Search papers, labs, and topics across Lattice.
The paper introduces CM2, a reinforcement learning framework that utilizes checklist rewards instead of verifiable outcome rewards to train agents for multi-turn, multi-step tool use. CM2 decomposes each turn's behavior into fine-grained binary criteria with evidence grounding, enabling more stable classification-style reward signals. Experiments in an LLM-simulated tool environment demonstrate that CM2 significantly outperforms supervised fine-tuning baselines on benchmarks like tau^-Bench, BFCL-V4, and ToolSandbox, achieving comparable or superior performance to similarly sized open-source models.
Forget hand-crafted reward functions: CM2 uses checklists to train tool-using agents, outperforming SFT baselines by up to 12 points on key benchmarks.
AI agents are increasingly used to solve real-world tasks by reasoning over multi-turn user interactions and invoking external tools. However, applying reinforcement learning to such settings remains difficult: realistic objectives often lack verifiable rewards and instead emphasize open-ended behaviors; moreover, RL for multi-turn, multi-step agentic tool use is still underexplored; and building and maintaining executable tool environments is costly, limiting scale and coverage. We propose CM2, an RL framework that replaces verifiable outcome rewards with checklist rewards. CM2 decomposes each turn's intended behavior into fine-grained binary criteria with explicit evidence grounding and structured metadata, turning open-ended judging into more stable classification-style decisions. To balance stability and informativeness, our method adopts a strategy of sparse reward assignment but dense evaluation criteria. Training is performed in a scalable LLM-simulated tool environment, avoiding heavy engineering for large tool sets. Experiments show that CM2 consistently improves over supervised fine-tuning. Starting from an 8B Base model and training on an 8k-example RL dataset, CM2 improves over the SFT counterpart by 8 points on tau^-Bench, by 10 points on BFCL-V4, and by 12 points on ToolSandbox. The results match or even outperform similarly sized open-source baselines, including the judging model. CM2 thus provides a scalable recipe for optimizing multi-turn, multi-step tool-using agents without relying on verifiable rewards. Code provided by the open-source community: https://github.com/namezhenzhang/CM2-RLCR-Tool-Agent.
Google Research