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

• URL: http://arxiv.org/abs/2511.04655v1
• Date: Thu, 06 Nov 2025 18:43:21 GMT
• Title: Benchmark Designers Should "Train on the Test Set" to Expose Exploitable Non-Visual Shortcuts
• Authors: Ellis Brown, Jihan Yang, Shusheng Yang, Rob Fergus, Saining Xie,
• Abstract summary: We find that models can ace many benchmarks without strong visual understanding.<n>This is especially problematic for vision-centric benchmarks that are meant to require visual inputs.<n>We adopt a diagnostic principle for benchmark design: if a benchmark can be gamed, it will be gamed.
• Score: 49.99400612296149
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Robust benchmarks are crucial for evaluating Multimodal Large Language Models (MLLMs). Yet we find that models can ace many multimodal benchmarks without strong visual understanding, instead exploiting biases, linguistic priors, and superficial patterns. This is especially problematic for vision-centric benchmarks that are meant to require visual inputs. We adopt a diagnostic principle for benchmark design: if a benchmark can be gamed, it will be. Designers should therefore try to ``game'' their own benchmarks first, using diagnostic and debiasing procedures to systematically identify and mitigate non-visual biases. Effective diagnosis requires directly ``training on the test set'' -- probing the released test set for its intrinsic, exploitable patterns. We operationalize this standard with two components. First, we diagnose benchmark susceptibility using a ``Test-set Stress-Test'' (TsT) methodology. Our primary diagnostic tool involves fine-tuning a powerful Large Language Model via $k$-fold cross-validation on exclusively the non-visual, textual inputs of the test set to reveal shortcut performance and assign each sample a bias score $s(x)$. We complement this with a lightweight Random Forest-based diagnostic operating on hand-crafted features for fast, interpretable auditing. Second, we debias benchmarks by filtering high-bias samples using an ``Iterative Bias Pruning'' (IBP) procedure. Applying this framework to four benchmarks -- VSI-Bench, CV-Bench, MMMU, and VideoMME -- we uncover pervasive non-visual biases. As a case study, we apply our full framework to create VSI-Bench-Debiased, demonstrating reduced non-visual solvability and a wider vision-blind performance gap than the original.

Related papers

• IF-RewardBench: Benchmarking Judge Models for Instruction-Following Evaluation [85.56193980646981]
  We propose IF-RewardBench, a comprehensive meta-evaluation benchmark for instruction-following.<n>For each instruction, we construct a preference graph containing all pairwise preferences among multiple responses.<n>Experiments on IF-RewardBench reveal significant deficiencies in current judge models.
  arXiv  Detail & Related papers  (2026-03-05T02:21:17Z)
• Visually Prompted Benchmarks Are Surprisingly Fragile [82.98001690512461]
  Key challenge in evaluating VLMs is testing their ability to analyze visual content independently from their textual priors.<n>We demonstrate how details in benchmark setup, including visual marker design and dataset size, have a significant influence on model performance and leaderboard rankings.<n>To mitigate this instability, we curate existing datasets to create VPBench, a larger visually prompted benchmark with 16 visual marker variants.
  arXiv  Detail & Related papers  (2025-12-19T18:26:58Z)
• How Many Code and Test Cases Are Enough? Evaluating Test Cases Generation from a Binary-Matrix Perspective [51.30005925128432]
  evaluating test cases automatically generated by Large Language Models (LLMs) is a critical yet challenging task.<n>Existing benchmarks suffer from high computational costs, score inflation, and a bias towards trivial bugs over rare, critical faults.<n>We introduce a framework that formalizes benchmark construction as finding an optimal diagnostic basis in a binary code-test matrix.
  arXiv  Detail & Related papers  (2025-10-09T18:29:24Z)
• Benchmark Profiling: Mechanistic Diagnosis of LLM Benchmarks [34.09939383415074]
  Benchmark Profiling decomposes benchmark performance into ten cognitively grounded abilities.<n>It explains why performance gains do not always translate into user-perceived competence.
  arXiv  Detail & Related papers  (2025-09-23T15:32:47Z)
• Do Large Language Model Benchmarks Test Reliability? [66.1783478365998]
  We investigate how well current benchmarks quantify model reliability.<n>Motivated by this gap in the evaluation of reliability, we propose the concept of so-called platinum benchmarks.<n>We evaluate a wide range of models on these platinum benchmarks and find that, indeed, frontier LLMs still exhibit failures on simple tasks.
  arXiv  Detail & Related papers  (2025-02-05T18:58:19Z)
• Leaving the barn door open for Clever Hans: Simple features predict LLM benchmark answers [10.786564839628952]
  Internal validity of AI benchmarks is crucial for ensuring they are free from confounding factors. We investigate the possibility that AI systems can solve benchmarks in unintended ways, bypassing the capability being tested.
  arXiv  Detail & Related papers  (2024-10-15T15:05:41Z)
• LiveXiv -- A Multi-Modal Live Benchmark Based on Arxiv Papers Content [62.816876067499415]
  We propose LiveXiv: a scalable evolving live benchmark based on scientific ArXiv papers.<n>LiveXiv accesses domain-specific manuscripts at any given timestamp and proposes to automatically generate visual question-answer pairs.<n>We benchmark multiple open and proprietary Large Multi-modal Models (LMMs) on the first version of our benchmark, showing its challenging nature and exposing the models true abilities.
  arXiv  Detail & Related papers  (2024-10-14T17:51:23Z)
• Do Text-to-Vis Benchmarks Test Real Use of Visualisations? [11.442971909006657]
  This paper investigates whether benchmarks reflect real-world use through an empirical study comparing benchmark datasets with code from public repositories. Our findings reveal a substantial gap, with evaluations not testing the same distribution of chart types, attributes, and actions as real-world examples. One dataset is representative, but requires extensive modification to become a practical end-to-end benchmark. This shows that new benchmarks are needed to support the development of systems that truly address users' visualisation needs.
  arXiv  Detail & Related papers  (2024-07-29T06:13:28Z)
• Scalable Performance Analysis for Vision-Language Models [26.45624201546282]
  Joint vision-language models have shown great performance over a diverse set of tasks. Our paper introduces a more scalable solution that relies on already annotated benchmarks. We confirm previous findings that CLIP behaves like a bag of words model and performs better with nouns and verbs.
  arXiv  Detail & Related papers  (2023-05-30T06:40:08Z)
• Zero-shot Model Diagnosis [80.36063332820568]
  A common approach to evaluate deep learning models is to build a labeled test set with attributes of interest and assess how well it performs. This paper argues the case that Zero-shot Model Diagnosis (ZOOM) is possible without the need for a test set nor labeling.
  arXiv  Detail & Related papers  (2023-03-27T17:59:33Z)
• On the Limits of Evaluating Embodied Agent Model Generalization Using Validation Sets [101.28658250723804]
  This paper experiments with augmenting a transformer model with modules that effectively utilize a wider field of view and learn to choose whether the next step requires a navigation or manipulation action. We observe that the proposed modules resulted in improved, and in fact state-of-the-art performance on an unseen validation set of a popular benchmark dataset, ALFRED. We highlight this result as we believe it may be a wider phenomenon in machine learning tasks but primarily noticeable only in benchmarks that limit evaluations on test splits.
  arXiv  Detail & Related papers  (2022-05-18T23:52:21Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.