Pith. sign in

REVIEW 4 major objections 5 minor 15 references

Two small vision-language models carry an internal signal of their own correctness that their verbalized confidence does not express.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 05:59 UTC pith:EVAF4SNQ

load-bearing objection Novel comparison with a real finding on internal confidence, but the 'cannot say so' claim rests on a single prompt template; needs an elicitation ablation before it can be stated as general. the 4 major comments →

arxiv 2607.22034 v1 pith:EVAF4SNQ submitted 2026-07-24 cs.CV cs.CLcs.LG

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

M M Asif Ferdous This is my paper
classification cs.CV cs.CLcs.LG
keywords vision-language modelsverbalized confidencetoken probabilitycalibrationerror detectionimage degradationdeferral signaluncertainty estimation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tests whether two small vision-language models can signal when they are wrong under realistic image degradation, comparing the confidence they state in words with the probability they internally assign to their own generated tokens. Across 3,800 multiple-choice predictions under six degradations, stated confidence in one model stays near constant (0.87–0.90) and detects errors at chance level, while internal token probability from the same predictions separates correct from incorrect answers with AUROC up to 0.99. The second model largely refuses to state any confidence at all, yet its internal signal still beats chance. Both models fail identically under severe underexposure: accuracy collapses while both confidence signals barely move. The paper's conclusion is that small VLMs encode usable self-knowledge their verbal output does not express, making internal probability the better deferral signal—except precisely in the severe condition where it would matter most.

Core claim

The central claim is that verbalized confidence and internal token probability are different quantities, and under realistic degradation they diverge sharply in small vision-language models. In Qwen2-VL-2B, verbalized confidence is effectively constant across all 19 conditions and error-detection AUROC hovers at chance (0.39–0.75, typically about 0.50), while mean token probability over the generated answer detects errors with AUROC 0.92–0.99. In SmolVLM, verbalized confidence was largely unobtainable across three prompt templates, yet internal probability again beat chance (AUROC 0.54–0.92). The paper frames this as evidence that small VLMs encode usable self-knowledge that their natural-la

What carries the argument

The key comparison is between two confidence signals computed on the same predictions: verbalized confidence, elicited by prompting the model for an integer 0–100 and normalizing it, and internal confidence, the mean probability the model assigned to each token it actually emitted, averaged over the answer. Error-detection AUROC is the operative metric, measuring whether a signal ranks correct answers above incorrect ones rather than merely matching accuracy levels. The divergence between these two signals on identical predictions is the mechanism that carries the argument.

Load-bearing premise

The load-bearing premise is that the single prompt template chosen for verbalized confidence fairly represents what these models can state; the paper concedes that the near-constant 0.90 in Qwen2-VL may be partly template-induced and that an elicitation ablation is required before the claim holds in full generality.

What would settle it

Run the same Qwen2-VL-2B predictions with a broad search over prompt templates for verbalized confidence and measure error-detection AUROC; if any template yields clearly above-chance discrimination (say, AUROC greater than 0.8) on the same data, the central claim that the model cannot say when it is wrong would be falsified.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Practitioners should prefer internal token probability over verbalized confidence as a deferral signal for small open-weight VLMs under mild-to-moderate degradation.
  • Post-hoc temperature scaling of verbalized confidence improves calibration under mild conditions but worsens it at severe low light, because rescaling a near-constant signal cannot recover information the signal never contained.
  • Deployments that may encounter severe underexposure need an upstream image-quality check rather than a downstream confidence threshold, since both signals go silent at exactly the point where accuracy collapses.
  • Expected calibration error can look well-calibrated while the confidence signal carries no usable discrimination; error-detection AUROC is the more diagnostic metric in this regime.
  • The verbalized/internal gap is only measurable because the models are open-weight; the paper's comparison recovers the token-level signal that proprietary settings exclude by construction.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If this divergence generalizes beyond the two models tested, then standard 'ask the model how confident you are' interfaces for small VLMs may be systematically misleading; uncertainty should be surfaced from the internal distribution instead, a design implication the paper leaves mostly implicit.
  • A testable extension suggested by the paper's own limitation is a broad elicitation ablation: if some prompt or fine-tuning strategy yields verbalized confidence with error-detection AUROC clearly above chance, the 'cannot say so' conclusion would be weakened to a prompt-selection artifact.
  • The shared failure under severe underexposure points toward a hybrid deferral rule that thresholds image-quality statistics (such as brightness) upstream and internal probability downstream; the paper gestures at this in its recommendation but does not test it.
  • Whether the verbalized/internal gap narrows with model scale is left untested; the paper itself calls this the most interesting open question it raises.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The manuscript reports an empirical comparison of two confidence signals in two small open-weight VLMs (Qwen2-VL-2B-Instruct and SmolVLM-Instruct) under six realistic image degradations at three severity levels, plus a clean baseline, on a 100-item Food101 subset posed as four-way multiple choice. Verbalized confidence is obtained by prompting the model to state an integer confidence; internal confidence is the mean token probability over the generated answer. The paper reports that Qwen2-VL verbalized confidence is nearly constant (0.87–0.90) and mostly at-chance for error detection (AUROC 0.39–0.75), while internal token probability separates correct from incorrect answers with AUROC 0.92–0.99. For SmolVLM, verbalized confidence was largely unobtainable across three prompt templates, while internal AUROC ranged 0.54–0.96. Under severe low light, accuracy collapses in both models while both confidence signals remain high, and internal error-detection intervals include chance. The paper concludes that internal mean token probability is the better deferral signal for small VLMs and recommends an upstream image-quality check under low-light conditions.

