Dynamic LLM Evaluation Metrics for Bias and Robustness Gaps

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing Large Language Models (LLMs) face challenges in contextual understanding, bias, and robustness, necessitating improved validation methods to ensure accurate, coherent, and unbiased responses.

Innovation Solution

A dynamic weighted metrics-based evaluation system that includes a contextual task analysis module, machine learning model training, and decision tree-based root cause analysis to assess and improve LLM responses, generating compliance certificates and NFTs using a smart contract blockchain.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If LLMs are trained on vast amounts of text data to improve response generation capabilities, then the model's language understanding and generation improve, but biases and contextual understanding issues worsen due to training data quality and composition

Engineering Contradiction:
Improveresponse generation capabilityVSAvoidcontextual understanding and bias
Core Design Contradiction:
ProductivityVSReliability

Solution Approach 1:

The system performs preliminary evaluation of LLM responses using multiple metrics before final output. Evaluation metrics are computed on generated responses to assess quality, relevance, and bias before the response is finalized, allowing pre-correction of problematic outputs

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The system implements feedback loops where LLM responses are evaluated using computed metrics, and this evaluation feedback is used to identify weaknesses and guide model improvements. The evaluation results feed back into the training process to reduce biases and improve contextual understanding

Inventive Principle:
Principle #23Feedback

2Adaptability or versatility

If LLMs are designed to be highly capable and general-purpose, then versatility improves, but vulnerability to adversarial attacks and robustness issues worsen

Engineering Contradiction:
Improvemodel capabilityVSAvoidadversarial attack vulnerability
Core Design Contradiction:
Adaptability or versatilityVSObject-affected harmful factors

Solution Approach 1:

The system applies preliminary protective actions by evaluating responses for adversarial vulnerabilities before output. Multiple evaluation metrics detect potential adversarial attacks and harmful content in advance, allowing the system to block or correct problematic responses before they reach users

Inventive Principle:
Principle #9Preliminary anti-action

Solution Approach 2:

The evaluation system acts as an intermediary between the LLM and end users. It mediates the response generation process by filtering and assessing outputs, providing a protective layer that reduces adversarial vulnerability while preserving model versatility

Inventive Principle:
Principle #24Intermediary (Mediator)

3Measurement precision

If comprehensive evaluation metrics are applied to assess LLM responses, then response quality assessment improves, but computational complexity and evaluation time worsen

Engineering Contradiction:
Improveresponse quality assessmentVSAvoidevaluation system complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The evaluation system is segmented into multiple independent metrics that can be computed separately and independently. Each metric assesses a specific aspect of response quality (e.g., relevance, bias, coherence), allowing modular evaluation that improves precision while managing complexity through division of labor

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The system computes multiple evaluation metrics, potentially more than strictly necessary (excessive action), to ensure comprehensive quality assessment. This approach prioritizes thorough evaluation over minimal complexity, accepting higher computational costs in exchange for more precise and reliable quality measurements

Inventive Principle:
Principle #16Partial or excessive action

Data Source

PatentEP4668131A1Method and system for dynamic weighted metrics-based evaluation and tokenization of large language models
Publication Date: 2025.12.24 TATA CONSULTANCY SERVICES LTD
  • EP4668131A1 patent drawingFigure 1
  • EP4668131A1 patent drawingFigure 2
  • EP4668131A1 patent drawingFigure 3A

AI summary

The embodiments of the present disclosure herein address unresolved problems of evaluation of LLM response quality and overall LLM models. Existing approaches for LLM evaluation and LLM response evaluation can be broadly categorized into automatic evaluation metrics, human evaluation, and adversarial testing. Embodiments herein provides a method and system for dynamically weighted selection of performance metrics for generation of LLM response score. Further, the system is configured method and system for generation of LLM maturity gap analysis and associated recommendation for improvement of LLM response score. Finally, the system generates a compliance certificate for every model (version) with a (threshold) level score and generates an NFT using a smart contract based blockchain, using metadata associated with the model and the evaluation metrics and results.