Curating Intelligence: How SUPERNOVA Redefines Data Selection for General Reasoning in LLMs

The past year has witnessed remarkable advances in mathematical and coding reasoning through Reinforcement Learning with Verifiable Rewards (RLVR). Models trained via RLVR on datasets like MATH and CodeForces have demonstrated sophisticated chain of thought capabilities, with improvements exceeding 50% on competitive mathematics benchmarks such as AIME24. Yet this success has remained stubbornly confined to formal domains. When researchers apply the same RLVR pipelines to general reasoning tasks requiring causal inference, temporal understanding, or pragmatic reasoning, the results often disappoint. Recent work demonstrates that OpenReasoner 7B, despite its mathematical prowess, actually degrades performance by 8% on Big Bench Extra Hard (BBEH), a benchmark designed to test exactly these broader reasoning capabilities.

The paper SUPERNOVA: Eliciting General Reasoning in LLMs with Reinforcement Learning on Natural Instructions by Suvarna et al. identifies the core bottleneck. It is not the RL algorithms themselves, but rather the geometry of the training data. The authors propose a systematic framework for mining and curating instruction tuning datasets to create verifiable rewards for general reasoning. Their findings challenge the prevailing assumption that reasoning