Significance. If the findings hold, the paper makes a useful empirical and methodological contribution: it directly compares verbalized and internal confidence on the same predictions in an open-weight setting, which is impossible for proprietary models, and it demonstrates that ECE can appear low while a confidence signal carries no discriminative information. The authors are unusually transparent about limitations, and the code and per-prediction data appear to be provided. However, the headline claim that small VLMs 'cannot say so' is only as strong as the prompt elicitation, and the evidence base is narrow (n≈100 per condition, one dataset, two ~2B models, and only one verbalized-confidence template per model). With an elicitation ablation and more robust error counts, the paper would be a solid empirical contribution; without them, the central claim is overgeneralized.

major comments (4)
  1. [§5.4, item 4; Appendix B] The central claim that verbalized confidence is uninformative rests on a single template per model, selected for parse rate. The Qwen2-VL template contains an explicit worked example 'Confidence: 85', which can anchor responses to a stock high value. The paper itself concedes (Section 5.4, item 4) that the near-constant 0.90 'may be partly template-induced' and that 'an elicitation ablation is required before the claim can be stated in full generality.' Because the title and abstract assert that the models 'cannot say so', this is a load-bearing unaddressed alternative. Either perform an elicitation ablation (varied phrasing, no worked example, different exemplars) or restrict the conclusion to 'with the tested template'.
  2. [§5.3; Appendix B (SmolVLM templates)] The statement that SmolVLM verbalized confidence is 'largely unobtainable' is based on three templates on a five-item pilot, with only one parseable response across fifteen attempts. This is insufficient to establish a general inability; it may reflect prompt mismatch. Moreover, the template was selected by parse rate, not informativeness, so a template that would have produced variable, error-discriminating confidence but lower parse rate would have been discarded. Please report per-condition parse rates, conduct a broader prompt search, or explicitly change the conclusion to 'no reliable verbalized confidence was obtained with the tested prompts' rather than 'unobtainable'.
  3. [§4.5; Table 3] The claim that 'both signals fail under severe low light' conflates calibration level with discrimination. For Qwen2-VL at low light s3, internal AUROC is 0.62 [0.48, 0.75] (not distinguishable from chance) while verbalized AUROC is 0.68 [0.54, 0.80], which is above chance. For SmolVLM internal AUROC is 0.54 [0.42, 0.66]. 'Failure' means different things for mean confidence (stays high while accuracy collapses) and for AUROC (near chance discrimination). A pre-specified criterion for 'failure' (e.g., CI includes 0.5 and mean shift < 0.05) should be stated, and both aspects reported separately.
  4. [Table 2 / §4.3; §5.4, item 2] The headline internal AUROC values of 0.92–0.99 in Qwen2-VL are often computed from very few incorrect answers. For clean (accuracy 0.99), jpeg s1/s2 (0.99), and rotation s3 (0.99), the error count is one to three, making bootstrap CIs and AUROC estimates fragile. Please report the number of correct and incorrect answers in every cell, and consider pooling over severity or using a rank-based paired test. Without this, the near-ceiling values are not yet robust evidence for the claimed gap.
minor comments (5)
  1. [§4.2; Table 4] The text reports mean verbalized confidence ranges (0.87–0.90) but Table 4 does not include a column for mean verbalized confidence. Adding such a column would let readers verify the constancy claim directly.
  2. [§3.6] The definition of AUROC does not specify how ties are handled. Since verbalized confidence is nearly constant, ties are frequent and the choice of tie-breaking (random, or half-count) can materially affect the reported values. Please state the convention.
  3. [Appendix B; §3.4] The prompt selection criterion should be stated more precisely: is 'parse rate' the only criterion, or was informativeness considered? The sentence 'the one with the highest parse rate retained' suggests parse rate alone, but the reader needs to know whether the chosen template also produced varied values on the pilot items.
  4. [References] The SmolVLM citation uses 'Hugging Face TB' as the author. The reference should include the full organizational author or a stable URL with a proper author field.
  5. [§5.2] The recommendation 'do not use stated confidence for deferral' is stated categorically. Given the acknowledged prompt sensitivity, this should be tempered to 'do not use stated confidence with the tested prompts' unless the requested ablation supports a stronger claim.

