Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

WISE-Flow: Workflow-Induced Structured Experience for Self-Evolving Conversational Service Agents

Yuqing Zhou, Zhuoer Wang, Jie Yuan, Hong Wang, Samson Koelle, Ziwei Zhu, Wei Niu

TL;DR

The paper addresses the reliability and adaptability of LLM-based, tool-augmented agents in user-facing services by introducing WISE-Flow, a workflow-induced structured experience framework. It induces grounded workflows with explicit prerequisites from multi-source interaction logs and uses retrieval at both the workflow and action levels to guide deployment, aided by a three-pass offline induction and progress-aligned online control. The approach introduces a new workflow representation and an evaluation metric $F_eta$ to jointly capture milestone coverage and error control, reporting consistent performance gains on ToolSandbox and τ^2-bench across base models. The findings demonstrate that environment feedback, structured experience representations, and trajectory-aggregated workflow induction substantially improve one-shot success, reduce errors, and enable cross-task transfer, underscoring the practical impact for scalable, self-evolving conversational services.

Abstract

Large language model (LLM)-based agents are widely deployed in user-facing services but remain error-prone in new tasks, tend to repeat the same failure patterns, and show substantial run-to-run variability. Fixing failures via environment-specific training or manual patching is costly and hard to scale. To enable self-evolving agents in user-facing service environments, we propose WISE-Flow, a workflow-centric framework that converts historical service interactions into reusable procedural experience by inducing workflows with prerequisite-augmented action blocks. At deployment, WISE-Flow aligns the agent's execution trajectory to retrieved workflows and performs prerequisite-aware feasibility reasoning to achieve state-grounded next actions. Experiments on ToolSandbox and $τ^2$-bench show consistent improvement across base models.

WISE-Flow: Workflow-Induced Structured Experience for Self-Evolving Conversational Service Agents

TL;DR

The paper addresses the reliability and adaptability of LLM-based, tool-augmented agents in user-facing services by introducing WISE-Flow, a workflow-induced structured experience framework. It induces grounded workflows with explicit prerequisites from multi-source interaction logs and uses retrieval at both the workflow and action levels to guide deployment, aided by a three-pass offline induction and progress-aligned online control. The approach introduces a new workflow representation and an evaluation metric to jointly capture milestone coverage and error control, reporting consistent performance gains on ToolSandbox and τ^2-bench across base models. The findings demonstrate that environment feedback, structured experience representations, and trajectory-aggregated workflow induction substantially improve one-shot success, reduce errors, and enable cross-task transfer, underscoring the practical impact for scalable, self-evolving conversational services.

Abstract

Large language model (LLM)-based agents are widely deployed in user-facing services but remain error-prone in new tasks, tend to repeat the same failure patterns, and show substantial run-to-run variability. Fixing failures via environment-specific training or manual patching is costly and hard to scale. To enable self-evolving agents in user-facing service environments, we propose WISE-Flow, a workflow-centric framework that converts historical service interactions into reusable procedural experience by inducing workflows with prerequisite-augmented action blocks. At deployment, WISE-Flow aligns the agent's execution trajectory to retrieved workflows and performs prerequisite-aware feasibility reasoning to achieve state-grounded next actions. Experiments on ToolSandbox and -bench show consistent improvement across base models.
Paper Structure (36 sections, 2 equations, 9 figures, 8 tables)

This paper contains 36 sections, 2 equations, 9 figures, 8 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Tool-augmented conversational agents and benchmarks.
  4. Self-evolving agents and memory-augmented experience reuse.
  5. Procedural experience.
  6. Methodology
  7. Multi-channel Interaction Logging
  8. Offline Workflow Induction
  9. Online Workflow-guided Agent Control
  10. Experimental Setup
  11. Benchmarks and Environments
  12. Baselines
  13. Models
  14. Evaluation Metric
  15. Results and Analysis
  16. ...and 21 more sections

Figures (9)

  • Figure 1: WISE-Flow pipeline.
  • Figure 2: Proportion of trial-and-error (TE) successes among all successful trials (lower is better). We decompose successful episodes into one-shot successes and successes after TE, and report the proportion of the latter for each method.(Refl. = Reflexion, REM = REMEMBERER, AutoG = AutoGuide, WISE= WISE-Flow)
  • Figure 3: Trajectory-wise vs. task-wise workflow induction on ToolSandbox. Higher Value is better. (We report milestone coverage as $1-\mathrm{MMR}$. Sim. = Similarity.)
  • Figure 4: Prompts of all baselines on ToolSandbox.
  • Figure 5: Evaluation of generalization across tasks on ToolSandbox. We compare the outcome composition of different methods, decomposing trials into one-shot success, trial-and-error success, and failure. WISE-Flow leave-one-task-out (WISE-LOTO) excludes the same-task experience from the retrieval pool.
  • ...and 4 more figures