acceptodds
Under review as a conference paper at ICLR 2027

Fisher–Rao Internal Decision Geometry for VLM Misbehavior Monitoring

Abstract

Detecting VLM misbehavior requires identifying incorrect or unsafe answers, including those generated under adversarial inputs and distribution shifts. Output confidence can remain high for such answers and does not fully describe the prediction process. Predictions with similar confidence may respond differently to small changes in decoder states, while their final distributions can conceal revisions across layers. We investigate whether these differences provide additional evidence of failure. Fisher–Rao geometry connects state variations to changes in predictive distributions and provides a shortest-path distance for comparing predictions across layers. Based on these properties, we propose Fisher–Rao Decision Geometry (FRDG), a training-free, white-box monitor. FRDG measures the strongest local predictive change under a perturbation budget relative to the state norm. It also measures accumulated changes across late-layer predictions beyond the direct distance between the first and last predictions. These measurements capture local sensitivity and intermediate revision, respectively, and their standardized summaries are combined into a risk score. We derive the local displacement interpretation and bounds on the additional layerwise change. Across four VLMs and nine benchmark subsets, the reported results show relative improvements of up to 12.42% in macro AUROC and 15.41% in macro AP over the strongest compared baseline for each model and metric. Component ablations show that the combined score improves failure ranking over either measurement alone.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Vision-Language Models, Misbehavior Monitoring

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_X4nLOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Fisher–Rao Decision Geometry (FRDG), a training-free, white-box monitor that assigns risk scores to completed VLM responses. It combines relative-state local predictive sensitivity, summarized by the leading eigenvalue of a state Fisher operator, with late-layer trajectory excess, defined as accumulated predictive distance minus the endpoint distance. Components are aggregated over answer positions and standardized using nominal reference responses. The paper derives a local perturbation interpretation and trajectory properties, presents a marginal nominal conformal alarm guarantee, and evaluates failure ranking on four VLMs and nine benchmark subsets. Reported macro AUROC/AP improve over the compared monitors, and ablations support combining the two measurements.

Critical strengths
  1. The local score has a precise interpretation under a specified relative state-perturbation budget. This is a coherent measurement distinct from evidential conflict or ignorance. Combining it with trajectory excess gives an interpretable monitor of completed answers without fitting a detector or generating additional answers.
  2. The reported gains merit attention: macro AUROC improves by 4.67–9.30 percentage points over the strongest baseline in the main macro comparison (Appendix D.1.2) across four VLMs, with trajectory excess adding 0.97–1.39 points over local sensitivity alone. The paper supplies extensive component and substitution controls, distinguishes nominal alarm guarantees from failure detection, and discloses variants outperforming the proposed construction. For the layer/position ablations, Appendix D.4.1 states that the main configuration was retained rather than replaced based on the test comparisons.
Critical weaknesses
  1. The methodological increment beyond existing internal-uncertainty monitoring remains limited. The local perturbation interpretation differs from EUQ, but the experiments support a useful family of sensitivity-plus-trajectory scores more strongly than the particular Fisher–Rao choices. In Table A40, replacing only the trajectory distance with L2 or JS matches or slightly improves macro AUROC on all four models; these controls retain the local Fisher component. The leading eigenvalue has no consistent AUROC advantage over the trace, although it achieves higher AP, so these alternatives are not equivalent across metrics. Several simpler aggregation variants also outperform the proposed combination. These observations do not invalidate the method, but weaken the case for its specific design. The paper does not yet clearly establish which failures its features distinguish beyond EUQ-CF, EUQ-IG, and ordinary confidence; the exploratory stratifications in Appendix C.5 do not provide a joint or tightly controlled comparison.

  2. The evaluation history and configuration behind parts of the central evidence remain unresolved. Appendix B.2 gives incompatible InternVL3 records: a 28-block decoder with expected indices 17/21/25/28 versus exported indices 19/24/29/32. The actual configuration producing that model’s scores is therefore unclear. Appendix D.2.2 also states that the reported intervals do not account for outcome-based selection recorded in the companion artifact and must not be interpreted as independent validation of a selected configuration. This needs reconciliation with the statements in Section 3.4/Appendix B.6 that test responses and behavior labels do not determine score parameters. Retaining the main configuration despite better reported variants is reassuring, but does not identify the scope or data used for the acknowledged selection. These are questions about experimental traceability and evaluation independence, not evidence of fabrication or proof that all results are invalid. Until resolved, they limit how strongly I can rely on the promising empirical improvements.

