💬 NLP

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

This study reveals that while small vision-language models possess accurate internal self-knowledge about their errors under realistic image degradation, they fail to verbalize this uncertainty, making internal token probability a superior signal for deferral than stated confidence, though both signals become unreliable under severe low-light conditions.

Original authors: M M Asif Ferdous

Published 2026-07-27

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: M M Asif Ferdous

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are walking through a bustling city with a very smart, but tiny, robot guide. This robot has a camera for eyes and a voice for speaking. Its job is to look at things around you and tell you what they are. But here's the catch: the city is messy. Sometimes the sun is too bright, sometimes it's pitch black, and sometimes your camera lens is smudged or shaking. In the world of artificial intelligence, these "messy" pictures are called degraded images.

When a robot like this gets confused by a bad picture, it needs to know when to say, "I don't know, ask a human," instead of confidently guessing the wrong answer. This ability to know when it is unsure is called uncertainty. There are two ways a robot can show this uncertainty. The first is verbalized confidence: the robot simply says out loud, "I am 90% sure!" The second is internal confidence: a secret, silent signal inside the robot's brain that measures how shaky its thoughts are, even if it doesn't say a word. Most people assume that if a robot says it's confident, it probably is. But what if the robot is just a bad liar? That is the big question this paper asks.

The Story of the Silent Signal

In this study, researchers put two small, open-source robot guides (called Qwen2-VL and SmolVLM) through a rigorous test. They didn't just show them clean, perfect photos. Instead, they took 100 pictures of food and deliberately ruined them in six different ways: squashing the quality like a JPEG file, blurring them with motion, dimming the lights until they were almost dark, adding blinding glare, tilting the camera, or shrinking and stretching the image. They did this at three levels of severity for each type of ruin.

The robots had to guess what the food was in every single ruined picture. Then, the researchers checked two things:

Did the robot say, "I'm confident!" or "I'm not sure"? (The Verbalized Signal)
What was the robot's secret, internal score of how likely it was to be right? (The Internal Signal)

The Big Surprise: The Robot That Won't Admit It's Wrong

The results were a bit like watching a magician who is terrible at their job but keeps bowing and saying, "That was perfect!"

The Verbalized Signal (The Robot's Mouth):
When asked to state their confidence, the robots were incredibly stubborn. No matter how terrible the picture was, no matter how blurry or dark it got, they kept saying they were about 87% to 90% sure. Even when the picture was so dark the robot was basically guessing randomly (getting the answer right only 22% of the time), it still confidently said, "I'm 87% sure!" It was as if the robot had a broken confidence dial stuck on "High." In fact, for one of the robots, asking it to give a confidence number was so difficult that it mostly just refused to answer, giving only the name of the food without any number at all.

The Internal Signal (The Robot's Brain):
However, when the researchers peeked inside the robot's brain to see its secret "token probabilities" (a fancy way of measuring how sure the robot's own math was), they found something totally different. The robot's internal signal was actually very honest! When the picture was clear, the internal score was high. When the picture was blurry or dark, the internal score dropped significantly. This secret signal was so good at spotting mistakes that it could tell the difference between a right answer and a wrong one with 92% to 99% accuracy.

The "Low Light" Trap

There was one scary moment in the story, though. When the researchers made the pictures extremely dark (the "severe underexposure" level), both signals failed. The robots' accuracy crashed (dropping from 99% down to 22%), and even the honest internal signal stopped working. It couldn't tell the difference between a right guess and a wrong one anymore. It was as if the darkness was so thick that even the robot's secret brain fogged up.

What This Means for You

So, what did the researchers learn?

Don't trust the robot's mouth. If you ask a small robot guide how sure it is, it will likely lie and say it's very sure, even when it's totally lost. The paper shows that this "stated confidence" is useless for knowing when to stop and ask a human for help.
Trust the robot's brain (mostly). The secret internal math is much better. It knows when it's confused and can signal that to a computer system, allowing the system to say, "Okay, I'm not sure, let's get a human to check this."
But watch out for the dark. Even the honest internal signal breaks down when the images are too dark. If you are using these robots in a dark room or at night, you can't rely on the robot to tell you it's failing. You need a separate "safety check" to make sure the picture isn't too dark before the robot even tries to look at it.

In short, small AI robots know when they are wrong deep inside their code, but they just won't say it out loud. They are like a student who gets a test question wrong but still raises their hand and says, "I'm totally sure of this!" The only way to catch them is to look at their secret scratch paper, not listen to what they say.

Technical Summary: Small Vision-Language Models Know When They Are Wrong But Cannot Say So

Problem Statement

As Vision-Language Models (VLMs) are increasingly deployed on consumer hardware and edge devices, they frequently encounter input images degraded by real-world conditions such as compression, motion blur, poor lighting, and glare. In these constrained environments, raw accuracy is insufficient for safe deployment; a system requires a reliable uncertainty signal to determine when to defer to a human rather than provide a potentially incorrect answer.

