Introduction
Large Multimodal Models (LMMs) are becoming increasingly capable, yet their deployment in high-stakes environments—such as assistive technology for the visually impaired—requires more than just accuracy. It requires calibration: the ability of a model to accurately gauge its own certainty. For a user asking about a medication label or an expiration date, a "don't know" response is far safer than a confident hallucination. This research explores how to improve model calibration in Visual Question Answering (VQA) by introducing AVG BLEU, a novel calibration score based on semantic consistency.
The Challenge of Uncertainty in VQA
In VQA, traditional calibration methods often struggle. Unlike simple classification, open-ended answers can be phrased in multiple correct ways (e.g., "05/25" vs. "May 2025"). Furthermore, real-world images, such as those from the VizWiz dataset, are often blurry, poorly framed, or missing critical information, making many questions inherently unanswerable. A well-calibrated model should assign low confidence to these cases.
Methods: Measuring Semantic Consistency
The researchers evaluated standard calibration techniques alongside a new proposed metric:
- Likelihood: The standard approach of looking at the probability the model assigns to its own generated tokens.
- Sampling Repetition: Generating multiple answers and seeing how often the most frequent answer appears.
- Sampling Diversity: Measuring how many distinct answers the model generates across multiple samples.
- AVG BLEU (Proposed): This method calculates the average BLEU score across all pairs of sampled answers. By measuring semantic similarity rather than exact string matches, it recognizes that "a can of soup" and "soup can" represent the same underlying certainty.
The study compared LMMs (e.g., LLaVA, Flamingo), which process images directly, against LLMs (e.g., PaLM 2, Falcon), which rely on text captions to understand the visual context.
Datasets
Evaluation focused on "answerability" using two primary benchmarks:
- VizWiz-VQA: Real-world photos taken by visually impaired users, often featuring low quality, blur, or ambiguous subjects.
- UNK-VQA: A synthetic dataset designed to test a model's ability to identify when a question cannot be answered based on the provided image.
Key Insights
- Visual Grounding Improves Calibration: LMMs generally outperform LLMs at knowing when they lack information. Direct visual input seems to act as a regularizer, preventing the overconfidence often seen in caption-reliant LLMs.
- Hallucination Tendencies: LLMs are more prone to "hallucinating" answers when information is missing from a caption, whereas LMMs more frequently (and correctly) signal that a question is unanswerable.
- Robustness of Likelihood in LMMs: While likelihood is often a poor uncertainty proxy in text-only models, it proved surprisingly effective for LMMs, often outperforming simple sampling methods.
Key Results
The primary metric for success was Coverage@Acc, which determines the percentage of questions a model can answer while maintaining a specific accuracy target (e.g., 70% or 80%).
As shown in the results table, AVG BLEU emerged as the most consistent performer across architectures. It improved coverage at 80% accuracy by 5 percentage points for the best LMM and by 8 percentage points at 70% accuracy for the best LLM. While more complex semantic metrics exist, the study found that the computationally efficient BLEU score was sufficient to significantly improve the reliability of VQA systems.
Limitations and Future Work
While promising, the authors note that sampling-based methods like AVG BLEU are computationally expensive, requiring multiple model passes for a single query. Future research should focus on more efficient approximations of semantic similarity. Additionally, expanding testing to a wider variety of datasets and using model-generated captions for LLM comparisons (rather than human-written "gold" captions) will be essential for real-world validation.
Summary updated