Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A Human-Computer Collaborative Tool for Training a Single Large Language Model Agent into a Network through Few Examples

Lihang Pan, Yuxuan Li, Chun Yu, Yuanchun Shi

TL;DR

This paper introduces EasyLAN, a human–computer collaborative tool that automatically evolves a single LLM agent into a network of agents (LAN) using a few training examples. The system decouples task knowledge (W-knowledge) across agents and enables human supervision to guide automated updates, yielding faster LAN construction and higher task performance. It formalizes an agent architecture with Input, Control, Execution, and Output modules, and delineates update strategies to address error causes while preserving prior training inputs. An empirical user study shows substantial reductions in interaction time and improvements in LAN effectiveness compared with a baseline, demonstrating the practical potential of computer-executed LAN updates under user supervision. The work highlights both the promise and current limits of LAN design, pointing to future directions in automatic data generation, recursion handling, and runtime monitoring to further enhance scalable multi-agent reasoning with LLMs.

Abstract

The capabilities of a single large language model (LLM) agent for solving a complex task are limited. Connecting multiple LLM agents to a network can effectively improve overall performance. However, building an LLM agent network (LAN) requires a substantial amount of time and effort. In this paper, we introduce EasyLAN, a human-computer collaborative tool that helps developers construct LANs. EasyLAN initially generates a LAN containing only one agent based on the description of the desired task. Subsequently, EasyLAN leverages a few training examples to update the LAN. For each example, EasyLAN models the gap between the output and the ground truth and identifies the causes of the errors. These errors are addressed through carefully designed strategies. Users can intervene in EasyLAN's workflow or directly modify the LAN. Eventually, the LAN evolves from a single agent to a network of LLM agents. The experimental results indicate that developers can rapidly construct LANs with good performance.

A Human-Computer Collaborative Tool for Training a Single Large Language Model Agent into a Network through Few Examples

TL;DR

This paper introduces EasyLAN, a human–computer collaborative tool that automatically evolves a single LLM agent into a network of agents (LAN) using a few training examples. The system decouples task knowledge (W-knowledge) across agents and enables human supervision to guide automated updates, yielding faster LAN construction and higher task performance. It formalizes an agent architecture with Input, Control, Execution, and Output modules, and delineates update strategies to address error causes while preserving prior training inputs. An empirical user study shows substantial reductions in interaction time and improvements in LAN effectiveness compared with a baseline, demonstrating the practical potential of computer-executed LAN updates under user supervision. The work highlights both the promise and current limits of LAN design, pointing to future directions in automatic data generation, recursion handling, and runtime monitoring to further enhance scalable multi-agent reasoning with LLMs.

Abstract

The capabilities of a single large language model (LLM) agent for solving a complex task are limited. Connecting multiple LLM agents to a network can effectively improve overall performance. However, building an LLM agent network (LAN) requires a substantial amount of time and effort. In this paper, we introduce EasyLAN, a human-computer collaborative tool that helps developers construct LANs. EasyLAN initially generates a LAN containing only one agent based on the description of the desired task. Subsequently, EasyLAN leverages a few training examples to update the LAN. For each example, EasyLAN models the gap between the output and the ground truth and identifies the causes of the errors. These errors are addressed through carefully designed strategies. Users can intervene in EasyLAN's workflow or directly modify the LAN. Eventually, the LAN evolves from a single agent to a network of LLM agents. The experimental results indicate that developers can rapidly construct LANs with good performance.
Paper Structure (37 sections, 7 figures, 7 tables)

This paper contains 37 sections, 7 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Decomposing Complex Tasks and Connecting Multiple LLM Agents
  4. Using Examples in LLM Applications
  5. Human-Computer Collaborative Tools
  6. LLM Agent Network (LAN) Design
  7. Input Module (IM) & Output Module (OM)
  8. Control Module (CM)
  9. Execution Module (EM)
  10. System Design
  11. Automated Update of LAN
  12. Initialization of the LAN
  13. Pipeline to update a LAN
  14. Details of the update strategies
  15. Ensuring the accuracy of previous training inputs
  16. ...and 22 more sections

Figures (7)

  • Figure 1: How EasyLAN trains a task-oriented LLM agent network (LAN) from a single LLM agent. (1) EasyLAN auto-generates an initial LAN that only contains a single LLM agent based on the task (e.g., translating French to English). A significant gap exists between the capabilities of the initial LAN and the task requirements. (2) A training example consists of an input and a ground truth. For a given training example, EasyLAN identifies discrepancies between the LAN's output and the expected output. For instance, when the input is a line of French poetry, "Vienne la nuit sonne l'heure, les jours s'en vont je demeure", the LAN fails to translate the text accurately while preserving the original rhyming scheme. (3) EasyLAN identifies the cause of the discrepancies and updates the LAN with respect to both the network architecture (e.g., splitting Translator into Literal Translator and Rhyming Polisher) and agent contents (e.g., adjusting the functionality of an agent). (n) EasyLAN iterates over a small set of training examples and constructs a satisfactory LAN.
  • Figure 2: An overview of an agent in a LAN. (a) The modules inside an agent. (b) an example LAN. "Literal Translator" and "Rhyming Polisher" are agents in the LAN. (c) Details of the "Rhyming Polisher" in (b). It receives the original French text and the output from the "Literal Translator" and computes a rhyming translation result."
  • Figure 3: The pipeline for updating the LAN. The light blue rectangles indicate update strategies.
  • Figure 4: An example of how EasyLAN updates a LAN. (0) EasyLAN decomposes the Translator into Literal Translator and Rhyming Polisher to ensure that the translation output can rhyme when necessary. (1) EasyLAN adds a Structure Refiner to adjust the syntactic structure of sentences (e.g., converting passive voice to active voice). (2) EasyLAN adds knowledge to the CM of the Rhyming Polisher to prevent unnecessary rhyming. (3) EasyLAN adds knowledge to the CM of the Rhyming Polisher to better identify whether the input sentence is rhyming. (4) EasyLAN adds knowledge to the EM of the Literal Translator to improve its capability to translate idioms. (5) EasyLAN splits the Literal Translator into Spoken Text Translator and Literary Text Translator to better cater to diverse translation needs. (6) EasyLAN adds a connection from the Structure Refiner to the Rhyming Polisher to ensure that the output from the Rhyming Polisher also adheres to the sentence structure defined by the Structure Refiner.
  • Figure 5: The user interface. Region 1 allows users to inspect and modify the LAN structure. Agents can be selected by clicking the pink rectangles. Region 2 facilitates inspection and editing of the selected agent's properties. In Region 3, users can provide training examples to EasyLAN. Region 4 offers insights into and intervention options for EasyLAN's automated LAN update workflow.
  • ...and 2 more figures