AgenticVBench · agentic video understanding

Contribute a task.
Co-author the papers.

One or multiple videos, one unambiguous question, and a deterministic scorer. Build one to the standard and it merges. Discuss your idea with our agent first — it takes a couple of minutes and saves you days.

Submit by July 31 to make the first batch

Why join the first batch

Real credit, and people who show up.

Closes July 31, 2026

Your name on the ICLR paper

We submit to ICLR this September. Every task that merges before the deadline goes in, credited to you. Land two and you co-author both the benchmark and the survey.

In person

We give back to the community

We co-host events in San Francisco and at the major ML conferences, together with frontier labs and other AI infrastructure companies. Come learn, share what you are building, and meet the people behind it.

No busywork

One good task is plenty

About 10 hours of focused work, on average. One task that passes review is worth more than ten thin ones.

What makes a task

Hard, long, and impossible to fake.

Long horizon

A real strong-agent attempt takes more than 50 tool-call turns — the answer lives across the whole video, not one lookup.

Hard for the right reason

A strong current agent scores below 0.10; your oracle scores 1.0 and an empty attempt near 0. Difficulty comes from the skill, not trick wording.

Deterministically scored

Pure code grades the answer — no VLM or LLM judge — strict enough that guessing scores about 0.

No shortcuts

Recall of famous footage, on-screen graphics, a single frame, or no media at all must not solve it.

Step 1 · discuss your proposal

Is your idea hard enough? Ask before you build.

Describe the video and the question you have in mind. The agent checks it against the bar — difficulty and long horizon first — and suggests the smallest change that would sharpen it.

Proposal agent · checks difficulty & long-horizon fit

Try one of these:

Steps 2–3 · build and calibrate

Build the task, then calibrate it.

Build in the repo, then run it against real agents and grade them. Iterate the two until every agent is well below the bar — that loop is the work.

Step 2 · build

  • Gather one or more real videos and write one unambiguous question.
  • Write the oracle solution — it scores 1.0, while an empty attempt scores near 0.
  • Write the deterministic verifier: pure code, no VLM or LLM judge.
  • Ablate the shortcuts — famous footage, on-screen graphics, or a single frame must not solve it.

Step 3 · calibrate (the loop)

  • Run the task with Antigravity, Codex, and Claude Code.
  • Save each full rollout and grade it with your verifier.
  • If any agent clears 0.10, or solves it in under ~50 turns, tighten the task and re-run.
  • Done when all three score below 0.10 over 50+ tool-call turns.

Step 4 · submit

Open a PR before July 31.

Open a pull request adding your task folder — the prompt, the verifier, the oracle, and your calibration bundle (the rollouts and scores for Antigravity, Codex, and Claude Code). A maintainer reviews and merges — and you are on the paper.