Circularity Check

0 steps flagged

No circularity: the paper's central comparison is a direct measurement with disclosed procedures; the prompt-sensitivity limitation is a validity caveat, not a circular derivation.

full rationale

The paper contains no derivation chain that reduces its conclusions to its inputs. The central claim—verbalized confidence is near-constant and near-chance for error detection while mean token probability separates correct from incorrect answers—is a direct empirical measurement on the same predictions. The internal confidence signal is defined before the accuracy comparison by equation (1), c_int = (1/T) Σ p_θ(y_t | y_<t, x), and is not fit to maximize AUROC. The verbalized-confidence signal is elicited with a prompt template, and the paper explicitly discloses that the template was selected by parse rate (Section 3.4), not by informativeness. The only fitted number, temperature T=0.50, is fitted on a held-out clean split and applied unchanged to degraded conditions (Section 4.7); this is out-of-sample evaluation, not circular fitting. The paper's own limitation section (Section 5.4, item 4) concedes that 'The near-constant 0.90 in Qwen2-VL may be partly template-induced. An elicitation ablation is required before the claim can be stated in full generality.' This is a limitation on external validity and generality, not a circularity: the claim is not defined in terms of the template, nor is the template derived from the claim. There are no self-citations, no imported uniqueness theorems, and no renamed known results masquerading as new derivations. The work is self-contained against an external benchmark and reports its measurements transparently, so the honest finding is no significant circularity.

Axiom & Free-Parameter Ledger

1 free parameters · 5 axioms · 0 invented entities

The central comparison is empirical and does not fit parameters to the headline result. The only fitted number is the temperature T=0.50 in the secondary scaling experiment, fit on a held-out clean split and disclosed. The main conclusions rest on domain assumptions: one dataset, two models, one chosen prompt template, and the treatment of mean token probability as internal confidence. No new entities or mechanisms are postulated.

free parameters (1)
  • Temperature T = 0.50
    Fitted on a held-out clean split for post-hoc temperature scaling of Qwen2-VL verbalized confidence (Section 4.7); applied to all degraded conditions. Disclosed, and not used in the central AUROC comparison.
axioms (5)
  • domain assumption Food101 4-way multiple-choice is representative of small-VLM deployed tasks
    Sole dataset used for all conclusions; generality to open-ended VQA and other domains is untested (Section 5.4, item 3).
  • domain assumption Mean token probability over the generated answer span is a valid internal-confidence signal
    Defined in Equation 1; no independent validation that this quantity corresponds to model self-knowledge; it is the core internal signal used for RQ3.
  • domain assumption The selected verbalized-confidence prompt is representative of natural use
    Prompt chosen from three candidates on a five-item pilot; paper explicitly flags that the near-constant 0.90 may be partly template-induced (Section 5.4, item 4).
  • domain assumption Qwen2-VL-2B and SmolVLM are representative of small open-weight VLMs
    Both are ≈2B parameters; whether the verbalized/internal gap narrows with scale is untested (Section 5.4, item 6).
  • domain assumption Greedy decoding yields stable and representative predictions
    Greedy decoding used throughout for reproducibility; no sampling-based diversity is explored (Section 3.1).

pith-pipeline@v1.3.0-alltime-deepseek · 9880 in / 15397 out tokens · 156791 ms · 2026-08-01T05:59:30.599052+00:00 · methodology

0 comments
read the original 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.

