LetTokenBurn
Sign in
Auto57%Harbor: pendingAgentic / Tool UseMultimodal

CineRecon

CineRecon: Benchmarking Frontier Models for Agentic Cinematic Reconstruction

A benchmark of 50 reference videos and 360 annotated clips evaluating agents' ability to reconstruct editable dynamic 3D scenes from video.

Paper

How to run

ltb · cinerecon
$ npm i -g lettokenburn-cli
$ ltb bench info cinerecon

External harness. No Harbor adapter has been generated for this benchmark yet. `ltb bench info` prints the known links.

Runs happen on your machine with your own subscription; only the score summary is pushed. CLI docs

Scores

12 sourced entries
HarnessSourceConfidenceLink
Claude Opus 5.576.5claude code 2.1.280IndependentmediumOct 12, 2026arxiv.org
GPT-6 Astra67.6codex cli 0.153.4IndependentmediumOct 12, 2026arxiv.org
Kimi K357.3kimi code cli 2.0.2IndependentmediumOct 12, 2026arxiv.org
Grok-4.654.7grok build 1.0.41IndependentmediumOct 12, 2026arxiv.org
DeepSeek V4.1 Flash53.9dsh 0.1.5-rc.3IndependentmediumOct 12, 2026arxiv.org
GLM-5.3 Flash53.8zcode cli 0.16.9IndependentmediumOct 12, 2026arxiv.org
Gemini 3.8 Flash52.8antigravity cli 1.2.11IndependentmediumOct 12, 2026arxiv.org
Qwen3.8 Max49.8qwen code 0.24.1IndependentmediumOct 12, 2026arxiv.org
Claude Opus 4.843.3claude code 2.1.278IndependentmediumOct 12, 2026arxiv.org
Qwen3.8 Flash42.7qwen code 0.24.1IndependentmediumOct 12, 2026arxiv.org
GPT-5.6 Sol39.7codex cli 0.153.4IndependentmediumOct 12, 2026arxiv.org
Gemini 3.1 Pro22.5antigravity cli 1.2.11IndependentmediumOct 12, 2026arxiv.org

Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.

Reproduced by the community

All runs
Loading community runs…

About

Comprises 50 videos from film, real-world footage, and games. Agents construct Blender scenes under Proxy, Simplified, or Detailed settings.

Scored by a vision-language judge across camera, static, and dynamic consistency. Evaluations reveal persistent failures in subject motion and interactions.