The Mirage of Calibrated Confidence: Trajectory-Independence of Verbalized Confidence in Vision-Language Models
Jisoo Yang, Jaeho Han, Trung X. Pham, Junyeong Kim
cs.AI
Quick verdict
Claims VLM verbalized confidence ignores its reasoning trajectory and calibration can worsen this; convincing diagnostics, though mainly one model family.
Summary
The paper asks whether verbalized confidence in VLMs is causally informed by the reasoning trajectory that produced the answer. Through three probes on Qwen3-VL-4B/8B and VL-Calibration-4B/8B - Jensen-Shannon divergence under trajectory-content substitution, attention-level token masking at five nested ratios, and pre/post-hesitation confidence around self-expressed doubt markers - the authors find confidence changes little in all cases and paradoxically rises at 100% masking. They formalize the property with the Trajectory-Grounding Score: TGS-self (confidence with vs. without the self-trajectory via full attention masking) and TGS-pair (confidence for good vs. bad trajectories along vision, reasoning, and answer axes), supported by TGS-Bench, 18,295 model-agnostic trajectory pairs over 10 benchmarks built from gold-conditioned Qwen2.5-VL-72B generations with single-error edits. Headline result: VL-Calibration cuts ECE (.402 -> .081 at 8B) yet yields negative average TGS-self (-.018) and reduced TGS-pair (.084 -> .065), with the vision axis weakest across all methods. Two-stage bootstrap CIs and Wilcoxon tests over 10 benchmark means support the main effects. Main limitations: evidence rests chiefly on the Qwen3-VL family (the InternVL3.5-8B cross-check shares its LM backbone), heavy filtering to small common-valid samples, and a diagnostic-only scope with no remediation proposed.
Review complete
Assessment at a glance
6
risk flags identified
● high 0
● medium 3
● low 3
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Verbalized confidence in the evaluated VLMs is largely trajectory-independent: it changes little when trajectory components are substituted (vision P(JSD<=0.1) up to 40-90%), attention-masked (confidence rises at 100% masking), or internally contradicted by the model's own hesitation markers (mean pre-post dip near zero, Figs. 13-14).
- Outcome-level calibration training does not imply trajectory-grounding: VL-Calibration cuts ECE (.402 -> .081 at 8B) while producing negative average TGS-self (-.018) and lower TGS-pair (.084 -> .065), i.e., calibration can induce anti-grounding.
- ECE and AUROC are structurally blind to this disconnect because they measure population-level confidence-accuracy agreement, which a model can achieve by learning per-difficulty base rates without ever inspecting its trajectory.
- Vision-axis grounding is consistently the weakest across all evaluated methods and scales (Fig. 7), identifying visual-perception rationale quality as the bottleneck.
Strengths
Triangulated causal and distributional design
Sec. 3.2-3.4 converge from three angles: content substitution (pairwise JSD, P(JSD<=0.1)), position-preserving attention masking at nested ratios with 3 seeds for partial conditions (A.2), and intra-trajectory hesitation analysis (Table 1: 22.5%/57.2% of trajectories contain markers). The masking-upturn-at-100% finding (Fig. 3) is a genuinely counterintuitive causal complement to the distributional evidence.
Signed metrics expose a failure mode invisible to ECE/AUROC
TGS's sign convention detects anti-grounding: Observation 3 reports negative TGS-self on 8/10 benchmarks at 4B and 9/10 at 8B after VL-Calibration, i.e., masking the model's own rationale raises its confidence - something ECE=.081 would certify as healthy.
Unusually rigorous benchmark validation
Honest, self-correcting reporting
Clear positioning and formalization
Weaknesses
Narrow architectural base for a field-general claim
Heavy sample attrition with unanalyzed exclusions
Uncontrolled fluency/style confound in pair construction
Answer-axis design may measure format insensitivity
Low statistical power and no multiplicity control
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What does 'trajectory-independent confidence' in VLMs mean?
It means the confidence a vision-language model verbally reports does not actually depend on the reasoning it just produced. The paper shows this by replacing parts of the reasoning (visual description, logic chain, or final answer) with different content, by hiding the reasoning from the model via attention masking, and by checking whether the model's own 'wait, let me recheck' moments lower its confidence. In all three tests, confidence barely moved, and hiding the entire reasoning sometimes raised it.
What are TGS-self and TGS-pair in this paper?
They are two proposed metrics for 'trajectory grounding' - whether confidence reflects the reasoning process. TGS-self compares the model's confidence when it can see its own generated reasoning versus when that reasoning is fully attention-masked; a positive score means the model reads its own rationale, near zero means trajectory-blindness, negative means anti-grounding. TGS-pair uses controlled good/bad trajectory pairs that differ in exactly one axis (vision, reasoning, or answer) and checks whether the model gives higher confidence to the good trajectory on each axis.
Why doesn't a low ECE (calibration error) mean a model's confidence is trustworthy?
ECE only checks whether stated confidence matches accuracy across a population - a model can be perfectly calibrated by learning how often it is right on each difficulty level, without ever looking at its own reasoning for a specific question. This paper shows exactly that failure: a model with ECE of .081 (seemingly excellent) still ignores its own chain of reasoning when producing confidence, so aggregate calibration scores miss whether individual confidence judgments are causally connected to the actual thought process.
Which models and benchmarks were used in this study?
The main experiments use Qwen3-VL-4B and 8B (base) and their calibration-trained derivative VL-Calibration, with additional checks on ConfTuner, RLCR, Verbalize, P(True), and SteerConf, plus InternVL3.5-8B (which shares Qwen3's language backbone). Evaluation spans 10 benchmarks - DynaMath, Geo3K, MathVerse, MathVision, MathVista, WeMath, LogicVista, A-OKVQA, MMK-12, and MMMU-Pro - with diagnostic subsets of MM-Vet and GQA; the proposed TGS-Bench contains 18,295 controlled trajectory pairs.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- All figures (Figs. 2-16) are unreadable from the extracted text; figure-dependent claims (masking curve shapes, per-axis JSD densities, scatter layouts) are judged from captions and prose restatements only.
- Table 2's column layout is garbled by extraction (Base/VL-Cal columns interleaved), so per-benchmark cells could not be verified; headline averages quoted in prose (.402->.081 ECE, -.018 TGS-self, .084->.065 TGS-pair) are internally consistent across Sec. 5.2 and Table 5.
- Some appendix content may be truncated (e.g., full MMK-12 provenance, rejection logs of TGS-Bench construction), which limits assessment of construction quality.
- Code and TGS-Bench are promised only upon acceptance, so reproducibility claims rest on the written procedures in Appendices A-C.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.