4 upvotes
Outcome RL lifts single-shot success, not broader reliability
What is What Outcome RL Buys: Finishing, Reliability, Rarely Reach?
Outcome reinforcement learning improves the pass@1 score for code generation on physics and math problems, but it does not increase pass@k when k is large. The study compares 20 released RL models ranging from 1.5 B to 32 B parameters against a chain-of-thought “thinking mode” under the same token budget.
Who it is for
Researchers and developers of code-generation tools should know RL helps finish a single attempt but adds little to overall reliability, guiding model-training choices.