What it does
RS-1 scores the options without fine-tuning or generating text. The reranker cannot handle every mechanical step in complex browser tasks, though. Hardcoded rules still match values to fields and recognize when the task has finished.
Benchmarks & research
RS-1
RS-1 uses an unmodified local Qwen3-Reranker-0.6B to score and select actions for quizzes, Wikipedia races, and browser automation.
Bookmark: RS-1
Keep this in your collection.
Leave a noteWhat would you try with this?
: RS-1
Only you can see your notes.
Screenshot unavailable. Open the experiment ↗
How you can use it
Think of the situation as a search query and the possible actions as documents to rank. RS-1 takes the highest-scoring action; the quiz and Wikipedia-race demos show that approach. For browser tasks, ordinary rules still have to fill fields and recognize completion.