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

TL;DR

<3-5 sentence high-level summary> TRoVe tackles the problem that temporal vision-language models sometimes rely on static, non-temporal features, causing systematic errors on time-based tasks. It automatically discovers these error-inducing static biases by clustering image-level representations, then scoring clusters by their impact on errors (ECS) and the model’s reliance on the static feature (SBS), outputting bias–class pairs. The authors validate TRoVe in synthetic settings with 101 VLMs and demonstrate a 28.6% improvement over baselines, then extend to real-world models and tasks, showing the biases surface across multiple domains and that bias-aware mitigation (via CoOp prompts) can greatly improve test-time accuracy (up to 111%). The work provides a practical framework for robustness evaluation and practical bias mitigation for temporal multimodal models, with code released for reproducibility.

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.

Figures (9)

• Figure 1: Temporal models often rely on the presence of static feature biases, such as background features or objects in the scene, when making predictions. In this example, the VideoCLIP-XL model wang-etal-2024-videoclip makes systematic prediction errors on the class swinging on something when trees are present.
• Figure 2: TRoVe is an automated approach for discovering error-inducing static feature biases learned by temporal VLMs. In this example, TRoVe identifies a static feature bias associated with trees, which results in degraded performance on the class label swinging on somethingkay2017kineticshumanactionvideo.
• Figure 3: Example five-image sequences with injected static visual features. Our evaluation framework considers three types of static visual features - background, object, and attribute - that can contribute to biased model predictions on temporal understanding tasks.
• Figure 4: [Left Panel] Qualitative examples of static feature biases discovered by TRoVe across various temporal VLMs and downstream tasks. [Right Panel] We demonstrate that TRoVe-discovered biases satisfy desired properties.
• Figure 5: TRoVe demonstrates strong performance across evaluation configurations with varying numbers of images per sequence.