The continued advancement of large language models (LLMs) is increasingly reliant on synthetic experience. However, effective data synthesis for interactive agents requires more than just executable trajectories; it demands interactions that provide genuine learning value. A new research paper introduces EvoSyn, a multi-agent learning framework designed to bridge the "feedback-to-decision gap."
The Static Synthesis Problem
Existing data synthesis methods often incorporate environmental feedback or learner failures, but they frequently struggle to translate these observations into specific changes for future data generation. EvoSyn addresses this by organizing synthesis around a "revisable state," which specifies which interaction factors can be deliberately generated and where synthesis efforts should be prioritized.
The EvoSyn Architecture
The system utilizes specialized roles to manage the synthesis process:
- Designer: Allocates synthesis directions and determines where effort is needed.
- Planner: Grounds these directions into executable experiences within the environment.
This structure enables within-space adaptation to target weak or under-covered regions, as well as synthesis-space expansion. The latter promotes newly discovered factors into explicit generation controls, allowing the system to evolve beyond its initial parameters.
Results and Performance
Experiments conducted on $ au^2$-Bench and CAR-Bench demonstrate that EvoSyn improves average task success by 13.3 points over the strongest synthesis baselines. The research indicates that the framework produces higher-quality and structurally broader synthetic experiences while identifying controllable factors that would have been missed without learner feedback.
