Benchmarks LongTail Challenge open · NeurIPS 2026

LongTail World Modeling

Can a world model imagine what happens next in a snowstorm, a road closure, or at night? Given the past of a long-tail scenario, generate the next five seconds of six-camera video, optionally together with the planned ego trajectory. Generated futures must look real and move consistently: across all cameras, and with the plan.

Nov 7 – Dec 3
2026, anywhere on earth
400
long-tail test scenarios
6 × 25
frames per scenario, 5 fps
2,000 €
prize money

The Task

For every test scenario, generate an H.264 video per camera with exactly 25 frames at 5 fps, covering 0.2 to 5.0 s into the future, at the KITScenes aspect ratio (e.g. 876 × 568). Only the front camera is required, but ego-motion coherence needs all six and perceptual quality weights the other five cameras too. A trajectory file with 25 (x, y) points per scenario is optional; without it, EMC, MMS and L2 are N/A. Since LongTail v2.1, the val and train splits include the future videos for training.

Five seconds of a snowy long-tail scenario: (a) ground truth, (b) Cosmos 3, (c) Epona (front camera only).

Leaderboard

Live · fetched from HuggingFace · best submission per method · ranked by PQ, ties broken by EMC × MMS

View full leaderboard ↗
Loading leaderboard…

Per-type columns show PQ per scenario type. N/A: submitted without a trajectory file.

Metrics

Submissions are ranked by perceptual quality. Ties are broken by EMC × MMS, which rewards videos whose implied ego motion matches a good plan.

PQ

Perceptual quality

↑ ranking metric

Each camera video is rated by Q-Align (OneAlign), following the perceptual quality metric of WorldMark, mapped to [0, 100] and normalized by the quality of real long-tail calibration videos. The front camera weighs 50%, each other camera 10%; missing or unreadable videos score 0, so front-only submissions reach at most half the maximum.

EMC

Ego-motion coherence

↑ tie-breaker (× MMS)

Recovers the ego motion implied by the videos with optical flow: yaw from the front and rear cameras, which must agree, speed from the four side cameras. Planned and recovered trajectories are split into 0–3 s and 3–5 s and classified into steering and acceleration actions with the SRAC heuristics; a class counts only at similar speed. Needs all six cameras; normalized by real reference videos.

MMS

Multi-maneuver score

↑ tie-breaker (× EMC)

Scores the submitted trajectory against reference maneuvers: expert-like (10), wrong speed (7), neglect instruction (4), off road without crashing (1), crash (0); unmatched 3.5. Comfort penalties of 1 each for jerk more than 44% and tortuosity more than 6% above the reference.

Example of the multi-maneuver score with reference trajectories for expert-like driving, wrong speed and crash
Multi-maneuver score: a plan is scored against the best-matching reference maneuver.

Prizes

🥇 First place 1000 €
🥈 Second place 600 €
🥉 Third place 400 €

Winners are encouraged to send a 4-minute video of their method, presented at the World Models in Physical AI workshop at NeurIPS 2026 and posted on this website.

Rules in Brief

  • Submissions from November 7 to December 3, 2026 (AoE), through the challenge Space on HuggingFace.
  • Teams are welcome; each participant may be part of one team only. One team member needs a HuggingFace account.
  • Up to 5 submissions per team and month; the submission with the highest PQ counts for the winners.
  • Upload a .zip of <scenario_id>/<camera>.mp4 videos (at most 3 GB, 50 MB per video) and optionally a .jsonl with 400 trajectories.
  • Evaluation takes about 20–30 minutes; signed-in users can track its progress in the Space.

The full rules, eligibility and data processing terms are in the challenge Space and apply.