LLM Debiasing via Chain-of-Thought and External Knowledge
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Large language models exhibit bias and inaccuracies due to outdated training data, leading to knowledge conflicts and computational inefficiencies, which conventional debiasing techniques fail to address effectively.
Innovation Solution
A debiasing module employs chain-of-thought prompting and external knowledge as an instrumental variable to estimate causal effects by generating factual and counterfactual prompts, detecting and mitigating bias in the internal knowledge representation of machine-learning models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If large language models are trained on vast amounts of training data, then the model's knowledge representation and language understanding improve, but bias and inaccuracies increase due to real-world scenarios exhibiting bias
Solution Approach 1:
The patent introduces an intermediary debiasing module that sits between the training data and the model's knowledge representation. This module processes training data through chain-of-thought prompting, creating a mediator layer that filters and adjusts the data before it influences the model's internal representations, thereby reducing bias while preserving accuracy
Solution Approach 2:
The patent implements feedback mechanisms where the model's generated responses are evaluated for bias, and this evaluation information is fed back into the training process. The debiasing module uses this feedback to adjust and refine the knowledge representation, creating a iterative process that continuously reduces bias while maintaining accuracy
2Object-generated harmful factors
If conventional debiasing techniques are used, then some bias reduction is achieved, but knowledge conflicts and computational inefficiencies remain
Solution Approach 1:
The patent applies preliminary action by performing chain-of-thought processing and bias detection during the training phase rather than after model deployment. This preliminary debiasing action prevents knowledge conflicts from arising in the first place, reducing the need for complex post-training adjustments and improving overall computational efficiency
Solution Approach 2:
The patent segments the debiasing process into distinct modules: chain-of-thought prompting module, bias detection module, and knowledge representation adjustment module. This segmentation allows each component to handle specific aspects of bias reduction independently, improving computational efficiency by avoiding the need to process all data through all operations simultaneously
Data Source
AI summary
Change-of-thought machine-learning model debiasing techniques and systems are described. A query is received and context data is produced based on the query, e.g., from an external source. A prompt is generated that includes the context data, the query, and a chain-of-though prompt, which is processed by a machine-learning model. A candidate result based on processing of the prompt using the machine-learning model. The candidate result includes a candidate answer and a chain-of-thought result describing reasoning indicated by the machine-learning model as used in generating the candidate answer.