How to Make Large Language Models More Confident in Their Answers

Snowflake's patent details a method to make large language models more trustworthy when extracting information from documents by precisely adjusting their confidence scores using a mix of mathematical techniques.

What it covers

This patent describes a system that improves how reliably large language models (LLMs) report their confidence when processing electronic text documents. First, it receives results from an ML model, including a confidence score (Claim 1). To address accuracy challenges, it then applies a special recalibration process to these confidence scores. This recalibration involves a combination of techniques: temperature scaling, selecting top-k probabilities from the softmax denominator, and polynomial regression (Claim 4). After adjusting the confidence scores, the system extracts specific pieces of information from the document and stores them in a database, each with its newly adjusted, more accurate confidence score (Claim 1). For example, if an LLM extracts a date from a contract, this method ensures the confidence score for that date is a more accurate reflection of the LLM's certainty.

What it doesn't cover

• —Does not cover improving confidence scores for machine learning models not specifically configured for document processing and understanding of electronic text documents.
• —Does not cover calibration methods that do not combine temperature scaling, softmax denominator top-k probability selection, and polynomial regression.
• —Does not cover systems that only provide a confidence score without then extracting individual elements of information and storing them with adjusted scores in a database.
• —Does not cover ML models that process non-textual data, such as images or audio, for confidence scoring.
• —Does not cover methods that improve selective prediction accuracy without also improving calibration quality.

The clever bit

The novelty lies in the specific combination of three distinct calibration techniques—temperature scaling, softmax denominator top-k probability selection, and polynomial regression—applied together to post-calibrate LLM confidence scores. This combined approach aims to improve both the calibration quality (how well the confidence score matches actual accuracy) and the selective prediction accuracy (how well the model knows when it's likely to be right or wrong) for document processing tasks.

Why it matters

Accurate confidence scores are crucial for large language models, especially when they are used to extract sensitive or critical data from documents. If an LLM reports high confidence for an incorrect answer, it can lead to bad decisions. This patent helps ensure that when an LLM extracts information, like financial figures or legal clauses, its stated confidence level is a true reflection of its accuracy, which is vital for secure machine learning systems and data platforms like Snowflake.

Real-world examples

• 1.Automated contract analysis platforms
• 2.Financial document processing systems
• 3.Legal discovery tools
• 4.Healthcare record summarization
• 5.Data extraction services in cloud platforms