TRoVe: Discovering Error-Inducing Static Feature Biases in Temporal Vision-Language Models

AI summary

Overview

Research area: Computer vision and vision-language modeling, specifically temporal understanding tasks (video activity recognition and longitudinal medical imaging) and model robustness/interpretability.

Technical level: Intermediate. The core idea is conceptually simple (cluster images, then score clusters), but evaluating it requires familiarity with vision-language models, contrastive embeddings, K-means clustering, and precision-based retrieval metrics.

Scope: The paper introduces TRoVe, an automated method for identifying the static visual cues (backgrounds, objects, attributes) that a temporal vision-language model wrongly relies on when classifying multi-image sequences, and it validates the method on a synthetic suite of 101 trained models plus 7 off-the-shelf models across 2 real-world tasks.

What This Paper Is About

Temporal vision-language models (VLMs) are supposed to classify things by analyzing how images change over a sequence — for example, recognizing a person climbing a tree from the motion across video frames. In practice, these models often take a shortcut and base their predictions on static features that happen to be present, such as a tree in the background, and then make systematic errors on other classes where that same feature appears. The paper's goal is to automatically discover which static features cause those errors for a given pretrained model, without needing any image-level annotations of those features.

Key Contributions

1. TRoVe, an automated bias-discovery method. Given a pretrained temporal VLM and a labeled validation dataset for a downstream classification task, TRoVe extracts candidate static features from the dataset and scores each one by how much it contributes to classification errors and how much the model relies on it.

2. A quantitative evaluation framework for temporal settings. Because ground-truth biases of trained models are normally unknown, the authors built a synthetic framework producing 101 temporal VLMs paired with ground-truth annotations of the injected static feature bias and the class label it harms. This addresses the paper's stated gap: prior error-discovery methods are designed for non-temporal settings and are hard to evaluate in temporal settings.

3. A demonstration on real-world models and tasks. TRoVe is applied to 7 off-the-shelf VLMs across 2 temporal understanding tasks — activity recognition on Kinetics400 (400 classes) and pneumonia progression classification on MS-CXR-T (2 classes) — surfacing previously unknown static feature biases.

4. A test-time mitigation strategy. The paper shows that knowledge of TRoVe-discovered biases can improve accuracy without data augmentation or VLM retraining, reporting improvements of up to 111% on an activity recognition task for sequences containing static features.

Main Findings

• TRoVe beats all evaluated baselines on the synthetic benchmark. It achieves a 28.6% improvement over the closest baseline, which the paper identifies as Confidence (an application of the method in Li et al. [23]).

• Near-perfect precision on some feature categories. On background features, TRoVe scores 100.0 precisely at Precision@10, 100.0 at Precision@25, 100.0 at Precision@100, and 95.0 R-precision. On attribute features it scores 100.0 / 100.0 / 99.3 / 89.3. On object features it scores 97.8 / 97.8 / 97.8 / 92.8.

• Non-temporal baselines transfer poorly to temporal inputs. When Domino, George, and Distilling Failures are naively adapted to multi-image sequences, they struggle to retrieve the specific images containing the static feature. For example, Domino scores 48.6 / 47.0 / 32.4 / 19.7 on background features and George scores 45.5 / 44.7 / 42.0 / 35.6, versus TRoVe's 100.0 / 100.0 / 100.0 / 95.0.

• TRoVe is consistent across feature types, while baselines vary. The paper states that baselines show significant performance variation across background, object, and attribute biases, whereas TRoVe remains strong across all three.

• Real-world biases found in activity recognition. For six pretrained VLMs evaluated on activity recognition, TRoVe identifies between 36 and 116 learned static feature biases per model. One example: for VideoCLIP-XL, TRoVe surfaces a cluster of babies, suggesting degraded performance on the class "sticking tongue out."

• Real-world biases found in medical imaging. For the one pretrained VLM evaluated on pneumonia progression classification, TRoVe identifies 4 learned static feature biases. A discovered cluster of chest X-rays with bilateral opacities, medical devices, and low lung volumes suggests degraded performance for BioViL-T on the class "improving."

