LLM Response Graph Clustering for Confidence and Variability

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Large Language Models (LLMs) exhibit susceptibility to hallucinations, inconsistent responses, and lack confidence in their predictions, posing a significant challenge for users and organizations relying on their outputs, particularly in scenarios requiring precision and reliability.

Innovation Solution

A method and system for computing variability and confidence scores for LLM responses by constructing graphs based on paraphrase questions, documents, and responses, and performing graph comparisons to determine edge sets and compute scores using cosine similarity, with graph clustering to enhance reliability.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If LLMs are used to generate responses for queries, then the productivity and coverage of natural language processing applications are improved, but the reliability and consistency of responses deteriorate due to hallucinations and variability

Engineering Contradiction:
Improveresponse generation capabilityVSAvoidresponse consistency
Core Design Contradiction:
ProductivityVSReliability

Solution Approach 1:

The system implements feedback by generating multiple responses for the same query and using graph-based comparison to identify consistent patterns. The variability score provides feedback on response reliability, allowing the system to select or regenerate responses based on consistency metrics rather than relying on a single LLM output.

Inventive Principle:
Principle #23Feedback

Solution Approach 2:

Instead of relying on a single LLM response, the system generates multiple responses (excessive action) and uses graph clustering to identify the most consistent subset. This partial approach of selecting only the most reliable responses improves overall reliability while maintaining productivity.

Inventive Principle:
Principle #16Partial or excessive action

2Reliability

If multiple LLM responses are generated and compared using graph construction, then the reliability measurement is improved, but the device complexity and computational requirements increase

Engineering Contradiction:
Improveresponse variability measurementVSAvoidgraph construction complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The system segments the complexity by dividing the response analysis into distinct graph components: query graphs representing input questions, document graphs representing source materials, and response graphs representing LLM outputs. Each graph type handles a specific aspect of the analysis, making the overall system more manageable and interpretable.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

Graph structures serve as intermediaries between raw LLM responses and reliability metrics. Instead of directly comparing text responses, the system transforms them into graph representations where similarity and consistency can be measured through edge comparisons and clustering algorithms, simplifying the complexity of direct text comparison.

Inventive Principle:
Principle #24Intermediary (Mediator)

3Measurement precision

If graph clustering is performed to determine dense regions and compute confidence scores, then the measurement precision of response reliability is improved, but the loss of time and computational resources increases

Engineering Contradiction:
Improveconfidence score accuracyVSAvoidcomputation time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The system performs preliminary graph construction and edge identification before executing the computationally intensive clustering operations. By pre-processing the data into structured graph formats with identified edges and nodes, the system reduces the computational burden during the actual confidence score calculation phase.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The system replaces direct text comparison mechanics with graph-based mathematical operations. Instead of computationally expensive string matching and semantic similarity calculations on raw text, the system uses graph theory operations like edge intersection, node matching, and clustering algorithms that are more efficient and scalable.

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

Data Source

PatentEP4672025A1Computing variability and confidence scores for responses generated by large language models (LLMS)
Publication Date: 2025.12.31 TATA CONSULTANCY SERVICES LTD
  • EP4672025A1 patent drawingFigure 1
  • EP4672025A1 patent drawingFigure 2
  • EP4672025A1 patent drawingFigure 3

AI summary

The rapid proliferation of Large Language Models (LLMs) across diverse organizations, domains, and modalities has revolutionized natural language processing applications. Despite their widespread adoption, a critical challenge persists: the inherent tendency of LLMs to hallucinate, exhibit substantial variability in responses, and often lack confidence in their predictions. Embodiments of the present disclosure provide system and method address the challenges associated with LLMs by identifying and selecting models for which various graphs such as query graph, response graph, and document graph are generated given one or more input queries and one or more documents. Various sets of edges are determined for computing variability score. Further, graph clustering is performed on response graph to compute a confidence score. The present disclosure enhances the reliability of LLM outputs, providing users with more consistent and trustworthy results across various applications.