Know What You Don't Know: Agentic Data Synthesis by Evolving Generation Space for Closed-Loop Learning
Quick Answer
This paper shows that EvoSyn, a multi-agent framework for data synthesis, enhances interactive agent learning by addressing the feedback-to-decision gap, achieving a 13.3-point increase in task success over existing baselines on $ au^2$-Bench and CAR-Bench.
Quick Take
This approach enables adaptive synthesis and the discovery of new controllable factors through learner feedback.
Key Points
- EvoSyn organizes data synthesis around revisable states for better learning interactions.
- The framework includes specialized roles for feedback translation and synthesis planning.
- Experiments show a 13.3-point improvement in task success over the strongest baseline.
- EvoSyn produces higher-quality synthetic experiences and identifies previously missed factors.
- The approach promotes synthesis-space expansion for enhanced generation control.
DeepSignal Analysis
What happened
The paper introduces EvoSyn, a multi-agent framework designed to enhance data synthesis for interactive agents. It addresses the feedback-to-decision gap by organizing synthesis around a revisable state that specifies interaction factors. The framework reportedly improves task success by 13.3 points on benchmarks.
Key evidence
- EvoSyn improves average task success by 13.3 points over existing baselines on the $\tau^2$-Bench and CAR-Bench.
- The framework organizes synthesis around a revisable state, allowing for deliberate generation of interaction factors.
- EvoSyn enables the discovery of new controllable factors through learner feedback, which were previously missed.
Why it matters
EvoSyn's approach to data synthesis could significantly enhance the learning capabilities of interactive agents by effectively utilizing feedback. This could lead to more robust models that adapt to new challenges and improve performance in real-world applications. The ability to discover new controllable factors may also expand the scope of tasks that these agents can handle.
What to watch
Paper Resources
📖 Reader Mode
~2 min readAbstract:The continued improvement of large language models increasingly depends on synthetic experience, yet effective data synthesis for interactive agents requires not only executable and verifiable trajectories, but also interactions that provide learning value for the current task agent. Existing methods increasingly incorporate environment feedback, data validation, or learner failures, but still face a feedback-to-decision gap: observed weaknesses do not directly specify how future synthesis should change, or whether they can be represented by the current synthesis space. To address this gap, we propose EvoSyn, a learning-feedback-driven multi-agent framework that organizes synthesis around a revisable state specifying what interaction factors can be deliberately generated and where synthesis effort should be allocated. A role-specialized multi-agent system translates learner feedback into synthesis revisions, while a Designer allocates synthesis directions and a Planner grounds them into executable experience. This process enables within-space adaptation toward weak or under-covered regions and synthesis-space expansion that promotes newly discovered factors into explicit generation controls. Experiments on $\tau^2$-Bench and CAR-Bench show that EvoSyn improves average task success by 13.3 points over the strongest synthesis baseline, while producing higher-quality and structurally broader synthetic experience and discovering controllable factors missed without learner feedback.
| Subjects: | Artificial Intelligence (cs.AI); Multiagent Systems (cs.MA) |
| Cite as: | arXiv:2610.12593 [cs.AI] |
| (or arXiv:2610.12593v1 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2610.12593 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Wei Yang [view email]
[v1]
Thu, 8 Oct 2026 18:17:20 UTC (5,980 KB)
— Originally published at arxiv.org
Want this in your inbox every morning?
Daily brief at your local 8am — bilingual EN/中文, free.
More from arXiv cs.AI
See more →VEPA: Verified Evolving Formal Policy for Lifelong Agent Guardrails
VEPA introduces a framework for verified evolving formal policies for LLM agents, enhancing safety by adapting guardrails based on deployment feedback. It outperforms existing methods across benchmarks like ASSEBench-Safety and ASSEBench-Security, ensuring deterministic policy enforcement even under challenging conditions.