anthology-of-the-sota

LIT-510: Gemini: A Family of Highly Capable Multimodal Models

1 min read

Properties5
StatusActive
Tagsanalysis-and-evaluation, multimodal-learning, in-context-learning
Cited as “introduced by” bySOTA-313 — Treat the inference-time decision procedure as part of what you are comparing, and report the ordering under each one
Cited as “paper” byNOTE-255 — Gemini 1.0: complete disclosure, and an abstract that picks the flattering row
Cited as “source” bySOTA-197 — Account for test-set proximity to the training data when evaluating, and state your contamination exposure, SOTA-313 — Treat the inference-time decision procedure as part of what you are comparing, and report the ordering under each one

Graph View

Gemini Team, Google (2023) — ARXIV-2312.11805, read as NOTE-255.

Standing

Filed for its evaluation appendices, not for its models. On architecture and data it says little a reader can act on — Transformer decoders, 32k context, multi-query attention, a pretraining set of “web documents, books, and code” plus image, audio and video. Nothing there is isolated and nothing is reproducible. What is filed is Appendix 10.2 and the contamination discussion in §5.1.

It measures three things about benchmarking that the record wanted and did not have, each reported voluntarily by the party they embarrass. They are in NOTE-255 and they support SOTA-313 and a second source for SOTA-197.

Which version was read. arXiv v5 (updated 2025-05-09) as served by ar5iv; the report was first posted 2023-12-19, and the published: date is the original posting per the record’s convention.

The record already used Gemini as a subject without holding it. LIT-383 evidences SOTA-233 on Gemini 1.0 Pro and 1.5 Flash, and until now the report describing those models was absent — the same trunk gap #243 was opened about.

Backlinks