Tag
SDAR enhances multi-turn agent training by integrating self-distillation with a sigmoid gate to selectively strengthen positive token-level guidance while mitigating negative teacher rejections, achieving significant improvements over GRPO across multiple benchmarks.