Researchers from Meta FAIR and the National University of Singapore have introduced SPICE (Self-Play In Corpus Environments), a novel reinforcement learning framework designed to help AI systems improve autonomously. SPICE operates with two AI agents: the Challenger, which generates challenging problems from a large corpus, and the Reasoner, which attempts to solve these problems without access to the source material. This approach prevents information symmetry and hallucination by grounding tasks in real-world documents, enabling AI systems to continuously learn from external sources rather than just their own outputs. SPICE allows the generation of diverse question formats and is applicable across multiple domains, overcoming limitations of previous self-play AI methods confined to specific fields. Evaluations show SPICE enhances reasoning capabilities significantly across various models and benchmarks, proving an effective self-improvement paradigm by creating an evolving curriculum through adversarial interactions between agents.
Back
