Training-Free Anti-Recomputation for Video Vision-Language Models
Research code, artifacts, and manuscript tooling for training-free anti-recomputation in video vision-language models.
The repo is organized around a small set of claim-bearing regimes:
- C-CEILING: component speedups survive to end-to-end latency only in proportion to the dense wall-clock share they own.
- C-PERSIST: after ingest, same-video follow-up queries can be much cheaper inside a tested cache-reuse envelope.
- C-VISION: bounded measured sparse-vision execution exists; broad sparse backends and sparse LM prefill remain open.
- candidate C-STREAM: native-rate streaming state reuse has a checked mixed/boundary bundle, but it is not an earned headline until a native policy beats matched baselines under cache-correctness and stale-cache tests.
The durable imported-target summary is in docs/claim-register.md, local reproduction status is in docs/reproduction-status.md, and raw history remains in git.
uv sync --locked --group dev --group research
uv run ruff format --check .
uv run ruff check .
uv run mypy src tests
uv run pytest
uv run python scripts/audit_artifact_integrity.pyFor local MLX / MLX-VLM research utilities:
uv sync --locked --group dev --group research --group vlmFor local corpus assets:
uv run python scripts/fetch_corpus.py --tier primary --encode
uv run python scripts/generate_synthetic_corpus.pyFor benchmark-native TOMATO / MVBench / VideoMME assets, start with docs/benchmark-setup.md. VideoMME uses checked manifest subsets and a separate subset fetch path documented in docs/videomme-download-handoff.md.
For the paper:
uv sync --locked --group dev --group research --group benchmark --group paper
brew install tectonic # macOS; any XeLaTeX/Tectonic install also works
make paper-doctor
make paper-sync
make paper-buildFor readers and reviewers:
- paper/arxiv/README.md: manuscript build and generated assets
- paper/claim-matrix.md: paper-facing claim truth table
- paper/publishability-status.md: current reviewer-facing claim inventory
- docs/reproduction-status.md: local reproduction status
- research/experiments/registry.md: phase/artifact ledger
For contributors and agents:
- AGENTS.md: canonical coding-agent guidance
- PLAN.md: current roadmap and open gates
- docs/README.md: durable docs router
- research/README.md: experiment-note workflow
- docs/methodology/performance.md: timing and denominator rules
.
├── docs/ stable methodology, setup, literature, and status
├── paper/ arXiv manuscript, generated assets, and paper claim ledgers
├── research/ dated experiment notes, registry, and checked artifacts
├── scripts/ reusable runners, analyzers, validators, and plotters
├── src/ importable codec_through package
└── tests/ unit tests for reusable code
Checked research artifacts remain in this repo when they directly support tables, figures, status claims, or review/publication previews with adjacent provenance notes. Large future bundles should use a manifest with checksums, but deleting current artifact evidence would make the paper harder to audit.
- label claims as reproduced here, imported target, or hypothesis
- separate semantic answer stability from real skipped work
- report denominators and setup costs explicitly
- keep negative results when they change the claim boundary
- use primary sources for literature and standards claims
Cite the paper as arXiv:2605.03351. CITATION.cff includes the preferred paper citation for GitHub's citation UI.
This repository is multi-licensed.
Code, scripts, tests, software configuration, and paper/build tooling are licensed under MIT.
Original documentation, research notes, manuscript source, generated paper
figures/tables, and non-code research artifacts created solely by this repo are
licensed under Creative Commons Attribution 4.0 International
(CC-BY-4.0).
Benchmark-derived preview artifacts carry provenance notes in their artifact folders.
