Small Vision-Language Models Know When They Are Wrong But Cannot Say So: A Two-Model Study of Stated versus Internal Confidence Under Realistic Image Degradation
It is concluded that small VLMs encode usable self-knowledge that their verbalized output does not express, that internal probability is therefore the better deferral signal in constrained deployment, and that neither signal should be trusted under severe low-light conditions.
Abstract
Vision-language models (VLMs) are increasingly deployed on consumer hardware where input images are degraded by compression, camera shake, and poor lighting. In such settings, a reliable uncertainty signal matters more than raw accuracy, because it determines when a system should defer rather than answer. We evaluate two small open-weight VLMs -- Qwen2-VL-2B-Instruct and SmolVLM-Instruct -- across six realistic photographic degradations at three severity levels, comparing two confidence signals: the confidence the model states in natural language, and the model's own mean token probability over its generated answer. Across 3,800 predictions, we find a large and consistent gap. Verbalized confidence in Qwen2-VL is almost constant (mean 0.87-0.90 across all conditions) and detects its own errors at chance level (AUROC 0.39-0.75, typically ~0.50), while internal token probability from the same model separates correct from incorrect answers with AUROC 0.92-0.99. In SmolVLM, verbalized confidence proved largely unobtainable: across three prompt templates, only one of five pilot attempts produced a parseable confidence value, while internal probability again yielded above-chance error detection (AUROC 0.54-0.92). Both models fail in the same place: under severe underexposure, accuracy collapses (0.99->0.22 for Qwen2-VL, 0.97->0.42 for SmolVLM) while both confidence signals barely move, and internal error-detection falls to chance. We conclude that small VLMs encode usable self-knowledge that their verbalized output does not express, that internal probability is therefore the better deferral signal in constrained deployment, and that neither signal should be trusted under severe low-light conditions.
This work measures, on identical inputs, how model scale and 4-bit quantization affect two confidence signals in the Qwen2-VL family: the confidence a model states in natural language, and its own mean token probability over the answer it generates, and argues that error-detection AUROC is the metric that exposes the difference between the two signals.
Modality reliance in VLMs is not fixed, but varies across tasks, evidence structures, models, and evaluation settings, which shows that modality reliance in VLMs is not fixed, but varies across tasks, evidence structures, and evaluation settings.
Vision-language models are evaluated by aggregate accuracy on multimodal benchmarks, a practice that implicitly assumes the model uses its visual input. We show this assumption fails on 40%--97% of samples across six VLMs and three perceptual benchmarks: blurring the question-relevant visual region leaves the next-token distribution nearly unchanged. We name this phenomenon the Visual Insensitivity Gap and quantify it with a per-sample Visual Sensitivity Index (VSI). The gap is a property of samples, not of models: VSI ranks correlate across models (grand-mean Spearman rho=+0.40, permutation p<10^-3), so the same samples are flagged insensitive by VLMs sharing no architectural detail beyond a contrastively pretrained vision tower. The mechanism is concrete: on the insensitive samples, a linear probe on each model's own vision tower distinguishes perturbed from clean images at 0.72--0.79 accuracy, yet the model's argmax token changes on only 2%--11% of the same samples, an encoder--LLM gap above 0.65 on every model. Mapping VSI's diagnostic utility cell by cell surfaces a strong regime (multi-choice reasoning on capable VLMs: AUROC=0.85--0.87) and a weak regime (well-calibrated factuality, where softmax confidence already leads). VSI is not a universal best abstention signal; it is a sample-intrinsic indicator of vision-ignoring failure, best used as a conditional ensemble component.
Vision-language models (VLMs) have made rapid progress in visual perception and increasingly support real-world tasks that depend on images. Many such tasks, however, require more than rec- ognizing what an image contains: a model must use visual evidence to make a complete decision whose parts jointly satisfy global constraints. We introduce COMPLEXITYWORLD, a benchmark of 390 tasks across 39 domain-inspired visual worlds and 29 decision categories. Each task is generated from a hidden structured specification, rendered as a visual scene, and scored by an exe- cutable verifier that accepts any feasible solution. Under direct inference, all evaluated models ex- cept GPT-5.6-Sol remain below 40% verifier ac- ceptance rate (VAR), while GPT-5.6-Sol reaches 75.6%. Performance improves substantially when the same decision information is made explicit in structured form, yet varies sharply across equiva- lent visual presentations. Agent scaffolds provide smaller, model-dependent gains. Together, these results reveal a persistent visual-to-decision bot- tleneck that additional inference alone does not remove.
We address few-shot object detection with vision-language models (VLMs) in out-of-domain settings such as aerial, industrial, and medical imagery, using only ten annotated images for supervision. Existing adaptation methods are discrete prompt optimization and LoRA fine-tuning. We revisit a third option: soft prompting, where a small number of continuous prompt tokens are optimized while the pretrained backbone remains frozen. We identify two key design choices. First, placing prompt tokens at the cross-modal boundary between visual and text tokens outperforms other placements (10.0 vs. 8.4 mAP). Second, initializing prompts from the empty space token outperforms semantic and random initialization. With these choices, one to three learned tokens (7,168 parameters on average) match the best LoRA configuration on Roboflow20-VL (14.2 mAP, 10-shot) while training over 20,000x fewer parameters. Soft prompting remains harder to optimize, exhibiting higher variance across random seeds. Unlike LoRA, however, it causes no forgetting: the LoRA rank matching our accuracy reduces NaturalBench VQA accuracy by 35% relative, rising to 56% at the largest rank, whereas soft prompting leaves pretrained performance unchanged. The learned tokens behave like prompts rather than weights. They transfer to a newer model without retraining (+0.8 mAP on Qwen3.5-9B) and can be verbalized into readable prompts competitive with prompt-search methods (matching DetPO and outperforming GEPA). The approach also extends beyond detection. On RoboCasa manipulation tasks, the frozen $\pi_{0.5}$ vision-language-action policy benefits from soft prompting, matching the LoRA baseline on two of three tasks when tokens are placed at the gradient bottleneck. These results suggest modern VLMs already encode much of what is needed for specialized domains; the challenge is learning how to ask.
Gautam Rajendrakumar Gare, Si-Ying Li, He-Wei Wang et al.· 0 citations
For the coarse attributes the authors study, MLLMs encode the visual evidence but cannot reliably control their reliance on it, indicating that for the coarse attributes they study, MLLMs cannot reliably control their reliance on it.
Jiaang Li, Chengzu Li, Zhaochong An et al.· arXiv.org· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
