single mention · post-training
Randomization of RL environments and prompts
Randomization varies the RL training environments and prompts, for example by drawing system prompts from a table of equivalent phrasings.1 A group seed controls which variant of a task the model sees, so the same task appears under different phrasings and languages.2
How this definition was written
Summarised by glm-5.3-flash from 1 source; every sentence is checked against the quote it cites.
The seed draws the system prompt from a table of equivalent phrasings
— tech reportOne workplace-assistant task under three group seeds
— tech report
Also called Randomisation.
Filed under post-training › agentic post-training