LLM Multi-Agent Cascades for Cost-Effective ML Task Solving

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing Large Language Model (LLM) agent systems are limited in flexibility and cost-effective solutions for complex machine learning tasks, often relying on expensive models like GPT-4 and lacking the ability to explore out-of-distribution solutions, while multi-agent systems have not been effectively utilized for nuanced problem-solving in real-world engineering challenges.

Innovation Solution

A multi-agent system using a combination of less expensive LLMs, such as Gemini-Pro, with cascaded invocation of stronger models like GPT-4, employs structured prompts and historical context retrieval to enhance decision-making, allowing cost-effective and scalable solutions for machine learning tasks.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If single-agent systems use larger and expensive models such as GPT-4, then performance on complex tasks improves, but cost increases significantly

Engineering Contradiction:
Improveperformance on complex tasksVSAvoidcost
Core Design Contradiction:
ReliabilityVSLoss of energy

Solution Approach 1:

The system divides the complex problem-solving process into multiple specialized agents (planner agent, coder agent, evaluator agent) rather than relying on a single powerful model. Each agent handles specific aspects of the task, reducing the need for any single agent to have excessive capabilities while maintaining overall system performance.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent implements a hierarchical structure where a planner agent generates high-level plans that are then executed by coder agents, which in turn may invoke smaller language models for specific sub-tasks. This nested arrangement allows the system to leverage multiple levels of reasoning and execution, achieving complex task performance without requiring every component to be a large expensive model.

Inventive Principle:
Principle #7Nested doll (Nesting)

2Adaptability or versatility

If single-agent systems rely on larger models, then capability to handle nuanced tasks improves, but cost increases

Engineering Contradiction:
Improvecapability to handle nuanced tasksVSAvoidcost
Core Design Contradiction:
Adaptability or versatilityVSLoss of energy

Solution Approach 1:

Different agents in the system are equipped with different capabilities appropriate to their specific roles. The planner agent focuses on high-level reasoning and strategy, the coder agent specializes in code generation, and the evaluator agent handles assessment. This specialization allows each agent to be optimized for its specific function rather than requiring all agents to have universal expertise, reducing overall cost while maintaining versatility.

Inventive Principle:
Principle #3Local quality

Solution Approach 2:

The planner agent acts as an intermediary that translates high-level task descriptions into detailed action plans that can be executed by coder agents. This intermediary layer enables the system to handle nuanced tasks by breaking down complex requirements into manageable steps, allowing smaller specialized agents to achieve what would otherwise require a single large model.

Inventive Principle:
Principle #24Intermediary (Mediator)

3Extent of automation

If AutoML-based approaches are used, then automation of ML tasks improves, but flexibility decreases due to predefined constraints

Engineering Contradiction:
Improveautomation of ML tasksVSAvoidflexibility
Core Design Contradiction:
Extent of automationVSAdaptability or versatility

Solution Approach 1:

The system employs dynamic plan generation and execution where the planner agent can adapt its approach based on the specific task requirements and feedback from evaluator agents. Rather than following rigid predefined workflows, the system dynamically adjusts its reasoning and execution strategies, allowing it to maintain high automation while remaining flexible to diverse ML tasks including data preprocessing, model selection, training, and evaluation.

Inventive Principle:
Principle #15Dynamics

Data Source

PatentEP4697233A1Cost effective multi-agent systems and methods for solving machine learning tasks using large language models
Publication Date: 2026.02.18 TATA CONSULTANCY SERVICES LTD
  • EP4697233A1 patent drawingFigure 1
  • EP4697233A1 patent drawingFigure 2
  • EP4697233A1 patent drawing

AI summary

Large Language Models (LLMs) excel in diverse applications including generation of code snippets but often struggle with generating code for complex Machine Learning (ML) tasks. Although existing LLM single agent-based approaches give varying performance depending on the task complexity, they purely rely on larger and expensive models such as Generative Pre-Trained Transformers (e.g., GPT-4). Present disclosure implements a LLM multi-agent-based system which leverages combination of experts using profiling, efficient retrieval of past observations, LLM cascades, and ask-the-expert calls to enable obtain task specification, actions to be performed along with associated justification and reasoning based on which appropriate planner(s) invoke respective agents for execution of a machine learning task.