Human video Elo vs physics scores

Short answer: On 14 Nov 2026 we compared published leaderboards, not model outputs. The Artificial Analysis text-to-video board (AA-Video-T2V v2.1, 34 models, Wan 3.0 first at Elo 1175.13) did not share four unambiguously matched models with Physics-IQ Verified (53 rows), Physics-IQ Original (32 rows) or VideoPhy-2 hard (7 models). Under our pre-registered rules we therefore computed no rank correlation. Physics-IQ lists image-to-video and video-to-video rows, not text-to-video; VideoPhy-2 lists older text-to-video names that do not appear under the same version on the human board.

By Evals Review editorial team · Snapshots 14 Nov 2026, 6:23 PM IST (Asia/Calcutta) · How we review. No affiliate links on this page. $0: public pages only. We did not run any video model.

Limits: read these before quoting a number

  • This is a comparison of published tables on one date, not a new leaderboard and not a run of any model.
  • Zero matched pairs means we do not report Spearman or Kendall figures. A missing correlation is not evidence that either board is wrong.
  • Input type matters. We did not treat an image-to-video physics score as if it belonged to a text-to-video Elo row, even when the marketing name was the same.
  • Two boards could not be read. LMArena Video returned HTTP 403 (a Cloudflare challenge). The official VBench-2.0 totals live in a Hugging Face dataset that returned HTTP 401. We did not invent those numbers.

Method (pre-registered)

The method was committed before the first snapshot: PREREGISTRATION.md (git commit eb28592, 2026-11-14 18:23:11 IST).

How we compared the public tables
Human boardArtificial Analysis AA-Video-T2V v2.1, text-to-video only, parsed from the official page payload. 34 models. The page also labelled "7 models added in the last 30 days".
Physics boardsOfficial Physics-IQ Verified and Original tables in the google-deepmind/physics-IQ-benchmark README (commit d3f3de1), and the official VideoPhy-2 hard-subset joint scores.
A matchSame model family and the same version token, the same input type (text-to-video only against T2V rows), and no best-of-N, reward or rewriter add-on on only one side. Ambiguous names stayed unmatched.
Correlation ruleSpearman and Kendall tau-b only if a pair of boards had 4 or more matches; a bootstrap interval only if there were 6 or more. Fewer than 4 matches: publish the tables, do not compute a correlation.
Watch the short video of this comparison (AI-voiced narration)

Results

Readable boards, 14 Nov 2026
BoardRows we parsedInput typeMatches with AA-Video-T2V v2.1
AA-Video-T2V v2.134 models (human Elo). First: Wan 3.0 at 1175.13, 7,787 appearances.text-to-video—
Physics-IQ Verified5336 image-to-video, 17 multiframe (video-to-video). No text-to-video row.0
Physics-IQ Original3222 image-to-video, 10 multiframe (video-to-video). No text-to-video row.0
VideoPhy-2 hard (joint)7 (open and closed tables)text-to-video action prompts0

VideoPhy-2 is the physics board that is actually text-to-video. Its listed models (Wan2.1-14B at 21.9 hard joint, plus CogVideoX-5B, Cosmos-Diff-7B, Hunyuan-13B, VideoCrafter-2, Ray2 and Sora) do not appear under those same version names on AA-Video-T2V v2.1, which is led by later names such as Wan 3.0.

Same family, still unmatched

Several human-board names sit near a physics-board name. Under the pre-registered rules they are not matches. The three exact-name pairs below differ by input type; the others also differ by version token.

Illustrative unmatched pairs, not a correlation
Human board (T2V Elo)Physics-IQ VerifiedWhy unmatched
Veo 3.1 Lite (rank 27, Elo 947.88)Veo 3.1 Lite (rank 46, score 31.83, i2v)Same version name; text-to-video board vs image-to-video physics row.
Veo 3.1 Fast (rank 26, Elo 954.76)Veo 3.1 Fast (rank 50, score 29.96, i2v)Same version name; text-to-video board vs image-to-video physics row.
MiniMax H3 Max (rank 8, Elo 1149.55)MiniMax H3 Max (rank 38, score 36.2, i2v)Same version name; text-to-video board vs image-to-video physics row.
Veo 3.1 (rank 25, Elo 963.37)Veo 3.1 Lite / Veo 3.1 Fast (i2v)Related names, not the same version token, and input types differ.
Wan 3.0 (rank 1, Elo 1175.13)Wan 2.2 14B / Wan 2.2 5B (i2v)Same family, different version (3.0 vs 2.2), and input types differ.
MiniMax H3 (768p) (rank 3, Elo 1160.91)MiniMax H3 (FL2VA) / MiniMax H3 Max (i2v)768p and FL2VA / Max are not the same version token; input types also differ.

Boards we could not read

  • LMArena Video ( arena.ai/leaderboard/video): HTTP 403 Cloudflare challenge on arena.ai/leaderboard/video and the related paths we tried. No ranks taken.
  • VBench-2.0 official totals: Official project page has no numeric table. Official VBench-2.0 results CSV in dataset Vchitect/vbench_leaderboard_submission returned HTTP 401. We did not invent a total.

Credit

The tables belong to their authors. Artificial Analysis publishes the Video Arena Elo boards. Physics-IQ is by INSAIT, Sofia University and Google DeepMind; Physics-IQ Verified is by Anates Labs with University of Technology Nuremberg, Tübingen AI Center, TU Munich / Helmholtz AI, and Google DeepMind. VideoPhy-2 is by UCLA and Google Research. VBench-2.0 is by Shanghai AI Laboratory, NTU S-Lab and collaborators. LMArena publishes the arena.ai video boards we could not read.

Data, scripts and licence

Download the pre-registration, source log with retrieval times, summary, parsed CSVs for AA-Video-T2V v2.1, Physics-IQ Verified, Physics-IQ Original and VideoPhy-2, plus raw HTTP snapshots under /data/tests/2026-11-14-human-votes-vs-physics-scores/raw/ (file index). The analysis script is in our repo under tests/human-votes-vs-physics-scores/analyze.py.

Licence: our method notes, parsed tables and this page text are CC BY 4.0. Attribution: Evals Review editorial team, Human video Elo vs physics scores, 14 Nov 2026, https://evalsreview.com/tests/human-votes-vs-physics-scores. The underlying leaderboards remain the property of their authors and are used here as published snapshots.