自我对弈仅在自我合成管道确保可学习信息增益时才能演化

Large language models (LLMs) enable the construction of systems that improve through self-evolving loops, yet many existing proposals are better characterized as self-play and often reach a plateau quickly. A central failure mode is the synthesis of more data without increasing learnable information for subsequent iterations. Experiments conducted on a self-play coding task reveal that sustainable self-evolution necessitates a self-synthesized data pipeline that ensures an increase in learnable information across iterations. The study identifies triadic roles played by self-evolving LLMs: the Proposer, which generates tasks; the Solver, which attempts solutions; and the Verifier, which provides training signals. Additionally, three system designs are identified that jointly target the enhancement of learnable information.