# Benchmark Designers Should "Train on the Test Set" to Expose Exploitable Non-Visual Shortcuts

arXiv: https://arxiv.org/abs/2511.04655 · HTML: https://mlsanity.com/paper/2511.04655/

## Key facts

- Cited by 29 papers in the map (data through 14 Oct 2026).
- Placed in MLLM training and visual instruction tuning, one of 16 directions in Multimodal LLMs / vision-language models.
- Builds on 41 of its 57 references that are in the map.
- First posted on arXiv 6 Nov 2025.

## Placed in

- Multimodal LLMs / vision-language models / [MLLM training and visual instruction tuning](https://mlsanity.com/mllm/mllm-training-and-visual-instruction-tuning/)

## Builds on (most cited first, 30 of 41)

- [Microsoft COCO: Common Objects in Context](https://mlsanity.com/paper/1405.0312/) (2014)
- [GPT-4o System Card](https://mlsanity.com/paper/2410.21276/) (2024)
- [HellaSwag: Can a Machine Really Finish Your Sentence?](https://mlsanity.com/paper/1905.07830/) (2019)
- [ScanNet: Richly-annotated 3D Reconstructions of Indoor Scenes](https://mlsanity.com/paper/1702.04405/) (2017)
- [LLaVA-OneVision: Easy Visual Task Transfer](https://mlsanity.com/paper/2408.03326/) (2024)
- [Making the V in VQA Matter: Elevating the Role of Image Understanding in Visual Question Answering](https://mlsanity.com/paper/1612.00837/) (2016)
- [VQA: Visual Question Answering](https://mlsanity.com/paper/1505.00468/) (2015)
- [MMBench: Is Your Multi-modal Model an All-around Player?](https://mlsanity.com/paper/2307.06281/) (2023)
- [Qwen2 Technical Report](https://mlsanity.com/paper/2407.10671/) (2024)
- [MMMU: A Massive Multi-discipline Multimodal Understanding and Reasoning Benchmark for Expert AGI](https://mlsanity.com/paper/2311.16502/) (2023)
- [Shortcut Learning in Deep Neural Networks](https://mlsanity.com/paper/2004.07780/) (2020)
- [Video-MME: The First-Ever Comprehensive Evaluation Benchmark of Multi-modal LLMs in Video Analysis](https://mlsanity.com/paper/2405.21075/) (2024)
- [Are We on the Right Way for Evaluating Large Vision-Language Models?](https://mlsanity.com/paper/2403.20330/) (2024)
- [Cambrian-1: A Fully Open, Vision-Centric Exploration of Multimodal LLMs](https://mlsanity.com/paper/2406.16860/) (2024)
- [ScanNet++: A High-Fidelity Dataset of 3D Indoor Scenes](https://mlsanity.com/paper/2308.11417/) (2023)
- [Annotation Artifacts in Natural Language Inference Data](https://mlsanity.com/paper/1803.02324/) (2018)
- [Thinking in Space: How Multimodal Large Language Models See, Remember, and Recall Spaces](https://mlsanity.com/paper/2412.14171/) (2024)
- [LLaVA-Video: Video Instruction Tuning With Synthetic Data](https://mlsanity.com/paper/2410.02713/) (2024)
- [Winoground: Probing Vision and Language Models for Visio-Linguistic Compositionality](https://mlsanity.com/paper/2204.03162/) (2022)
- [MMMU-Pro: A More Robust Multi-discipline Multimodal Understanding Benchmark](https://mlsanity.com/paper/2409.02813/) (2024)
- [The Hateful Memes Challenge: Detecting Hate Speech in Multimodal Memes](https://mlsanity.com/paper/2005.04790/) (2020)
- [ARKitScenes: A Diverse Real-World Dataset For 3D Indoor Scene Understanding Using Mobile RGB-D Data](https://mlsanity.com/paper/2111.08897/) (2021)
- [Don't Just Assume; Look and Answer: Overcoming Priors for Visual Question Answering](https://mlsanity.com/paper/1712.00377/) (2017)
- [SWAG: A Large-Scale Adversarial Dataset for Grounded Commonsense Inference](https://mlsanity.com/paper/1808.05326/) (2018)
- [Hypothesis Only Baselines in Natural Language Inference](https://mlsanity.com/paper/1805.01042/) (2018)
- [Counterfactual VQA: A Cause-Effect Look at Language Bias](https://mlsanity.com/paper/2006.04315/) (2020)
- [RUBi: Reducing Unimodal Biases in Visual Question Answering](https://mlsanity.com/paper/1906.10169/) (2019)
- [Adversarial Filters of Dataset Biases](https://mlsanity.com/paper/2002.04108/) (2020)
- [Overcoming Language Priors in Visual Question Answering with Adversarial Regularization](https://mlsanity.com/paper/1810.03649/) (2018)
- [Towards Interpreting and Mitigating Shortcut Learning Behavior of NLU Models](https://mlsanity.com/paper/2103.06922/) (2021)

