SUPERNOVA: Reinforcement Learning Extends LLM Reasoning Beyond Math and Code to General Tasks

arXiv·medium signalXBlueskyLinkedInCopy link

SUPERNOVA addresses the fundamental constraint that RLVR is limited to formal domains (math, code) by generating high-quality verifiable training data spanning diverse reasoning skills like causal inference and temporal understanding. The method uses natural language instructions rather than formal verifiers, enabling RL-based reasoning improvements on general tasks where ground-truth verification was previously unavailable. This is a significant step toward reasoning LLMs that generalize beyond narrow benchmarks.