VideoMM: Adaptive Macro-Micro Inference for Efficient Video MLLMs¶
L1 core-mechanism diagnostic. Executes macro selection and micro activation without loading a full video MLLM or extrapolating LongVideoBench performance.
Paper information¶
| Field | Details |
|---|---|
| Paper | VideoMM: Adaptive Macro-Micro Inference for Efficient Video MLLMs |
| Organization | First-author affiliation: University of Science and Technology of China |
| First public release | 2026-09-15 (arXiv v1) |
| Authors' code | Public implementation |
| Adapter / method | videomm |
| Local reproduction code | src/auto_research/foundation_latest_20260916_followup.py |
Original paper¶
Motivation and method¶
Locate relevant regions on a low-resolution Macro Proxy, then activate high-fidelity Micro Tokens only if macro-level consensus is insufficient.
flowchart LR
I[Public input and current state] --> M[VideoMM mechanism]
M --> A[Auditable intermediates]
A --> O[Output or state update]Key figure from the paper¶
Figure 2: Core architecture and module connections. From the original paper; copyright remains with the authors.
Core computation and code mapping¶
The reference kernel preserves the decisive ranking, gating, preference direction, credit weighting, or state transitions and reports auditable statistics. Three-seed fixtures test invariants and edge cases; randomized mini-suite scores are not paper capability results.
Paper results¶
Paper accuracy, speed, training loss, and benchmarks are cited as original results only. Local scores are not directly compared or extrapolated to online gains.
Local reproduction¶
See the three-seed results. The common receipt marks diagnostic_only=true, ineligible for formal capability ranking.
CUDA ran on an NVIDIA A100; see the sanitized GPU receipt. It omits hostnames, SSH aliases, and driver/build strings.
Reproduction boundaries¶
Local execution covers macro selection and micro activation, not a full video MLLM or LongVideoBench outcomes.