LLM Self-Evaluation for Reliable Selective Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Large Language Models (LLMs) face reliability issues in high-stakes decision-making scenarios due to the potential for inaccurate responses, lacking the ability to reliably determine the correctness of their generated answers.
Innovation Solution
The ASPIRE framework trains LLMs on a question answering task using self-evaluation to learn the distinction between correct and incorrect answers, incorporating a selection score that combines the likelihood of the generated answer with a self-evaluation score for selective prediction, employing parameter-efficient fine-tuning techniques to reduce computational cost.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If LLMs are trained to improve reliability and accuracy in high-stakes decision-making, then the reliability of responses is improved, but the computational cost and complexity of the system increases
Solution Approach 1:
The patent segments the LLM training process into two distinct phases: (1) task-specific training to learn the domain knowledge, and (2) self-evaluation training to learn correctness assessment. This segmentation allows the system to acquire reliability capabilities without requiring complete retraining of the entire model, thereby managing computational complexity while improving reliability.
Solution Approach 2:
The patent applies preliminary action by first training the LLM on task-specific data to establish baseline performance, then subsequently training on self-evaluation data. This staged approach allows the model to build foundational skills before acquiring evaluation capabilities, reducing overall computational burden compared to simultaneous training on all objectives.
2Measurement precision
If LLMs are trained to distinguish correct from incorrect answers through self-evaluation, then the ability to provide accurate responses is improved, but the training time and computational resources increase
Solution Approach 1:
The patent extracts the self-evaluation capability as a separate training objective from the main task performance. By isolating the correctness detection function into its own training phase with specific evaluation metrics, the system achieves precise answer correctness detection without requiring excessive training time on combined objectives.
Solution Approach 2:
The patent employs parameter changes by adjusting training hyperparameters, learning rates, and data sampling strategies during self-evaluation training. These parameter optimizations enable the model to learn correctness detection efficiently, reducing training time while maintaining high measurement precision in answer evaluation.
3Manufacturing precision
If a selection score combining likelihood and self-evaluation is used for selective prediction, then the accuracy of predictions is improved, but the computational cost increases
Solution Approach 1:
The patent merges two complementary signals - the model's inherent likelihood estimation and the learned self-evaluation score - into a unified selection score for selective prediction. This combination leverages both probabilistic reasoning and explicit correctness assessment, achieving high prediction accuracy. The computational cost is managed by efficiently integrating these signals through weighted combination rather than requiring separate independent evaluation processes.
Data Source
AI summary
Aspects of the disclosure are directed to methods, systems, and computer readable media for adaptation with self-evaluation to improve selective prediction in large language models (LLMs), generally referred to as ASPIRE. ASPIRE includes training LLMs on a portion of training data from a question answering task to learn self-evaluation, e.g., learn to distinguish whether a generated answer is correct or not. ASPIRE further includes a selection score that combines a likelihood of that generated answer is correct with a self-evaluation score for selective prediction. ASPIRE demonstrates improved selective prediction performance with less computational cost.