Clarification on scorer-code versions and leaderboard scores

Hi AVGen-Bench authors,

I’m Mukesh from NVIDIA, and I’m evaluating AVGen-Bench results for comparison with the public leaderboard.

I am seeing differences between my reproduced scores and the public leaderboard values, particularly for LTX-2.3 (Outputs imported from here).

My understanding is that the public LTX-2.3 scores were generated in March, while the OSS Gemini scorer code received substantial updates in May, including the robustness change and the official-client change. It appears that the static leaderboard values may not have been recomputed after those updates.

For reference, below are examples of the differences I observe for LTX-2.3:

Could you please confirm:

• Which scorer-code version was used for the public LTX-2.3 leaderboard scores?
• Are score differences expected when using the current OSS Gemini scorers?
• Are there plans to recompute or version the public leaderboard results?

Thank you.

Start by comparing the linked LTX-2.3 outputs with the scorer changes in commits 75343b04ca49dfbb0ea014eb6c695bc66945fc4f and 57d8a37bbfb8c4c704f4780be76841c7e58219fd. Check the repository's scorer entry points and leaderboard data to identify the version used; done when the public-score provenance and any recomputation or versioning plan are documented.