Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

CRMWeaver: Building Powerful Business Agent via Agentic RL and Shared Memories

Yilong Lai, Yipin Yang, Jialong Wu, Fengran Mo, Zhenglin Wang, Ting Liang, Jianguo Lin, Keping Yang

TL;DR

The paper tackles the challenge of building robust business agents capable of operating over complex, multi-relational enterprise data. It proposes CRMWeaver, a framework that combines synthetic data generation, a two-stage training pipeline (SFT followed by RL with DAPO), and a long-term memory module to reuse solution guidelines. Experiments on the CRMArena-Pro benchmark show that a lightweight Qwen-3-4B model achieves competitive B2B and B2C performance, with memory augmentation improving reliability and generalization. The approach reduces reliance on very large models while delivering practical capabilities for real-world business scenarios.

Abstract

Recent years have witnessed the rapid development of LLM-based agents, which shed light on using language agents to solve complex real-world problems. A prominent application lies in business agents, which interact with databases and internal knowledge bases via tool calls to fulfill diverse user requirements. However, this domain is characterized by intricate data relationships and a wide range of heterogeneous tasks, from statistical data queries to knowledge-based question-answering. To address these challenges, we propose CRMWeaver, a novel approach that enhances business agents in such complex settings. To acclimate the agentic model to intricate business environments, we employ a synthesis data generation and RL-based paradigm during training, which significantly improves the model's ability to handle complex data and varied tasks. During inference, a shared memories mechanism is introduced, prompting the agent to learn from task guidelines in similar problems, thereby further boosting its effectiveness and generalization, especially in unseen scenarios. We validate the efficacy of our approach on the CRMArena-Pro dataset, where our lightweight model achieves competitive results in both B2B and B2C business scenarios, underscoring its practical value for real-world applications.

CRMWeaver: Building Powerful Business Agent via Agentic RL and Shared Memories

TL;DR

The paper tackles the challenge of building robust business agents capable of operating over complex, multi-relational enterprise data. It proposes CRMWeaver, a framework that combines synthetic data generation, a two-stage training pipeline (SFT followed by RL with DAPO), and a long-term memory module to reuse solution guidelines. Experiments on the CRMArena-Pro benchmark show that a lightweight Qwen-3-4B model achieves competitive B2B and B2C performance, with memory augmentation improving reliability and generalization. The approach reduces reliance on very large models while delivering practical capabilities for real-world business scenarios.

Abstract

Recent years have witnessed the rapid development of LLM-based agents, which shed light on using language agents to solve complex real-world problems. A prominent application lies in business agents, which interact with databases and internal knowledge bases via tool calls to fulfill diverse user requirements. However, this domain is characterized by intricate data relationships and a wide range of heterogeneous tasks, from statistical data queries to knowledge-based question-answering. To address these challenges, we propose CRMWeaver, a novel approach that enhances business agents in such complex settings. To acclimate the agentic model to intricate business environments, we employ a synthesis data generation and RL-based paradigm during training, which significantly improves the model's ability to handle complex data and varied tasks. During inference, a shared memories mechanism is introduced, prompting the agent to learn from task guidelines in similar problems, thereby further boosting its effectiveness and generalization, especially in unseen scenarios. We validate the efficacy of our approach on the CRMArena-Pro dataset, where our lightweight model achieves competitive results in both B2B and B2C business scenarios, underscoring its practical value for real-world applications.

Paper Structure

This paper contains 23 sections, 8 equations, 3 figures, 6 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Preliminaries
  3. Methodology
  4. Synthesis Data Construction
  5. Agent Setup of CRMWeaver
  6. Two-stage Training Recipe
  7. SFT for Agentic Model Cold Starts
  8. Reinforcement Learning to Generalization
  9. Improving Reasoning with Long-term Memory
  10. Experiments
  11. Experiments Setup
  12. Main Experiments
  13. Impact of Reinforcement Learning with Agents
  14. The Frequency of Updates to Long-Term Memory
  15. Related Works
  16. ...and 8 more sections

Figures (3)

  • Figure 1: Challenges of business agents in handling diverse user needs and interacting with heterogeneous data sources.
  • Figure 2: The overall pipeline of CRMWeaver. We first construct both simple and complex synthetic data by leveraging correlations among data records to support model training (§ \ref{['sec:data']}). The training recipe adopts a two-stage paradigm, including SFT for model initialization and reinforcement learning for generalization (§ \ref{['sec:training']}). Finally, the inference process incorporates a memory enhancement mechanism (§ \ref{['sec:memory']}).
  • Figure 3: Performance comparison with SFT and RL on CRMArena-Pro across different tasks.