LLM Self-Evaluation to Reduce Hallucination-Driven Compute Load
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Large language models (LLMs) consume significant resources and are prone to hallucinations, leading to the generation of factually inaccurate responses, which necessitate additional follow-up inputs to correct errors, increasing computational resource usage.
Innovation Solution
Implementing a self-evaluation and progressive refinement process to generate responses that adhere to specific instructions, where candidate responses are evaluated against the input instructions, and refined iteratively until they meet the required criteria, reducing the need for follow-up inputs by ensuring higher quality outputs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If LLMs are used to generate responses to NL based inputs, then natural language processing tasks can be performed, but computational resources are consumed significantly and hallucinations occur leading to factually inaccurate responses
Solution Approach 1:
The patent implements a self-evaluation mechanism where the LLM critiques its own candidate responses against the original instructions. This feedback loop allows the system to identify hallucinations and inaccuracies, then generate refined responses that better meet the requirements, thereby improving response accuracy while maintaining productivity
Solution Approach 2:
The system performs preliminary evaluation of candidate responses before final output by comparing them against the instructions. This preliminary action identifies potential hallucinations early, allowing the system to correct issues before presenting the final response to the user
2Productivity
If LLMs generate responses that may include hallucinations, then response generation can be completed, but additional follow-up inputs are required to correct errors, increasing computational resource usage
Solution Approach 1:
The system performs preliminary self-evaluation and refinement of responses before final output. By checking candidate responses against instructions in advance and correcting potential hallucinations, the system reduces the need for follow-up user inputs, thereby decreasing computational resource consumption while maintaining response generation speed
Solution Approach 2:
The LLM performs self-evaluation and self-correction of its own responses without requiring external intervention. This self-service mechanism identifies and corrects hallucinations autonomously, reducing the need for additional computational resources spent on follow-up corrections
Data Source
AI summary
Implementations relate to improving instruction following capabilities of large language models (LLMs) using instruction decomposition, self-evaluation, and optionally progressive refinement. Processor(s) of a system can: obtain natural language (NL) based input, generate a plurality of candidate responses and evaluate the candidate responses based on instructions included in the NL based input, using an LLM, and progressively refine the candidate responses until it is determined that one or more termination criteria are satisfied. In some implementations, the NL based input can be received from a client device. In these implementations, a given candidate response that is progressively refined can be rendered for presentation at the client device and responsive to the NL base input. In additional or alternative implementations, the NL based input can be obtained from database(s). In these implementations, a given candidate response that is progressively refined can be utilized in fine-tuning of the LLM.