AI brief
A LessWrong post describes a toy experiment in which Qwen3.5-2B was allowed to choose its own reward during GRPO training on introspection-adjacent prompts. The author says most design choices were not ablated and lists caveats at the end of the post.
Written by AI from LessWrong's published text.
Source
“Epistemic status: This is mostly a toy project rather than a thorough scientific investigation. Most design choices are not ablated. We list some caveats at the end of the post.”
Interpretation or community commentary rather than straight reporting.
Read original story ↗