Conversation for Non-verifiable Learning: Self-Evolving Large Language Models through Meta-Evaluation

Training large language models (LLMs) for non-verifiable tasks—such as creative writing, dialogue, and ethical reasoning—remains challenging due to the absence of ground-truth labels. While LLM-as-Judge approaches offer a scalable alternative to human feedback, they face a fundamental limitation: performance is constrained by the evaluator's own quality. If the judge cannot recognize good solutions, it cannot provide useful training signals, and evaluation biases (e.g., favoring verbosity over quality) remain unaddressed. This motivates meta-evaluation—the ability to evaluate and improve the evaluator itself. We introduce CoNL, a framework that unifies generation, evaluation, and meta-evaluation through multi-agent self-play. Our key insight: critique quality can be measured by whether it helps others improve their solutions. In CoNL, multiple agents sharing the same policy engage in structured conversations to propose, critique, and revise solutions. Critiques that enable other agents' solution improvements earn a diagnostic reward, creating explicit supervision for meta-evaluation and enabling joint optimization of generation and judging capabilities through self-play, without external judges or ground truth. Experiments on various benchmarks show that CoNL achieves consistent improvements over self-rewarding baselines while maintaining stable training.

Some tasks have a clear ground-truth answer, like math or coding problems. We can train LLMs on those by checking whether the answer is right. But many tasks don't work that way, like writing a poem, giving health advice, or thinking through an ethical question. There is no ground-truth answer to check against. A common workaround is to have another LLM grade the work. The problem is that if the grader has bad taste, say it tends to pick the longer answer over the better one, training just bakes in that bias. Nothing is checking the grader itself. Our idea borrows from how Wikipedia improves over time: someone writes a draft, others spot problems, and the draft gets revised. We set up four LLM agents, all sharing the same model, to do something similar. They each propose an answer, rank and critique each other's answers, then revise based on the feedback. The key insight is that a critique only counts as useful if it actually leads to a better revised answer. That single rule gives us a training signal for both writing and judging at the same time, without ever needing a labeled correct answer. We tested this on creative writing, health advice, research planning, and hard reasoning problems. It clearly beats other label-free methods and comes close to methods that train with the ground-truth labels. This matters because most real-world tasks don't come with such labels, and getting experts to write them is slow and expensive.