Gemini 1.5
Google DeepMind — Gemini
rich detail 8 method fields published
Evaluation
Adversarial long-context needle-in-a-haystack safety evaluation
Adversarial version of text needle-in-a-haystack: adversarial needle (T2T policy-violating prompt) inserted in Paul Graham essay haystack across 25 settings (1k-1M tokens, varying depth); automated short-context T2T evaluation applied
How it was run
Open a release to see the setup, scoring, and other details that the lab published. Only fields the lab actually disclosed are shown.