## Built on by (most cited first, 29 of 29)

- [Scaling Spatial Intelligence with Multimodal Foundation Models](https://mlsanity.com/paper/2511.13719/) (2025)
- [Beyond Language Modeling: An Exploration of Multimodal Pretraining](https://mlsanity.com/paper/2603.03276/) (2026)
- [Thinking with Geometry: Active Geometry Integration for Spatial Reasoning](https://mlsanity.com/paper/2602.06037/) (2026)
- [SpatialClaw: Rethinking Action Interface for Agentic Spatial Reasoning](https://mlsanity.com/paper/2606.13673/) (2026)
- [S-Agent: Spatial Tool-Use Elicits Reasoning for Spatial Intelligence](https://mlsanity.com/paper/2606.20515/) (2026)
- [Revealing Perception and Generation Dynamics in LVLMs: Mitigating Hallucinations via Validated Dominance Correction](https://mlsanity.com/paper/2512.18813/) (2025)
- [Cambrian-P: Pose-Grounded Video Understanding](https://mlsanity.com/paper/2605.22819/) (2026)
- [Diagnosing Corruption-Induced Reliability Failures in Vision-Language Models](https://mlsanity.com/paper/2511.19032/) (2025)
- [What does RL improve for Visual Reasoning? A Frankenstein-Style Analysis](https://mlsanity.com/paper/2602.12395/) (2026)
- [Native Active Perception as Reasoning for Omni-Modal Understanding](https://mlsanity.com/paper/2606.19341/) (2026)
- [Towards Spatial Supersensing in the Wild](https://mlsanity.com/paper/2607.13681/) (2026)
- [MMGist: A Comprehensive Multimodal Benchmark for 2027](https://mlsanity.com/paper/2606.22437/) (2026)
- [Towards Physics of Multimodal Pretraining: Knowledge Flow, Modality Synergy, Early Unification, and Recipes](https://mlsanity.com/paper/2608.05000/) (2026)
- [Solving Spatial Supersensing Without Spatial Supersensing](https://mlsanity.com/paper/2511.16655/) (2025)
- [From Where Things Are to What They Are For: Benchmarking Spatial-Functional Intelligence in Multimodal LLMs](https://mlsanity.com/paper/2605.02130/) (2026)
- [Ablate-to-Validate: Are Vision-Language Models Really Using Continuous Thought Tokens?](https://mlsanity.com/paper/2605.21642/) (2026)
- [Beyond 3D VQAs: Injecting 3D Spatial Priors into Vision-Language Models for Enhanced Geometric Reasoning](https://mlsanity.com/paper/2605.30231/) (2026)
- [Imaginative Perception Tokens Enhance Spatial Reasoning in Multimodal Language Models](https://mlsanity.com/paper/2606.03988/) (2026)
- [4DP-QA: Scalable QA for 4D Perception in Vision Language Models](https://mlsanity.com/paper/2606.11568/) (2026)
- [From Hallucination to Grounding: Diagnosing Visual Spatial Intelligence via CRISP](https://mlsanity.com/paper/2606.26535/) (2026)
- [From Accuracy to Visual Dependence: Auditing and Filtering Modality Collapse in Traffic VideoQA](https://mlsanity.com/paper/2606.30220/) (2026)
- [Lifting Unlabeled Internet-level Data for 3D Scene Understanding](https://mlsanity.com/paper/2604.01907/) (2026)
- [CaMo: Camera Motion Grounded Evaluation and Training for Vision-Language Models](https://mlsanity.com/paper/2605.20165/) (2026)
- [REKEY: Metadata-Grounded Visual-Key Regeneration for Contamination-Resilient VQA Evaluation](https://mlsanity.com/paper/2606.20736/) (2026)
- [Judging by the Cover: Cleaning LLM Truthfulness Benchmarks to Avoid Surface-Level Feature Leakage](https://mlsanity.com/paper/2609.13003/) (2026)
- [PhysBrain 1.5: From Vision-Language Models to Physical Foundation Models](https://mlsanity.com/paper/2609.14973/) (2026)
- [ReTAC: Residual Test-Time Agentic Computation for Video Spatial Reasoning](https://mlsanity.com/paper/2610.16315/) (2026)
- [EvoSpatial: Experience-Driven Skill and Tool Evolution for Spatial Reasoning Agents](https://mlsanity.com/paper/2610.16679/) (2026)
- [GeoThinker-F: Endogenous Predictive World Representations for Fast Spatial Intelligence](https://mlsanity.com/paper/2610.20592/) (2026)

## Abstract (arXiv)

Multimodal LLMs can answer many questions in vision-centric benchmarks without looking at the image, using linguistic priors and skewed answer distributions. Blind evaluation shows what a model already answers this way, but not what it could learn from the test set's own regularities. Extending partial-input auditing (e.g., hypothesis-only baselines in natural language inference), we argue that benchmark designers should "train on the test set": probe the artifact they release for exploitable patterns. Our Test-set Stress-Test (TsT) cross-validates a text-only Qwen2-7B on the test set's questions and answer options, yielding a benchmark-level score and a per-question bias score s(x); a random forest on hand-crafted features adds a fast, interpretable audit. On the template-based VSI-Bench and CV-Bench, the held-out score is 17.9 and 13.1 points above the model's own zero-shot score. On MMMU it learns little, even though GPT-4o correctly answers 52.3% of its multiple-choice questions without the image, so blind success there reflects pretrained knowledge rather than learnable test-set patterns. Iterative Bias Pruning (IBP) removes the questions with the highest s(x) and re-diagnoses; on VSI-Bench it widens a fine-tuned model's vision-blind gap more than random removal at the same rate. We also release VSI-Bench-Debiased, which lowers a fine-tuned model's blind score from 44.7 to 32.0 while its vision score falls only from 57.1 to 48.7.