Uncertainty Calibration and Abstention in Vision Language Models

24 concepts18 papers1 proposal

When a Vision-Language Model (VLM) is asked to describe an image, it usually responds with remarkable fluency. However, this linguistic confidence is highly deceptive. If you present a VLM with an unanswerable question—such as identifying an object hidden in a blurry photo, or diagnosing a rare eye disease from an out-of-domain medical scan—it will rarely admit ignorance. Instead, it will confidently fabricate a plausible-sounding lie.

This tendency toward overconfidence poses severe risks in safety-critical deployments like autonomous robotics, medical diagnostics, and assistive technologies for the visually impaired. To build truly reliable multimodal systems, researchers are shifting focus away from simply making models smarter, and toward making them uncertainty-aware. The goal is to give VLMs a mathematically grounded way to say: "I don't know."

The Root of Overconfidence: Linguistic Fluency vs. Visual Grounding

Why do VLMs hallucinate with such high confidence? The core issue stems from a structural mismatch between how these models process visual data and how they generate text.

When a VLM generates an answer, standard uncertainty estimation methods look at the probability distribution of the output tokens. However, as Erum Mushtaq et al. demonstrate in their work on HARMONY, output token probabilities often reflect fluent linguistic patterns learned during text pre-training rather than actual comprehension of the visual input. A model might assign a high probability to a word simply because it makes grammatical sense in context, even if it is entirely unsupported by the image.

Furthermore, the model's internal vision branch often does perceive the correct information, but it gets drowned out during the decoding phase. Dingchen Yang et al. discovered that when a VLM hallucinates, it frequently assigns high confidence scores to both the correct and incorrect visual tokens simultaneously. The model struggles to separate the true visual signal from deceptive noise, leading to over-compliance and a failure to reject invalid premises.

Probing the Internal State for Hidden Knowledge

If the final text output cannot be trusted, how can we detect when a model is guessing? Researchers have found that the model's internal states and early token distributions contain "hidden knowledge" that reveals its true uncertainty.

1. The First-Token Logit Distribution

One of the most immediate indicators of model confusion happens at the very beginning of text generation. Qinyu Zhao et al. showed that the logit distribution of the very first token contains a wealth of hidden information. Even if the model ultimately outputs a confident hallucination, the entropy and distribution of that first token can accurately signal whether a question is unanswerable or deceptive.

2. Probabilistic Embeddings

Standard contrastive models like CLIP map images and text into rigid, deterministic vectors. This approach lacks a native way to express ambiguity. To fix this, Aishwarya Venkataramanan et al. introduced GroVE, a framework that uses Gaussian Process Latent Variable Models to transform these rigid embeddings into probabilistic distributions. By representing inputs as distributions rather than points, the model can naturally quantify its own epistemic uncertainty. Similarly, the authors of Intra-Class Probabilistic Embeddings for Uncertainty Estimation in Vision-Language Models leverage the distribution of visual features within specific classes to flag misclassifications without requiring expensive retraining.

3. Multi-Modal Fusion of Gut Feelings

Rather than relying on text or vision alone, frameworks like HARMONY train a small auxiliary transformer to fuse the VLM's internal "gut feeling" (its hidden layer activations) with its external confidence (token probabilities). This allows the system to catch instances where the text generation layer is confidently running away from what the visual layers actually see.

Benchmarking the Ability to Abstain

To systematically fix this overconfidence, the research community has developed benchmarks specifically designed to penalize models that refuse to admit ignorance.

   [ Input Prompt ] ---> ( Is the information present in the image? )
                                /                         \
                             YES                           NO
                             /                             \
             [ Standard VQA Task ]               [ Unsolvable Problem Detection ]
             Goal: Correct Answer                Goal: Abstain / Say "I don't know"

A foundational framing for this is Unsolvable Problem Detection (UPD), introduced by Atsuyuki Miyai et al.. UPD tests a model's capacity to not answer when given absurd, missing, or contradictory visual-textual pairs.

Subsequent benchmarks have highlighted just how pervasive this failure mode is:

  • Domain Gaps: Zhenyue Qin et al. introduced LMOD, showing that state-of-the-art models like GPT-4o fail to recognize the boundaries of their medical competence, attempting to diagnose complex ophthalmic conditions rather than abstaining.
  • Textual Underspecification: The FOCUS benchmark by Kankan Zhou et al. evaluates how well models handle ambiguous sentences that require precise visual integration, finding that models often hallucinate interpretations rather than asking for clarification.
  • Real-World Constraints: Benchmarks like TUBench (Xingwei He et al.) and VizWiz-LF (Mina Huh et al.) focus on unanswerable questions stemming from poor image quality or missing context, particularly emphasizing the harm caused to blind users who cannot visually verify a model's confident but false assertion.
  • Robotic Risks: In embodied AI, ZhiYuan Feng et al. demonstrate that over-compliance causes robots to execute unsafe actions because the underlying VLM hallucinates spatial relationships rather than rejecting an invalid command.

Calibrating the Response: Abstention vs. Coverage

Fixing overconfidence requires a delicate balancing act. If a model is penalized too harshly for making mistakes, it may become overly cautious, leading to unnecessary abstention—refusing to answer questions it actually possesses the internal knowledge to solve.

To navigate this trade-off, researchers are developing inference-time calibration strategies:

  • Semantic Perturbations: The VL-Uncertainty framework by Ruiyang Zhang et al. measures consistency by applying slight blurs to images or rephrasing questions. If the model's answers vary wildly under these minor perturbations, it indicates high internal confusion, triggering an automatic abstention.
  • Retrieval Contrastive Decoding (RCD): To actively suppress visual noise during decoding, Dingchen Yang et al. propose retrieving similar images and mathematically "subtracting" the common hallucinatory patterns they induce, isolating the true visual cues.
  • Active Interrogation: To prevent excessive caution, Jack Hessel et al. developed ReCoVERR. When a VLM expresses low confidence, an external LLM "interrogates" the VLM for supporting evidence. If the VLM can successfully justify its hunch, the system overrides the initial hesitation, safely increasing the model's coverage without sacrificing accuracy.

By shifting from raw accuracy metrics to confidence-weighted metrics—such as those proposed by Khyathi Chandu et al. in CertainlyUncertain—the field is moving toward a paradigm where a model's ability to accurately gauge its own ignorance is treated as just as important as its ability to provide the right answer.

Key Foundational Work

Know What You do Not Know: Verbalized Uncertainty Estimation Robustness on Corrupted Images in Vision-Language Models (2025) — contributed 3 concept elements to this direction.
Unsolvable Problem Detection for Vision Language Models (2024) — contributed 3 concept elements to this direction.
CertainlyUncertain: A Benchmark and Metric for Multimodal Epistemic and Aleatoric Awareness (2025) — contributed 2 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_vlm_overconfidence_29c56fc5,
  title = {VLM overconfidence},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/vlm-overconfidence/rdr_29c56fc5432748119bfd3b865f36478f},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_vlm_overconfidence_29c56fc5_v1,
  title = {VLM overconfidence},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/vlm-overconfidence/rdr_29c56fc5432748119bfd3b865f36478f/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake