Visual Orchestration Tax in Agentic VLM Pipelines: Auditing and Certifying Visual Evidence Reuse
4.20T1 sourcearXiv cs.MA
Source record
Published by arXiv cs.MA (T1 source). The original is at https://arxiv.org/abs/2610.08170.
Pipeline notes
The summary and note below are generated by the signal pipeline — they are Beyond Desk’s reading, not quotations from the source.
SummaryPaper identifies 'visual orchestration tax': in agentic VLM pipelines, the same static image is repeatedly reconstructed as image-conditioned requests across specialist agents, causing 66.8-75.6% touch redundancy. Proposes audit metrics (M1, M2) and SharedVisCache, a contract-aware reuse hook that preserves outputs while reducing physical-layer vision calls 4x.
Why it mattersNames a measurable redundancy mode in multi-agent VLM stacks and ships a contract-checked cache that cuts compute calls without changing outputs. Useful before scaling agent count.
Cited by
No citations on record.
