PaperlayerView exact revision on arXiv

Source-linked AI summary

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

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

arXiv:2510.25333v2cs.CL

TL;DR

Business agents face intricate data relationships and heterogeneous tasks, while satisfactory performance has depended on costly models. CRMWeaver combines synthetic-data training, reinforcement learning, and inference-time shared memories. On CRMArena-Pro, its lightweight model achieves performance comparable to larger models across B2B and B2C scenarios.

  • Problem

    Business agents must handle intricate enterprise-data relationships and heterogeneous tasks, from statistical queries to knowledge-based question answering, in settings where satisfactory performance has depended on costly models.

    Source: §1 · p. 1

  • Method

    CRMWeaver combines synthesized training data and two-stage supervised initialization plus DAPO-based reinforcement learning with inference-time retrieval of guidelines from successful similar-task trajectories.

    Source: §1 · p. 1, §7 · pp. 6–7

  • Results

    A Qwen3-4B model achieves performance comparable to Qwen3-235B-Instruct and Gemini 2.5-Pro on CRMArena-Pro across B2B and B2C scenarios.

    Source: §1 · p. 2

  • Takeaways & Limitations

    CRMWeaver is presented as a practical, cost-efficient approach for deploying reliable business agents across diverse and evolving real-world tasks.

    Source: §7 · p. 7

    The reinforcement-learning setup focuses on single user queries with multiple tool turns, and rollout constraints restrict context length and tool-call count.

    Source: Limitation · p. 7

Abstract

from arXiv · show

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.

1 Introduction

Business agents must handle intricate enterprise data and heterogeneous tasks, while satisfactory performance has depended on costly models. CRMWeaver combines synthetic-data training, reinforcement learning, and shared memories; its lightweight model is reported comparable to larger models on CRMArena-Pro across B2B and B2C scenarios.

1 Introduction

  • Business agents query enterprise databases and internal knowledge bases, facing intricate data dependencies and tasks ranging from statistical queries to knowledge-based question answering.The introduction notes that satisfactory performance in these settings had been achieved only by proficient but prohibitively costly models.Source: §1 · p. 1
  • CRMWeaver combines synthesized complex training data with rejection sampling, trajectory distillation, and DAPO-based reinforcement learning.The two-stage training pipeline first bootstraps multi-turn query solving, then applies reinforcement learning to improve generalization in complex business environments.Source: §1 · pp. 1–2
  • At inference, CRMWeaver retrieves guidelines from successful trajectories for similar tasks and injects them into context, supporting stability across tasks, especially unseen ones.Source: §1 · p. 2
  • A Qwen3-4B model achieves performance comparable to Qwen3-235B-Instruct and Gemini 2.5-Pro on CRMArena-Pro across B2B and B2C scenarios.Source: §1 · p. 2

2 Preliminaries

The paper situates its work in Salesforce-based business-agent benchmarks and formalizes task solving as agent–environment interaction. The setup also trains on CRMArena and evaluates on the distinct CRMArena-Pro environment.

2 Preliminaries

  • CRMArena models Salesforce customer service with 16 interconnected objects and nine tasks; CRMArena-Pro expands to 25 objects and 19 tasks across Service, Sales, and Configure-Price-Quote.Source: §2 · p. 2
  • Training on CRMArena and evaluation on CRMArena-Pro use entirely distinct environments to assess generalizability while mitigating data leakage.Source: §2 · p. 2
  • Task solving is modeled as a partially observable Markov decision process, with Execute, Date Calculation, and Answer as the available tools.Execute queries records via SQL or SOSL; Date Calculation supports temporal reasoning, and Answer submits the final response.Source: §2 · p. 2

3 Methodology

CRMWeaver combines synthetic-data construction and two-stage training with inference-time long-term memory to address complex and unseen business tasks. Its methodology spans data synthesis, ReAct-based agent setup, supervised fine-tuning, reinforcement learning, and memory-guided inference.

