Entropy-Based LLM Output Selection for Text Fluency Detection

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Machine-generated text often lacks fluency, appearing robotic and unnatural due to high entropy, making it difficult to select the most suitable large language model (LLM) for generating natural language text.

Innovation Solution

An entropy-based technique is employed to compute an entropy score for output texts generated by multiple LLMs, selecting the LLM with low entropy for producing fluent natural language by training an entropy model on fluent training samples.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If LLM generates text based on word frequency and statistical prediction, then text generation capability is achieved, but text fluency and naturalness deteriorate

Engineering Contradiction:

Improvetext generation capabilityVSAvoidtext fluency

Core Design Contradiction:

ProductivityVSReliability

Solution Approach 1:

The patent implements feedback by computing entropy scores for generated text and using this information to select or adjust LLM outputs. The entropy calculation provides feedback about the fluency quality of generated text, enabling the system to iteratively improve text selection by comparing multiple candidates and choosing those with lower entropy scores that indicate more natural language patterns.

Inventive Principle:

Principle #23Feedback

2Reliability

If multiple LLMs are evaluated to select the most fluent output, then text fluency improves, but computational complexity and time increase

Engineering Contradiction:

Improvetext fluencyVSAvoidcomputational complexity

Core Design Contradiction:

ReliabilityVSDevice complexity

Solution Approach 1:

The patent replaces complex human judgment mechanisms with an automated entropy-based evaluation system. Instead of requiring manual assessment of text fluency, the system uses mathematical entropy calculations to objectively measure and compare the naturalness of text generated by multiple LLMs, significantly reducing computational complexity while maintaining evaluation accuracy.

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

3Adaptability or versatility

If LLM relies on training data from various sources, then language coverage improves, but text naturalness deteriorates due to high entropy

Engineering Contradiction:

Improvelanguage coverageVSAvoidtext naturalness

Core Design Contradiction:

Adaptability or versatilityVSReliability

Solution Approach 1:

The patent changes the selection parameter from simple word frequency or probability to entropy-based fluency measurement. By using entropy scores as the selection criterion, the system transforms how LLM outputs are evaluated, prioritizing text with lower entropy that exhibits more natural language patterns over text that merely has high probability based on training data frequency.

Inventive Principle:

Principle #35Parameter changes

Data Source

PatentUS20260064970A1Entropy-based detection of the fluency of machine-generated text

Publication Date: 2026.03.05 MICROSOFT TECHNOLOGY LICENSING LLC

AI summary

An entropy-based technique is used to select a large language model capable of generating fluent natural language text. An entropy model, trained on fluent natural language samples, is used to determine the entropy of a large language model based on an output text generated by the large language model. The entropy of a machine-generated natural language text is used to quantify the amount of information that the large language model holds with respect to the tokens and context of an input text segment. The entropy score of a model is then used to select a large language model capable of generating fluent text or to select the most fluent machine-generated output text produced by a set of large language models.