Gemini 1.5 Technical Report

Showing top 8 of 141. See full list below.

• “we introduce the Gemini 1.5 family of models, representing the next generation of highly compute-efficient multimodal models capable of recalling and reasoning over fine-grained information from millions of tokens of context, including multiple long documents and hours of video and audio.”

• “not released in the Gemini 1.0 models.”

• “mitigations include: Safety filters with established thresholds to set responsible default behaviors.”

• “available to the model.”

• “future work. These models also show improvements in jailbreak robustness and do not respond to “garbage” token attacks, but they do respond to handcrafted prompt injection attacks – potentially due to their increased ability to follow the kind of instructions in the prompt injection.”
• “future work. 9.5. Assurance Evaluations Assurance evaluations are our ‘arms-length’ internal evaluations for responsibility governance decision- making (Weidinger et al., 2024).”