Seirênes: Self-Play RL Framework Boosts LLM Reasoning Against Distractions

Researchers have introduced Seirênes, a self-play reinforcement learning framework that turns contextual interference into a training signal for large language models (LLMs). The framework addresses the fragility of LLMs when faced with non-idealized contexts—such as superfluous information, tangential instructions, or incidental correlations—that differ from clean benchmark distributions. Seirênes uses a parameter-shared adversarial self-play loop where a single model both constructs plausible distracting contexts that expose its own reasoning blind spots and solves problems by discerning the essential task from these perturbations. This approach co-evolves more resilient reasoners by pitting competing objectives against each other. The work is detailed in a preprint on arXiv (2605.11636).