Caption-once, Frames-on-Demand: Visual-Need Routing for Budget-Aware Agentic Long Video Understanding¶

L1 core-mechanism diagnostic. Executes a caption/query gate and top-k frame selection, not a long-video VLM.

Paper information¶

Field	Details
Paper	Caption-once, Frames-on-Demand: Visual-Need Routing for Budget-Aware Agentic Long Video Understanding
Organization	First-author affiliation: Queen Mary University of London
First public release	2026-09-10
Authors' code	No public author repository was found on 2026-09-14.
Adapter / method	`frames-on-demand`
Local reproduction code	`src/auto_research/foundation_latest_20260914.py`

Original paper¶

Motivation and method¶

Build caption memory once for each video. The visual-need router reads relevant frames within a fixed budget only when text evidence is insufficient.

flowchart LR
  I[Input and public state] --> M[Frames-on-Demand mechanism]
  M --> A[Auditable intermediates]
  A --> O[Output or update]

Key figure from the paper¶

Figure 1: Overview of the paper's method and key components. From the original paper; copyright remains with the authors.

Core computation or operation¶

The local deterministic reference kernel implements the defining operation and reports gate, selection, projection, graph-edge, or state-update statistics. Tests cover formula invariants and boundaries.

Paper results¶

The paper reduces visual frame use by roughly an order of magnitude while remaining competitive. This is not directly comparable to the scaled-down local run. No online gains are inferred without an online A/B test.

Local reproduction¶

See the three-seed metrics. All candidates share the public fixtures, budget, and seeds. These results verify actual execution of core state transitions only.

Reproduction boundaries¶

The caption/query gate and top-k frame selection run locally. No long-video VLM was run.