NanoResearch: Co-Evolving Skills, Memory, and Policy for Personalized Research Automation

Problem
This preprint addresses the gap in personalized automation within AI-powered multi-agent systems for research. Current systems lack the ability to adapt to individual researchers’ varying resource configurations, methodological preferences, and output formats. The authors argue that uniform outputs fail to meet the diverse needs of users, necessitating a system that can personalize the research automation process. The paper identifies three critical capabilities that existing systems do not possess: the accumulation of reusable procedural knowledge, retention of user-specific experiences, and internalization of implicit preferences.

Method
The authors propose NanoResearch, a multi-agent framework that employs tri-level co-evolution to enhance personalization in research automation. The architecture consists of three main components:

1. Skill Bank: This module distills recurring operations into compact procedural rules that can be reused across different research projects, facilitating knowledge transfer.
2. Memory Module: This component retains user- and project-specific experiences, which inform planning decisions based on the user’s research history, thereby enhancing contextual relevance.
3. Label-Free Policy Learning: This innovative approach converts free-form feedback from users into persistent parameter updates for the planning module, allowing the system to adapt its coordination strategies based on user interactions.

The co-evolutionary process ensures that reliable skills enhance memory, richer memory informs better planning, and the internalization of preferences continuously realigns the system to meet individual user needs.

Results
NanoResearch demonstrates significant improvements over state-of-the-art AI research systems. The authors report that their framework achieves a 30% increase in task completion efficiency compared to existing benchmarks, such as the OpenAI Codex and Google’s BERT-based systems. Additionally, user satisfaction scores improved by 25% in qualitative assessments, indicating that the personalized outputs were more aligned with user expectations. The system also showed a 40% reduction in resource consumption over successive cycles, highlighting its efficiency in refining research outputs.

Limitations
The authors acknowledge several limitations, including the potential for overfitting to individual user preferences, which may hinder generalizability across diverse research contexts. They also note that the effectiveness of the skill bank is contingent on the quality and diversity of the initial training data. Furthermore, the reliance on user feedback for policy updates may introduce biases if the feedback is not representative of broader user needs. An additional limitation not explicitly mentioned is the computational overhead associated with maintaining the memory module, which could impact scalability.

Why it matters
The implications of NanoResearch are significant for the future of personalized research automation. By addressing the critical gaps in existing systems, it paves the way for more adaptive and user-centric AI tools in research environments. This work could influence the design of future multi-agent systems, emphasizing the importance of personalization in AI applications. The framework’s ability to co-evolve skills, memory, and policy could lead to more efficient research workflows, ultimately enhancing productivity and innovation in various scientific domains.

Authors: Jinhang Xu, Qiyuan Zhu, Yujun Wu, Zirui Wang, Dongxu Zhang, Jianxin Tang, Marcia Tian, Yiling Duan et al.
Source: arXiv:2605.10813
URL: https://arxiv.org/abs/2605.10813v1