While generative VLMs offer two potential routes to uncertainty estimation—verbalized confidence (the model stating a confidence score in natural language) and internal confidence (the mean probability of the generated tokens)—the relationship between these signals under realistic image degradation has not been systematically examined for small, open-weight models. Prior work on proprietary models could not access internal token probabilities, leaving a gap in understanding whether small models possess usable self-knowledge that their text output fails to express.

Methodology

The study evaluates two small open-weight instruction-tuned VLMs: Qwen2-VL-2B-Instruct and SmolVLM-Instruct. Both models were run in fp16 on a single NVIDIA T4 GPU using greedy decoding.

Experimental Setup:

Dataset: A 100-item subset of the Food101 test set, formatted as four-option multiple-choice questions.
Degradations: Six realistic photographic degradation families were applied at three severity levels each:
1. JPEG Compression
2. Motion Blur
3. Low Light (brightness reduction + noise)
4. Glare (brightness amplification)
5. Rotation
6. Resampling (downscale/upscale)
Total Predictions: 3,800 predictions (1,900 per model across 19 conditions including the clean baseline).

Confidence Signals:

Verbalized Confidence: Models were prompted to output an integer confidence score $[0, 100]$ , normalized to $[0, 1]$ .
Internal Confidence: Calculated as the mean probability assigned to each token emitted during the generation of the answer span ( $c_{int} = \frac{1}{T} \sum p_\theta(y_t | y_{<t}, x)$ ).

Metrics:
The primary metric for evaluating uncertainty utility was Error-Detection AUROC (the probability that a randomly chosen correct answer receives a higher confidence score than a randomly chosen incorrect one). Secondary metrics included Accuracy, Expected Calibration Error (ECE), and Brier score.

Key Results

1. Accuracy Degradation

Both models demonstrated robustness to most degradations (compression, blur, glare, rotation, resampling), maintaining high accuracy. However, severe low-light conditions caused a catastrophic collapse in performance:

Qwen2-VL-2B: Accuracy dropped from 0.99 (clean) to 0.22 (severe low light).
SmolVLM: Accuracy dropped from 0.97 (clean) to 0.42 (severe low light).

2. Verbalized Confidence is Uninformative

Qwen2-VL-2B: Verbalized confidence remained nearly constant (mean 0.87–0.90) across all conditions, regardless of accuracy. Even when accuracy collapsed to 0.22, the model stated a confidence of 0.87. Consequently, verbalized confidence performed at or near chance level (AUROC $\approx$ 0.50) in error detection.
SmolVLM: Verbalized confidence was largely unobtainable. Across three prompt templates, only one of five pilot attempts produced a parseable confidence value; the model typically returned bare answers regardless of the instruction to state confidence.

3. Internal Confidence Detects Errors

In stark contrast to verbalized output, internal token probability provided a strong signal for error detection:

Qwen2-VL-2B: Internal AUROC ranged from 0.92 to 0.99 in clean and moderately degraded conditions, significantly outperforming verbalized confidence.
SmolVLM: Internal AUROC was above chance (0.54–0.92) in most conditions, though the signal was noisier and weaker than in Qwen2-VL.

4. Failure Under Severe Low Light

Both confidence signals failed precisely when they were needed most. Under severe underexposure (where accuracy collapsed), internal AUROC dropped to chance levels (0.62 for Qwen2-VL, 0.54 for SmolVLM). While internal confidence was slightly more responsive to accuracy drops than verbalized confidence, neither signal could reliably detect the catastrophic failure mode.

5. Calibration Metrics vs. Diagnostic Utility

Expected Calibration Error (ECE) was low for both signals in most conditions, creating a misleading impression of reliability. The study demonstrates that ECE is less diagnostic in this context because a near-constant confidence signal can appear well-calibrated if the model is generally accurate, even if the signal carries no information about specific errors. Error-detection AUROC is identified as the superior metric for assessing deferral utility.

Significance and Claims

The paper concludes that small VLMs encode usable self-knowledge that their verbalized output does not express.

For Deployment: Practitioners should not rely on stated confidence for deferral in small open-weight VLMs, as it is either at chance level or unobtainable. Instead, internal token probability should be used as the deferral signal, as it is free to compute, requires no retraining, and provides usable error detection under mild-to-moderate degradation.
Safety Limitation: Neither signal should be trusted under severe low-light conditions, as both fail to track the collapse in accuracy. Deployments facing such conditions require explicit upstream image-quality checks rather than downstream confidence thresholds.
Methodological Contribution: The study highlights that verbalized confidence and internal probability are distinct quantities that diverge sharply under degradation. It argues that future work on verbalized confidence in small VLMs should prioritize error-detection AUROC over calibration error measures.

The authors note that their findings are specific to the ~2B parameter class and that the gap between verbalized and internal confidence at larger scales remains an open question. Additionally, the study is limited to a single dataset (Food101) and synthetic degradations approximating real-world artifacts.