Questions for authors
  1. Which configurations were selected using outcomes, on which data partitions, and at what stage? Please reconcile Appendix D.2.2 with Section 3.4/Appendix B.6 and identify which main results and ablations were evaluated on data untouched by that selection.
  2. What exact InternVL3 checkpoint, decoder depth, and block indices generated the reported scores, and can the affected results be confirmed after reconciling the records? Separately, please clarify whether Table 3’s spectral checks use the actual full-dimensional, full-vocabulary model operators and how these relate to the metadata for local checks in Appendix D.3.2.
  3. The manuscript includes a custom CF+IG control . Beyond this aggregate comparison, can the recorded scores establish additional failure information conditional on EUQ-CF, EUQ-IG, and output confidence? A joint or matched analysis with an evaluation protocol fixed independently would help assess whether the contribution extends beyond improved aggregate ranking. This is a request for evidence supporting the current contribution, rather than a suggestion for a new application.
Reviewer_ZgcmOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes a training-free Fisher–Rao Decision Geometry (FRDG), which is a white-box monitor for identifying incorrect or unsafe VLM responses. The method combines two signals: local predictive sensitivity, measured by the leading eigenvalue of the scaled state Fisher operator, and trajectory excess. These signals are combined into a final response risk score. Experiments on 4 VLMs and 9 benchmark subsets demonstrate improvements in misbehavior detection.

Critical strengths

S1. The paper provides mathematical explanations based on Fisher–Rao geometry and clearly distinguishes the resulting geometric properties from answer correctness.

S2. The experiments cover multiple VLMs and different types of misbehavior, with relatively detailed experimental analyses.

Critical weaknesses

W1. The supplementary material is excessively long (up to 73 pages including main paper), and the large volume of experimental results makes it difficult to assess the paper’s core claims. The experiments in the appendix and the writing of appendix itself looks like an output of the auto research. The experiments should be presented more concisely and organized around the specific claims they support.

W2. In the detailed ablation study in Table A40, L2 distance excess and JS distance excess achieve results comparable to or slightly better than FRDG. Moreover, the leading eigenvalue does not consistently outperform the trace. These results do not sufficiently justify the necessity of several key design choices.

Questions for authors

Please address the concerns in the weaknesses.

Reviewer_CHwqOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

FRDG is a training-free, white-box monitor for incorrect or unsafe VLM answers. It combines two signals: local Fisher sensitivity at the final decoder state, under a perturbation budget relative to the state norm, and the excess Fisher–Rao path length across late-layer predictions. Across 4 VLMs and 9 benchmarks, it reports the best macro AUROC (84–87%) and AP among the compared monitors.

Critical strengths
  1. The method is principled and cheap, with clean theory and an efficient matrix-free Lanczos estimator.

  2. The paper reports honestly where FRDG loses and where alternative score variants do better.

Critical weaknesses
  1. The main signal’s interpretation does not match the motivation. Under RMSNorm, the scaling factor st2=∥ht∥2s_t^2 = \|h_t\|^2 cancels the normalization Jacobian, so that At≈d P⊥DγW⊤CtWDγP⊥A_t \approx d\, P_\perp D_\gamma W^\top C_t W D_\gamma P_\perp The score then depends only on the output distribution, the unembedding matrix, and the direction of the hidden state. This makes it closer to a vocabulary-weighted dispersion of the output distribution than to “state sensitivity.” The paper never compares the score against entropy or confidence at matched levels.

  2. The authors disclose provenance problems themselves. The appendix states that experimental provenance is not independently established and refers to “outcome-based selection.” The InternVL3 layer indices (29, 32) are impossible for a 28-layer decoder, and the latency numbers are analytic rather than measured. The conformal threshold also fails in practice: TPR is about 0% on Qwen and LLaVA at α=0.05\alpha = 0.05 because of a mismatch between the reference and test distributions.

Questions for authors
  1. Were any configuration choices (layers, number of positions, combination rule) selected using test outcomes? What exactly does “outcome-based selection” in the companion artifact refer to? What is the correct InternVL3 layer mapping, and do the results change once it is fixed?

  2. Can you report measured latency and memory on common hardware, and the sampling parameters actually used for SC, SE, and EUQ?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.