- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper presents CompoWorld, a framework that composes reusable software services into verifiable cross-service tasks for LLM agent training. It combines automated environment synthesis, dependency-based task generation, SFT, and a completion-focused RL reward. Experiments on eight benchmarks show an average improvement of 9.17 points over the Qwen3.6-35B-A3B backbone.
Critical strengths
- Practical framework: The reusable service library and automated task verification provide a promising approach to scalable agent training.
- Encouraging results: The single-versus-composed environment comparison suggests that cross-service tasks can improve agent performance under a fixed data budget.
Critical weaknesses
- Weak compositional dependency modeling. Random walks do not guarantee meaningful cross-service reasoning. The generated tasks may primarily increase tool-call length rather than require genuine compositional generalization.
- Limitations of the proposed reward. Completion-Focused Rubric Reward relies on pass-rate estimates from only eight rollouts and provides no discrimination when all rollouts fail a criterion. Its advantage over uniform rubric rewards remains insufficiently established.
Questions for authors
- How do you ensure that generated dependencies require genuine cross-service reasoning?
- How robust is the proposed reward to rollout-group size and universally failed criteria?
2 comments
Thank you again for your constructive feedback. We have provided detailed responses to your concerns regarding compositional dependencies and the completion-focused reward.
We would greatly appreciate it if you could take another look at our rebuttal and let us know whether it addresses your concerns. If any questions remain, we would be happy to provide further clarification.
Thank you very much for your evaluation of our work and your constructive comments.
W1&Q1: We would like to clarify that Random Walk does not simply randomly concatenate services; rather, it is used to construct cross-service workflows with explicit data dependencies and state dependencies. Specifically, the inputs required by a downstream service must come from the execution results (Observation) of preceding services and cannot be obtained independently. In addition, we require that at least one service be visited repeatedly, while avoiding consecutive visits to the same service, thereby introducing dependencies caused by state changes—for example, the Agent needs to revisit a service after operations on other services in order to obtain the updated state.
Figure 5(a) further verifies the value of compositional-environment training: under the conditions that the number of training trajectories (1K), the number of environments, and the SFT pipeline are kept consistent, compositional-environment training achieves better results on all four benchmarks. Appendix I also provides complete cross-service execution trajectories as concrete examples.
W2&Q2: We would like to clarify that, for Criteria that no Rollouts pass, our method indeed cannot further distinguish between different trajectories. This is essentially the same as in GRPO: when all Rollouts receive the same reward, the within-group relative advantage is zero, and no effective optimization signal can be provided.
Regarding Rollout Group Size, our method adopts the same within-group sampling and relative estimation mechanism as GRPO, so its sensitivity to Group Size has the same nature. Increasing the number of Rollouts can improve the stability of completion-rate estimation, but it also increases sampling cost. Group Size = 8 is a common setting for GRPO.
The core advantage of our method lies in the fine-grained reward signal. For Criteria where some Rollouts succeed and some fail, we increase their relative weight so that the model focuses on optimizing capabilities that have not yet been stably mastered; for Criteria where all Rollouts succeed or all fail, no additional discriminative signal is produced. Therefore, our method essentially refines the within-group relative optimization in GRPO to the Criterion level, and thus will certainly provide stronger supervision signals than GRPO.