3.1 Synthesis Data Construction

  • A record graph and random walks provide relational paths that an LLM expands into synthetic queries, supporting reasoning across interconnected tables.The graph represents records as nodes and inter-table relationships as edges; each query is built iteratively from adjacent record pairs.Source: §3.1 · pp. 2–3
  • The training corpus combines simple and task-specific examples with complex multi-relational queries, statistical analysis, and knowledge-base question answering.Simple data builds basic business QA and analysis skills, while task-specific data covers nine CRMArena tasks.Source: §3.1 · p. 3

3.2 Agent Setup of CRMWeaver

  • CRMWeaver uses ReAct cycles in which the model selects thoughts and actions from interaction history, using tools to query business data, calculate dates, and submit answers.The action space includes execute, date calculation, and answer tools.Source: §3.2 · pp. 3–4

3.3.1 SFT for Agentic Model Cold Starts

  • GPT-4.1 generates candidate execution trajectories, which are filtered for answer correctness and adequate execution before supervised fine-tuning.Filtering removes incorrect final answers and very short trajectories for samples whose answer is “None.”Source: §3.3.1 · p. 4

3.3.2 Reinforcement Learning to Generalization

  • Reinforcement learning rolls out trajectory groups and uses DAPO with rule-based rewards to refine the agent through multi-turn environment interactions.Rewards include format validity, exact-match answers for data analysis, and F1 for QA tasks.Source: §3.3.2 · p. 4, §3.2 · p. 4

3.4 Improving Reasoning with Long-term Memory

  • Long-term memory retrieves relevant workflow guidelines for analogous queries and consolidates new guidelines from resolved queries when no sufficiently similar precedent exists.A verification step determines whether a generated guideline is added to the indexed memory pool.Source: §3.4 · pp. 4–5

4 Experiments

CRMWeaver is evaluated on CRMArena-Pro in B2B and B2C scenarios. It achieves competitive average scores and outperforms all baselines on database-related tasks.

4.1 Experiments Setup

  • CRMArena-Pro evaluation uses 3.8k single-turn test samples across 19 tasks, with F1 for knowledge QA and exact match for other tasks.The tasks span database querying and numerical computation, information retrieval and textual reasoning, workflow execution, and policy compliance.Source: §4.1 · p. 5

4.2 Main Experiments

  • The assessment compares CRMWeaver with baseline models on CRMArena-Pro across B2B and B2C domains.The benchmark contains 19 distinct subtasks.Source: §4.2 · p. 5
  • CRMWeaver achieves average scores of 55.6 in B2B and 57.1 in B2C compared with more advanced LLMs, and outperforms all baselines on database-related tasks, scoring 73.0 (B2B) and 72.8 (B2C).Source: §4.2 · pp. 5–6

5 Analysis

Ablations show large gains over the backbone and performance decreases when any training stage is removed. Memory updates are infrequent, while indexed memories are used across a broad range of tasks.

5.1 Ablation Studies for CRMWeaver

  • 179.4% (from 19.9 to 55.6) B2B and 181.2% (from 20.3 to 57.1) B2C average-score gains are reported over the backbone.The comparison is on CRMArena-Pro.Source: §5.1 · p. 6
  • Removing either training stage decreases performance, indicating that each stage contributes to the overall results.Source: §5.1 · p. 6

5.2 The Frequency of Updates to Long-Term Memory

  • Indexed memories are used across a broad range of tasks even though Long-term Memory updates are relatively infrequent.Source: §5.2 · p. 6
  • About 3% of B2B tasks trigger memory updates; a few-shot LLM classifier prevents excessive updates for knowledge-QA tasks.Low question similarity can otherwise make indexed-memory access less likely for some tasks, such as Knowledge QA.Source: §5.2 · p. 6

6 Related Works

Prior work on business agents progresses from synthetic tasks toward realistic enterprise environments, while long-term memory research explores varied representations and agentic reinforcement learning extends tool use and interaction.

Business Agents and Benchmarks

  • Business-agent benchmarks progressed from simplified webpages to enterprise-platform tasks, with CRMArena modeling expert-verified Salesforce CRM tasks across interconnected objects.CRMArena-Pro further expands scale and task diversity, described as the most comprehensive business-agent evaluation to date.Source: §6 · p. 6

Long-term Memory of Language Agent

Agentic Reinforcement Learning

  • Agentic RL research includes rule-based models such as Deepseek-R1 (Guo et al., 2025), tool-augmented reasoning such as Search-R1 (Jin et al., 2025) and ReTool (Feng et al., 2025), and web-interaction systems using data synthesis and end-to-end RL (Wu et al., 2025a; Li et al., 2025; Tao et al., 2025).Source: §6 · p. 6

7 Conclusion

CRMWeaver combines synthesized training data, two-stage learning, and inference-time long-term memory for business agents. The paper presents this combination as a practical, cost-efficient approach for reliable agents across diverse and evolving real-world tasks.

