Negative Feedback Learning for LLM Network Troubleshooting Agents

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Large language models (LLMs) face challenges in network monitoring and control due to delays in responding to queries and potential mistakes, which hinder their use in self-healing networks.

Innovation Solution

A device uses negative feedback learning to adjust prompts for LLM-based agents, allowing them to perform tasks without significant delays and improve accuracy by learning from successful troubleshooting trajectories.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If LLM-based agents are used for network troubleshooting tasks, then the ability to perform complex tasks is improved, but response time increases significantly

Engineering Contradiction:

Improvetask performance capabilityVSAvoidresponse time

Core Design Contradiction:

Adaptability or versatilityVSLoss of time

Solution Approach 1:

The system pre-generates multiple possible responses and trajectories before the user sees the final answer. LLMs generate several potential troubleshooting paths in advance, then select the best one, reducing the perceived response time while maintaining complex task capability.

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The troubleshooting process is divided into multiple independent trajectories or paths. Each trajectory represents a possible sequence of actions the LLM can take, allowing parallel generation and evaluation of different approaches to the same problem.

Inventive Principle:

Principle #1Segmentation

2Extent of automation

If LLM-based agents are used for network monitoring, then automation capability is improved, but mistake rate increases

Engineering Contradiction:

Improveautomation capabilityVSAvoidmistake rate

Core Design Contradiction:

Extent of automationVSReliability

Solution Approach 1:

The system implements feedback loops where generated trajectories are evaluated for correctness. Failed trajectories provide negative feedback that helps identify and correct mistakes, improving the reliability of automated decisions over time while maintaining high automation levels.

Inventive Principle:

Principle #23Feedback

Solution Approach 2:

Mistakes and failures are converted into valuable training data. The system learns from erroneous trajectories by analyzing what went wrong and adjusting future generations, turning reliability problems into improvement opportunities.

Inventive Principle:

Principle #22Blessing in disguise (Convert harm into benefit)

3Manufacturing precision

If multiple troubleshooting trajectories are generated, then task completion accuracy is improved, but computational complexity increases

Engineering Contradiction:

Improvetask completion accuracyVSAvoidcomputational complexity

Core Design Contradiction:

Manufacturing precisionVSDevice complexity

Solution Approach 1:

The system generates a limited number of trajectories that is sufficient for achieving good accuracy without overwhelming computational resources. Rather than exhaustively exploring all possible paths, it generates enough diverse trajectories to ensure high probability of finding correct solutions.

Inventive Principle:

Principle #16Partial or excessive action

Data Source

PatentUS20250278575A1Using negative feedback learning on a language model-based network troubleshooting agent

Publication Date: 2025.09.04 CISCO TECHNOLOGY INC

AI summary

In one implementation, a device obtains an indication of a failure by a language model-based agent for a computer network to perform a first task requested by a first prompt. The device determines a feedback metric that quantifies how critical the failure is. The device identifies a subsequent prompt for the language model-based agent to perform a new task of a similar type as the first task. The device adjusts, based on the feedback metric, the subsequent prompt to avoid the language model-based agent failing the new task.