment without increasing model scale.

CLIP models learn transferable multi-modal features via image-text contrastive learning on internet-scale data. They are widely used in zero-shot classification, multi-modal retrieval, text-to-image diffusion, and as image encoders in large vision-language models. However, CLIP's pretraining is dominated by images paired with short captions, biasing the model toward encoding simple descriptions of salient objects and leading to coarse alignment on complex scenes and dense descriptions. While recent work mitigates this by fine-tuning on small-scale long-caption datasets, we identify an important common bias: both human- and LLM-generated long captions typically begin with a one-sentence summary followed by a detailed description. We show that this acts as a shortcut during training, concentrating attention on the opening sentence and early tokens and weakening alignment over the rest of the caption. To resolve this, we introduce DeBias-CLIP, which removes the summary sentence during training and applies sentence sub-sampling and text token padding to distribute supervision across all token positions. DeBias-CLIP achieves state-of-the-art long-text retrieval, improves short-text retrieval, and is less sensitive to sentence order permutations. It is a drop-in replacement for Long-CLIP with no additional trainable parameters.