• Discovered biases were independently validated. For activity recognition, using an open-vocabulary object detector to annotate babies, VideoCLIP-XL accuracy was lower by 15.4 points on the class "sticking tongue out" when babies were present, and the model was highly confident on static sequences containing babies. Incorrectly classified sequences in that class were predicted as "baby waking up" and "carrying baby." For pneumonia, using radiology reports and a domain-specific entity extraction tool, BioViL-T accuracy was lower by 16.7 points when severe-pneumonia markers were present, with high confidence on static sequences containing them.

• Non-temporal VLMs show more static biases than temporal VLMs. Across four non-temporal VLMs evaluated on activity recognition, TRoVe discovers an average of 134.5 ± 38.8 static feature biases per model, versus an average of 84.5 ± 27.3 for the six temporal VLMs.

• Mitigation improves Accuracy@5 on affected sequences. On Kinetics400 videos containing the top 20 TRoVe-identified static features, VideoCLIP-XL goes from 51.7 to 94.4 on label ỹ and from 82.2 to 86.7 overall; ViCLIP-B goes from 45.3 to 95.8 on label ỹ and from 73.4 to 77.7 overall; ViCLIP-L goes from 71.4 to 96.9 on label ỹ and from 77.1 to 80.7 overall.

Methodology in Plain English

The approach has two stages.

Stage one — find candidate static features. TRoVe takes a labeled validation set of multi-image sequences and breaks each sequence into its individual images. To get an image-level representation from a model that expects sequences, it builds a "static sequence" by repeating a single image as many times as that image's original sequence length, effectively erasing all motion, and feeds that through the model's vision encoder. It then groups all these image embeddings into clusters using spherical K-means with cosine distance, automatically picking the number of clusters that maximizes the Silhouette score. Each cluster represents a recurring visual feature, such as trees or babies.

Stage two — score each cluster. Two scores are computed per cluster and class label:

• The Error Contribution Score (ECS) subtracts the model's accuracy on sequences containing at least one image from the cluster from its accuracy on sequences containing none, for a given class label. It ranges from -1 to 1, and large positive values mean the feature is associated with degraded performance on that class.
• The Static Bias Score (SBS) looks only at images from mispredicted sequences of that class, feeds each one in as a repeated static sequence, and averages the model's calibrated confidence (via temperature scaling) on the wrong label it produced. It ranges from 0 to 1. The reasoning: a temporal task by definition needs changing content, so a model that is confident given an unchanging input is likely leaning on a static bias.

The two are summed into the TRoVe score, and clusters are ranked.

How it was evaluated. The synthetic framework generates sequences of a blue circle (diameter 10 pixels) on a 60×60 black background, moving north, south, west, or east, with captions describing the direction. One of three static features is injected — a red background, a red 15×15-pixel rectangle placed at a random position, or a red circle — and made highly prevalent in "moving south" sequences and rare elsewhere, so the trained model picks up the shortcut. Hyperparameters varied include the feature type, sequence length, feature prevalence, and how many images per sequence show the feature. After quality verification, this yields 101 VLMs with known ground-truth biases. Performance is measured with Precision@K (K = 10, 25, 100) — the share of the top-K returned images that truly depict the ground-truth bias — plus R-precision, where K equals the total number of images in the validation set annotated with the ground-truth bias.

How it was applied in the real world. For the mitigation experiment, TRoVe produces a ranked list of clusters and associated error-prone labels. The authors then use CoOp to learn prompts that maximize accuracy on validation sequences containing at least one image from cluster C, keeping all VLM parameters frozen. At test time, the clustering model first checks whether an incoming sequence contains an image from cluster C; if so, the learned prompts are used, otherwise default prompts are used. Performance is reported as Accuracy@5.

Why This Matters

Impact on research. The paper reframes robustness analysis for temporal VLMs: instead of mitigating static biases during training through optimization tricks or data augmentation, it treats bias discovery as a post-hoc diagnostic on an already-trained model. It also provides a reusable synthetic evaluation framework, since ground-truth biases of pretrained models are normally unavailable — a barrier the paper explicitly identifies as blocking quantitative progress in this area. The finding that non-temporal VLMs carry more static biases on average (134.5 ± 38.8) than temporal VLMs (84.5 ± 27.3) gives a concrete baseline for future work.

