Vision-language models enable training-free video anomaly detection by answering questions about video segments. VAD benchmarks, however, require a scalar anomaly score for each segment and evaluate the resulting ranking using the AUROC or AP. A VLM-based detector should therefore define an answer interface: the answer scale specifies the admissible answers, and the readout rule maps the model's output distribution to a score. Because this interface can change the evaluated ranking, it is part of the detector rather than a formatting detail. The generated readout uses only the most likely answer, whereas the probability readout uses the full distribution over admissible answers. Across four 7-8B VLMs, the probability readout outperforms the generated readout for every tested combination of answer scale, benchmark, and metric, with average gains ranging from 5 to 13 points across the four benchmark-metric pairs. The gap arises because the generated readout keeps only one answer value per segment, so segment with different answer distributions can receive the same score and lose their relative order. We call this loss of relative order generated-answer rank compression. Even when the answer scale allows 91 answers, the generated readout produces only 4-18 distinct scores, whereas the probability readout retains substantially finer score resolution. The advantage persists under every decoding strategy, prompt wording, and joint scoring-explanation prompt we test. The answer interface is therefore a consequential component of VLM-based VAD and should be explicitly specified and evaluated.
Frozen vision foundation models are commonly evaluated through a single global image embedding, but this interface can conflate missing information with information lost at readout time. We study this distinction by keeping a pretrained vision encoder frozen and varying only the readout applied to its final patch tokens. We compare standard global readouts against a lightweight foveated readout, which attention-pools patch tokens using a learned or question-conditioned query, and against an oracle readout with access to the annotated target region. We evaluate these interfaces on three localized binding problems: a controlled synthetic color--shape binding task under clutter, a color-free crowded shape-detection variant, and a GQA-derived natural-image task where paired questions ask for the colors of different same-category objects in the same image. Global readouts perform near perfectly when the synthetic target appears alone, but collapse under clutter and counterfactual target edits, whereas the foveated readout recovers most of the oracle-accessible signal. On the GQA-derived task, question-independent global image vectors improve only modestly over question-only priors, while question-conditioned foveation substantially improves paired localized color accuracy. A counterfactual nuisance-to-signal ratio explains the synthetic failures: global pooling dilutes localized label-changing evidence while exposing the probe to nuisance variation from irrelevant objects. These results indicate that apparent spatial blindness in frozen vision models can arise from the global embedding interface rather than from an absence of spatial information in the frozen patch tokens.
Learned dynamics models often answer global physical questions, such as fault severity or impact stiffness, by pooling a per-step feature sequence into one readout vector. This sequence-to-global interface creates an under-studied temporal credit problem: with only trajectory-level supervision, a model can predict accurately in training conditions while reading from abundant smooth correlates rather than the brief physical events that determine the target. We call this failure temporal credit dilution. It is not exposed by the training loss and is not removed by standard physics-informed residuals, because the error lies in where the global readout assigns functional credit. We introduce Credit-in-Event, an interface-level probe for measuring how much pooled credit lands on event steps, and prove in closed form that a pooled linear reader routes credit to a spurious background channel as the event fraction shrinks. We then propose CREST, a training-free and label-free readout that estimates a transient event core from learned features and re-anchors the pooled representation through event-versus-rest contrast. Across simulated gear and impact systems, recurrent and attention encoders, and public bearing vibration data, CREST reduces out-of-distribution error while restoring event credit. Ablations show that stable-step selection and receptive-field shrinking fail, confirming that the gain comes from event-core credit re-anchoring rather than a generic locality or stability prior.