Technical Summary: SpurLens: Automatic Detection of Spurious Cues in Multimodal LLMs

Problem Statement

While unimodal vision models are well-documented to rely on spurious correlations (non-essential input attributes), the extent to which Multimodal Large Language Models (MLLMs) exhibit similar biases remains unclear, despite their language supervision. The authors hypothesize that many critical failures in MLLMs, specifically object hallucination (generating factually incorrect content) and object recognition failures, stem from a reliance on spurious cues rather than true object recognition.

Existing methods for detecting these biases often require human supervision, are limited to predefined features, or rely on computationally expensive caption co-occurrence analysis. Furthermore, it is unclear whether these biases are fundamental to the vision encoder, the language model, or their fusion, and whether simple prompting strategies can mitigate them.

Methodology: SpurLens Pipeline

The paper introduces SpurLens, an automated, unsupervised pipeline designed to identify spurious visual cues and quantify their impact on MLLM performance. The pipeline operates in three main stages:

1. Proposing Spurious Features:

   • For a given target object $t$, the system leverages GPT-4 to generate a list of potential spurious cues (objects or background elements) that commonly co-occur with $t$.
   • The generated list undergoes rigorous filtering via GPT-4 to ensure features are:
     • Physical and visually discernable.
     • Not part of the target object itself (e.g., "screen" is not a cue for "laptop").
     • Not inseparable from the target object.
     • Detectable by open-set object detectors.

2. Identifying and Ranking Images:

   • An open-set object detector (OWLv2) is used to scan a large image dataset for the presence of the proposed spurious features.
   • Images are ranked based on the confidence scores of the detected spurious cues. This creates a "spuriosity" ranking for each feature, allowing the system to select images with the highest presence of a cue (Top-K) and the lowest presence (Bottom-K).

3. Computing Spurious Gaps:

   • The ranked images are fed into the target MLLM with binary prompts (e.g., "Is there a [target object] in the image?").
   • Two core metrics are calculated:
     • Perception Accuracy (PA) Gap: The difference in recognition accuracy between images containing the spurious cue and those without it. A large positive gap indicates over-reliance on the cue for recognition.
     • Hallucination Rate (HR) Gap: The difference in hallucination rates (false positives) between images containing the cue and baseline images. A large positive gap indicates the cue triggers hallucinations.
   • The feature yielding the largest gap is identified as the most influential spurious cue for that class.

Key Contributions

• Automated Discovery: Unlike prior work requiring human annotation or fixed feature sets, SpurLens automatically discovers interpretable spurious cues across a broad range of objects and datasets without human supervision.
• Dual Failure Mode Analysis: The study systematically quantifies two distinct failure modes caused by spurious bias:
  1. Over-reliance: Performance drops significantly when spurious cues are removed.
  2. Hallucination Amplification: Spurious cues trigger object hallucinations at rates significantly higher than baseline.
• Diagnostic Depth: The framework isolates the source of bias through ablation studies, examining whether the bias originates in the vision encoder, the language model, or the fusion process, and tests the efficacy of various mitigation strategies.

Experimental Results

The authors evaluated five MLLMs (Qwen2-VL, Qwen-3.5, Llama-3.2, LLaVA-v1.6, and GPT-4o-mini) across three datasets (HardImageNet, ImageNet Subset, and COCO).

1. Object Recognition Failures

• Removing spurious cues leads to significant accuracy drops across all models.
• Example: GPT-4o-mini's accuracy on COCO dropped from 88.0% (with cues) to 67.6% (without cues), a 20.4% decrease.
• The magnitude of the drop varies by model and dataset, with LLaVA-v1.6 showing a 16.0% gap on COCO.

2. Object Hallucination

• Spurious cues dramatically amplify hallucination rates.
• Example: In COCO, the hallucination rate for GPT-4o-mini increased from 1.4% (no cues) to 11.2% (with cues), a 9.8x increase.
• In the ImageNet subset, the increase was even more pronounced, with some models showing up to an 18-fold increase in hallucination rates when spurious cues were present.

3. Ablation Studies & Mitigation

• Token Dropping: Even when visual tokens corresponding to the target object were artificially removed, models continued to hallucinate the object's presence at high rates (e.g., HR > 30% for some models), suggesting the bias is not solely dependent on visual evidence.
• Vision Encoder Isolation: Training binary classifiers on vision encoder embeddings alone revealed significant spurious gaps, indicating that the bias exists within the visual representation itself, independent of the language model.
• Prompting Strategies: The authors tested several mitigation techniques, including prompt ensembling, Chain-of-Thought (CoT) reasoning, and explicitly informing the model of potential spurious cues.
  • Result: While some strategies offered minor situational improvements, none significantly reduced the Spurious Gap. Even "cheating" by providing the model with the exact strongest spurious cue identified by SpurLens failed to eliminate the bias.

Significance and Claims

The paper claims that spurious bias is a deeply rooted, fundamental issue in current MLLMs that persists across different architectures and datasets. The authors assert that:

1. Current MLLMs are highly susceptible to spurious correlations, leading to systematic failures in both recognition and generation.
2. Simple mitigation strategies are insufficient. Language-based interventions (prompting) cannot fully resolve the underlying reliance on spurious cues, as the bias appears embedded in the vision encoder's feature space.
3. Rigorous evaluation is needed. The persistence of these biases calls for more robust evaluation methods and mitigation strategies beyond current prompting techniques to ensure the reliability of MLLMs in real-world applications.

The study concludes that SpurLens provides a scalable, interpretable tool for diagnosing these failures, but the root causes require more than just surface-level adjustments to model prompting.