Real-world applications:

• Clinical deployment of longitudinal imaging models. A model like BioViL-T that lowers its accuracy by 16.7 points when devices or bilateral opacities appear could mislead radiologists about whether pneumonia is improving.
• Video content moderation and activity recognition. Biases tied to people or objects (e.g., babies) mean systems may mislabel actions whenever those objects appear.
• Pre-deployment auditing of any temporal VLM. The method requires only a trained model and a labeled validation set, so it can be run as a standard audit step.
• Targeted, low-cost model repair. Because the mitigation freezes all model parameters and only learns prompts, teams can improve affected cases without retraining or new data collection.

Industry relevance. Organizations deploying vision-language systems on video or longitudinal imagery need to know where those systems fail before shipping them. TRoVe offers an automated alternative to manual inspection with pixel-wise interpretability tools such as GradCAM, which the paper notes require extensive human effort and become harder as sequence length grows. The prompt-learning mitigation is attractive commercially because it avoids retraining costs.

Future Directions

• Generalizing beyond contrastive VLMs. The mitigation case study focuses specifically on contrastive temporal VLMs; whether the same prompt-learning repair works for other architectures is left open.
• Better real-world validation of discovered biases. The authors state that evaluating discovered biases in real-world settings is challenging and rely on image-level pseudolabels plus downstream performance gains as proxies; stronger validation methods remain an open problem.
• Handling features that appear in only a fraction of a sequence. The paper highlights this as a core difficulty and tests it in the synthetic framework, but the limits of the approach as sequences grow longer and features become sparser are not fully characterized in the provided content.
• Extending beyond classification tasks. The formal problem setting is restricted to downstream temporal understanding tasks formulated as classification problems; other temporal formulations are not addressed.

Target Audience

Researchers and engineers working on vision-language models, video understanding, and longitudinal medical imaging who need to audit model behavior before deployment. It is also relevant to robustness and interpretability researchers interested in shortcut learning and systematic error discovery, and to practitioners who want a low-cost, training-free way to improve performance on the specific cases where their models fail. Readers should be comfortable with embeddings, clustering, and retrieval-style evaluation metrics; the paper's terminology is otherwise self-contained.

Note: the provided paper content is truncated (Appendix A cuts off mid-sentence), and the sizes of the real-world evaluation datasets for Kinetics400 and MS-CXR-T are not reported in the available text.

Authors’ abstract

Vision-language models (VLMs) have made great strides in addressing temporal understanding tasks, which involve characterizing visual changes across a sequence of images. However, recent works have suggested that when making predictions, VLMs may rely on static feature biases, such as background or object features, rather than dynamic visual changes. Static feature biases are a type of shortcut and can contribute to systematic prediction errors on downstream tasks; as a result, identifying and characterizing error-inducing static feature biases is critical prior to real-world model deployment. In this work, we introduce TRoVe, an automated approach for discovering error-inducing static feature biases learned by temporal VLMs. Given a trained VLM and an annotated validation dataset associated with a downstream classification task, TRoVe extracts candidate static features from the dataset and scores each feature by (i) the effect of the feature on classification errors as well as (ii) the extent to which the VLM relies on the feature when making predictions. In order to quantitatively evaluate TRoVe, we introduce an evaluation framework consisting of 101 trained temporal VLMs paired with ground-truth annotations for learned static feature biases. We use this framework to demonstrate that TRoVe can accurately identify error-inducing static feature biases in VLMs, achieving a 28.6% improvement over the closest baseline. Finally, we apply TRoVe to 7 off-the-shelf VLMs and 2 temporal understanding tasks, surfacing previously-unknown static feature biases and demonstrating that knowledge of learned biases can aid in improving model performance at test time. Our code is available at https://github.com/Stanford-AIMI/TRoVe.