Search papers, labs, and topics across Lattice.
This paper introduces Reinforcement Learning with Self-Verifiable Rewards (RLSVR), a novel paradigm that transforms open-ended tasks into verifiable proxy environments to generate automatic reward signals. By utilizing a multi-agent self-play environment called SpyRL, agents engage in a voting process to identify a designated spy, allowing for fully verifiable rewards while maintaining relevance to output quality. Experiments demonstrate that RLSVR significantly outperforms existing self-improvement methods on non-verifiable tasks and consistently enhances performance on verifiable reasoning tasks, thereby broadening the applicability of RLVR techniques.
Task transformation enables LLMs to achieve self-improvement in open-ended tasks without the biases and costs associated with human evaluators.
Reinforcement Learning with Verifiable Rewards (RLVR) has driven recent progress in reasoning-oriented large language models (LLMs) by enabling large-scale optimization. However, its applicability remains largely limited to domains such as mathematics and coding, where correctness can be deterministically verified. Open-ended tasks instead often rely on human preferences, reward models, or LLM-based judges, introducing evaluation bias, judge capability bottlenecks, and additional inference costs.Drawing on the principle of self-supervised learning, which constructs pretext tasks to derive supervision from the data itself, we propose Reinforcement Learning with Self-Verifiable Rewards (RLSVR), a task-transformation-based training paradigm for extending RLVR to open-ended tasks. RLSVR transforms open-ended tasks into verifiable proxy environments whose internal rules and interaction outcomes automatically generate reward signals. We instantiate RLSVR with SpyRL, a multi-agent self-play environment inspired by Who Is the Spy?. Agents receive asymmetric information, complete the same target task, and vote to identify a designated spy. Because the spy identity is predetermined, voting outcomes provide fully verifiable rewards, while successful identification remains closely related to output quality. Experiments on text summarization, creative writing, and mathematical reasoning show that SpyRL outperforms existing self-improvement methods on non-verifiable tasks and yields consistent gains on verifiable reasoning tasks. These results demonstrate that task transformation can extend scalable RLVR-based self-improvement beyond inherently verifiable domains. Models and code have been released at https://github.com/wangqinsi1/SpyRL.