7 Conclusion

  • CRMWeaver combines synthesized training data, supervised initialization, DAPO-based reinforcement learning, and inference-time long-term memory for knowledge reuse.Source: §7 · pp. 6–7
  • The combined approach is presented as practical and cost-efficient for deploying reliable business agents across diverse and evolving real-world tasks.Source: §7 · p. 7

Limitation

The framework is limited to single-query user interactions, and its memory module pairs online retrieval with offline consolidation to build task guidelines.

Limitation

  • The framework has not been extended to multi-turn user interactions, where users provide a sequence of queries.Its training setup supports one user query with multiple tool-call and observation turns.Source: Limitation · p. 7

A Long-term Memory Details

  • Algorithm 1 pairs online memory-guided inference, which retrieves a top-1 memory candidate, with offline memory consolidation, where BUILDMEMORY obtains a memory unit E.Source: A Long-term Memory Details · p. 10

A.1 Memory Indexing

  • The memory guideline is generated from trajectories using o3-mini for reasoning and GPT-4.1 with a prompt for guideline construction.The prompt is described in Table 4.Source: §A.1 · p. 10

A.2 Evaluation Protocol for Memory Module

Evaluation permits within-run memory updates under a continual-deployment setting, while randomizing query order and averaging results across three runs.

A.2 Evaluation Protocol for Memory Module

  • Memory updates are permitted during each evaluation run, reflecting a continual-deployment setting rather than a strictly frozen memory.The updates follow the offline consolidation procedure.Source: §A.2 · p. 10
  • Test queries are randomly permuted and evaluated over three independent runs, with performance reported as their average.Updates occur infrequently, about 3% on average in B2B, and results remain stable across random orderings.Source: §A.2 · p. 10

B Implementation Details

The implementation specifies RL and inference settings, while CRMArena-Pro provides coverage across business scenarios and four task categories.

B Implementation Details

  • RL uses rollout size G = 16, εlow = 0.2, and εhigh = 0.28; inference uses temperature 0.2 and at most 20 interaction turns.Experiments use SQLite for SQL, the Salesforce API for SOSL, and two nodes with 8 Nvidia H20 GPUs each.Source: B Implementation Details · p. 10

C Details of CRMArena-Pro

  • CRMArena-Pro covers B2B and B2C scenarios and four business-skill categories: data, text, workflow execution, and policy compliance.The categories assess database querying and numerical computation, information retrieval and textual reasoning, workflow execution, and policy compliance.Source: C Details of CRMArena-Pro · pp. 10–11
  • Table 3 provides detailed dataset statistics for CRMArena-Pro.The supplied passage does not specify the table’s individual values.Source: C Details of CRMArena-Pro · p. 10, Table 3 · p. 11

D Synthesis Data Construction

The appendices provide prompts and examples spanning complex-query synthesis, memory-guideline construction, and agent trajectories.

D.1 Prompt for Complex Queries Generation

  • Tables 6 and 7 contain prompts for generating complex queries.The section introduces examples of complex synthesis data.Source: §D.1 · p. 11, §D.2 · p. 11

D.2 Complex Synthesis Data Examples

  • The complex-query examples include a multi-clue transcript search and a product-identification task grounded in linked business events.One example asks for a chat transcript ID; another asks for a product’s unique catalog reference.Source: §D.2 · p. 11

E.1 Long-term Memory Example

  • Table 5 illustrates a memory unit whose indexed query key is paired with a corresponding guideline value.The caption identifies the key as the query used to match similar queries.Source: §E.1 · p. 11, Table 5 · p. 12

E.2 Full Trajectories Example

  • The trajectory example shows an agent following lead-routing criteria through territory lookup and subsequent tool interactions.The example includes a database query and its returned territory record.Source: §E.2 · pp. 11, 15
Loading 2510.25333v2…