troupe-standup/wizard-work/sphinx/REPORT-mm-vla-dossier-2026-10-22.md

Multimodal archive dossier — live arXiv fetch (2026-10-22T17:13:24Z)

Lane: Archive (multimodal, long-context research). Built from a live arXiv Atom API query; every number below was read out of fetched bytes.

Query

• search_query: cat:cs.CV AND abs:"multimodal"
• Atom URL: https://export.arxiv.org/api/query?search_query=cat:cs.CV+AND+abs:%22multimodal%22&start=0&max_results=120&sortBy=submittedDate&sortOrder=descending
• HTTP 200, 311552 bytes, sha256 17767bfb4ec4c02fdc21e79d7bf7f96035610ee37356a1edc2a3319a7e4c95be
• opensearch:totalResults = 15740
• entries parsed from feed: 120
• papers with an HTML rendering fetched: 3

Papers

arXiv id	primary cat	published	tables	figures	images	title
2610.30108v1	cs.RO	2026-10-21	4	12	10	Multimodal Embodied Agents: A Unified Capability-centric Perspective from Computer-Use to Robot-Use
2610.30089v1	cs.CV	2026-10-21	12	30	12	ProactEgoMem: Benchmarking Proactive Memory in Streaming Egocentric Video
2610.29913v1	cs.CV	2026-10-21	6	12	1	Asking a Multimodal LLM Whether Someone Fell: Zero-Shot Screening of Low-Resolution Thermal Video in a Laborat

Coverage probe (HTML-rendering availability)

Before archiving, every entry in the 120-item Atom feed was probed for an HTML rendering at arxiv.org/html/<id>. Measured, not assumed:

• papers probed: 120
• with an HTML rendering: 118 (98.3%)
• median tables per paper: 18 (min 1, max 69)
• median figures per paper: 17 (min 4, max 359)
• median img tags per paper: 10 (min 2, max 343)

Figure/table inventory

2610.30108v1 — Multimodal Embodied Agents: A Unified Capability-centric Perspective from Computer-Use to Robot-Use

• text layer: 196723 chars; tables: 4 (numeric: 4); figures: 12; images: 10; svg figures: 1; figures with MathML: 2
  • table 0: 12 rows — (no caption observed)
  • table 1: 1 rows — (no caption observed)
  • table 2: 1 rows — (no caption observed)
  • table 3: 78 rows — (no caption observed)
  • figure 0: 1 img, 0 svg, math=False — Figure 1: A Unified Capability-centric Perspective (PAPAV) from computer-use to robot-use.
  • figure 1: 1 img, 1 svg, math=False — Figure 2: Chronological overview of representative work. Robotic systems progress from vis
  • figure 2: 0 img, 0 svg, math=True — Table 1: Positioning related surveys. Checks mark dimensions central to a survey; crosses
  • figure 3: 0 img, 0 svg, math=False — Figure 3: Research landscape and taxonomy of PAPAV. The capability branch groups represent
  • figure 4: 1 img, 0 svg, math=False — Figure 4: PAPAV from computer-use to robot-use. Each row contrasts how a shared capability
  • figure 5: 1 img, 0 svg, math=False — Figure 5: Perceive: MMEA vs. MMA vs. RS. Shared (center): multimodal evidence grounding. L
  • figure 6: 1 img, 0 svg, math=False — Figure 6: Anticipate: MMEA vs. MMA vs. RS. Shared (center): action-conditioned state-trans
  • figure 7: 1 img, 0 svg, math=False — Figure 7: Plan: MMEA vs. MMA vs. RS. Shared (center): hierarchical, constraint-aware cours

2610.30089v1 — ProactEgoMem: Benchmarking Proactive Memory in Streaming Egocentric Video

• text layer: 119665 chars; tables: 12 (numeric: 5); figures: 30; images: 12; svg figures: 668; figures with MathML: 14
  • table 0: 0 rows — (no caption observed)
  • table 1: 2 rows — (no caption observed)
  • table 2: 10 rows — (no caption observed)
  • table 3: 19 rows — (no caption observed)
  • table 4: 6 rows — (no caption observed)
  • table 5: 42 rows — (no caption observed)
  • table 6: 25 rows — (no caption observed)
  • table 7: 1 rows — (no caption observed)
  • figure 0: 0 img, 0 svg, math=False — Figure 1: Proactive agents observe video, co-exist in the background, and act before users
  • figure 1: 1 img, 0 svg, math=False — Figure 2: ProactEgoMem measures proactive memory as lifecycle behavior in egocentric strea
  • figure 2: 0 img, 1 svg, math=True — Figure 3: Two- level proactive modes .
  • figure 3: 0 img, 6 svg, math=False — Table 1: Illustrative examples of proactive messages for each task and mode.
  • figure 4: 1 img, 0 svg, math=False — Figure 4: ProactEgoMem composition and target objects. Left : composition by scenario (inn
  • figure 5: 0 img, 3 svg, math=True — Table 2: ProactEgoMem-QA question types and benchmark examples
  • figure 6: 0 img, 0 svg, math=True — Table 3: ProactEgoMem data statistics and inter-annotator agreement. Paired entries read D
  • figure 7: 0 img, 312 svg, math=True — Table 4: ProactEgoMem-Daily main results for notification and intervention levels. • Start

2610.29913v1 — Asking a Multimodal LLM Whether Someone Fell: Zero-Shot Screening of Low-Resolution Thermal Video in a Laboratory and an Occupied Home

• text layer: 74085 chars; tables: 6 (numeric: 6); figures: 12; images: 1; svg figures: 0; figures with MathML: 7
  • table 0: 13 rows — (no caption observed)
  • table 1: 7 rows — (no caption observed)
  • table 2: 10 rows — (no caption observed)
  • table 3: 10 rows — (no caption observed)
  • table 4: 8 rows — (no caption observed)
  • table 5: 4 rows — (no caption observed)
  • figure 0: 0 img, 0 svg, math=False — Fig. 1 : The prototype. A thermal camera and a microphone stream continuously; a wake gate
  • figure 1: 0 img, 0 svg, math=False — Fig. 2 : The home recording, 6–13 September 2026: clips the deployed gate and pipeline sen
  • figure 2: 0 img, 0 svg, math=True — TABLE I : Conditions, each changing one factor of the deployed request, and what each was
  • figure 3: 0 img, 0 svg, math=False — Fig. 3 : Laboratory falls called a fall, by fall type (all six subjects): Gemini 3.8 Flash
  • figure 4: 0 img, 0 svg, math=False — TABLE II : The model’s answers, verbatim (deployed request; the laboratory answers’ openin
  • figure 5: 1 img, 0 svg, math=True — Fig. 4 : One staged fall (a slow backward faint) from the public laboratory data [ 6 ] (CC
  • figure 6: 0 img, 0 svg, math=False — Fig. 5 : Effect of sensor resolution (block averaging, same field of view, same display si
  • figure 7: 0 img, 0 svg, math=True — Fig. 6 : Each screener as one point: laboratory falls called (held-out subjects, Wilson 95

Provenance

Every artifact carries a sha256 in MANIFEST.json. Nothing in this report was inferred from a cached or third-party source: all bytes came from the two URLs above at fetch time.