Figures

Figures reproduced from arXiv: 2607.22034 by M M Asif Ferdous.

Figure 1
Figure 1. Figure 1: Accuracy as a function of corruption severity for both models. Five of six degradation [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Error-detection AUROC for verbalized (red) versus internal (blue) confidence in Qwen2- [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Qwen2-VL-2B under increasing low light. Accuracy (black) falls from 0.99 to 0.22 [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Reliability diagrams for both confidence signals on clean images (Qwen2-VL-2B). Both [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

15 extracted references · 8 linked inside Pith

  1. [1]

    arXiv preprint arXiv:2504.03440 , year =

    Know What You Do Not Know: Verbalized Uncertainty Estimation Robustness on Corrupted Images in Vision-Language Models , author =. arXiv preprint arXiv:2504.03440 , year =

    Pith/arXiv arXiv

  2. [2]

    The Twelfth International Conference on Learning Representations (ICLR) , year =

    Can LLMs Express Their Uncertainty? An Empirical Evaluation of Confidence Elicitation in LLMs , author =. The Twelfth International Conference on Learning Representations (ICLR) , year =

  3. [3]

    arXiv preprint arXiv:2305.14975 , year =

    Just Ask for Calibration: Strategies for Eliciting Calibrated Confidence Scores from Language Models Fine-Tuned with Human Feedback , author =. arXiv preprint arXiv:2305.14975 , year =

    Pith/arXiv arXiv

  4. [4]

    arXiv preprint arXiv:2405.02917 , year =

    Overconfidence is Key: Verbalized Uncertainty Evaluation in Large Language and Vision-Language Models , author =. arXiv preprint arXiv:2405.02917 , year =

    Pith/arXiv arXiv

  5. [5]

    arXiv preprint arXiv:1907.07484 , year =

    Benchmarking Robustness in Object Detection: Autonomous Driving When Winter is Coming , author =. arXiv preprint arXiv:1907.07484 , year =

    Pith/arXiv arXiv 1907

  6. [6]

    arXiv preprint arXiv:2302.09664 , year =

    Semantic Uncertainty: Linguistic Invariances for Uncertainty Estimation in Natural Language Generation , author =. arXiv preprint arXiv:2302.09664 , year =

    Pith/arXiv arXiv

  7. [7]

    International Conference on Machine Learning (ICML) , pages =

    On Calibration of Modern Neural Networks , author =. International Conference on Machine Learning (ICML) , pages =. 2017 , publisher =

    2017

  8. [8]

    Advances in Neural Information Processing Systems , volume =

    Can You Trust Your Model's Uncertainty? Evaluating Predictive Uncertainty Under Dataset Shift , author =. Advances in Neural Information Processing Systems , volume =

  9. [9]

    arXiv preprint arXiv:1903.12261 , year =

    Benchmarking Neural Network Robustness to Common Corruptions and Perturbations , author =. arXiv preprint arXiv:1903.12261 , year =

    Pith/arXiv arXiv 1903

  10. [10]

    arXiv preprint arXiv:2207.05221 , year =

    Language Models (Mostly) Know What They Know , author =. arXiv preprint arXiv:2207.05221 , year =

    Pith/arXiv arXiv

  11. [11]

    arXiv preprint arXiv:2409.12191 , year =

    Qwen2-VL: Enhancing Vision-Language Model's Perception of the World at Any Resolution , author =. arXiv preprint arXiv:2409.12191 , year =

    Pith/arXiv arXiv

  12. [12]

    2024 , howpublished =

    SmolVLM , author =. 2024 , howpublished =

    2024

  13. [13]

    Computer Vision -- ECCV 2014 , series =

    Food-101 -- Mining Discriminative Components with Random Forests , author =. Computer Vision -- ECCV 2014 , series =

    2014

  14. [14]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume =

    Obtaining Well Calibrated Probabilities Using Bayesian Binning , author =. Proceedings of the AAAI Conference on Artificial Intelligence , volume =

  15. [15]

    Monthly Weather Review , volume =

    Verification of Forecasts Expressed in Terms of Probability , author =. Monthly Weather Review , volume =