Determine whether targeted fine-tuning resolves the temporal capability gap

Conduct a targeted fine-tuning probe on Vision-Language Models using physical-AI and densely timestamped video supervision to determine whether weak temporal localization and dense video captioning reflect a capability gap or an elicitation gap.

Background

The paper compares model families with progressively greater amounts of physical-AI training data and finds highly uneven transfer. Tracking, object localization, and event verification improve substantially, whereas Temporal Localization and Dense Video Captioning improve only slightly.

The authors hypothesize that bounding-box supervision is abundant while densely timestamped video supervision is scarce. They explicitly leave unresolved whether the poor temporal results arise from a fundamental capability limitation or from an elicitation problem that could be corrected through targeted fine-tuning.

References

A targeted fine-tuning probe would distinguish a capability gap from an elicitation gap, and we leave this test to future work.

— VANTAGE-Bench: Evaluating the Infrastructure AI Gap in Vision-Language Models  (2609.09396 - Bhat et al., 8 Sep 2026) in Section 4.3, “Returns to Scale and the Temporal Lag”

Future work will evaluate these dimensions and examine whether more accurate motion captions provide better supervision for downstream learning.

— Beyond Binary Preferences: Graded Preference Optimization for Limb-Motion Captioning  (2609.36628 - Wang et al., 29 Sep 2026) in Section 6, Conclusion; reiterated in Section 11, "Limitations and Future Work" (Scope of Dense Caption Evaluation)

Whether denser temporal sampling or longer sequence horizons can yield further improvements remains an open question.

— Mid-Training Language Models on Raw Video  (2610.11019 - Hwang et al., 8 Oct 2026) in Section Discussion