Compare With JoyAI and Mage

[Learning-ML-Steve]

Thanks for your fantastic work! As a new learner of this domain, I found there are also some good work in this domain, like JoyAI-VL-Interaction from JoyAI, and Mage-VL from Microsoft. I really hope the authors could make the comparison together of these paper.

[ShacklesLay]

Thanks! Both are good pointers.

JoyAI-VL-Interaction is already a baseline in Table 6 (OVO-Bench panel): 59.2 vs. MOSS-VL-Realtime's 70.2. It's average-only because JoyAI publishes one overall OVO-Bench number with no subset breakdown — it's in the GitHub README benchmark table (OVOBench 59.20), not the arXiv report.

Mage-VL is comparable through OVO-Bench, but per category rather than on the headline average. Its "Avg." is the mean of two categories (RTVP + BT); ours is the standard three-category mean including FAR. Mage-VL also evaluates under SimpleStream's recent-window protocol (the four most recent frames at 1 fps), while we run 1 fps over the accumulated stream up to 768 frames — which is why the shape differs.

OVO-Bench	MOSS-VL-Realtime	Mage-VL-4B
Real-Time Visual Perception	75.9	79.84
Backward Tracing	72.6	48.15
Forward Active Responding	62.1	not reported
Avg — RTVP + BT (their convention)	74.25 *	64.00
Avg — all three (OVO-Bench convention)	70.2	n/a

* recomputed on Mage-VL's convention; ours as reported is 70.2. Mage-VL is 4B; MOSS-VL-Realtime is 11.3B with an 8.2B language backbone.

On proactive behaviour there's no shared number yet. FAR isn't in the Mage-VL paper — it evaluates response timing on SoccerNet-Caption instead. That's the axis our wins concentrate on (FAR 62.1, OmniMMI Proactive Alerting 66.0, StreamingBench Proactive Output 60.0), so a shared FAR number would be the most useful thing to add.

We'll add a related-work note on both, plus a protocol/Avg footnote on OVO-Bench, in the next revision. Thanks for raising it!