LLM Authorship Traceability Using Distributed Ledger Feedback
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Large language models (LLMs) often generate inaccurate, hazardous, or dangerous responses due to their black-box nature, making it difficult to trace the source or authorship of the generated content.
Innovation Solution
A computing platform trains a LLM using iterative clustering and records each clustering iteration on a distributed ledger, allowing for the identification and remediation of errors, and updates the model based on accuracy thresholds.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If LLM is trained using self-supervised learning on vast amounts of unlabeled text, then the model can generate responses quickly, but the accuracy and reliability of the generated responses deteriorate
Solution Approach 1:
The patent implements a feedback mechanism where the LLM's generated responses are evaluated against ground truth data, and the model is retrained using supervised fine-tuning based on the identification of erroneous responses. This closed-loop feedback system continuously improves response accuracy while maintaining the ability to generate responses efficiently.
Solution Approach 2:
The patent performs preliminary supervised fine-tuning training on a subset of labeled data before deploying the model for production. This preliminary action establishes a baseline level of accuracy and enables the model to generate reliable responses while maintaining fast generation speeds through the pre-trained parameters.
2Device complexity
If LLM generates responses using a black box approach, then the system operates simply, but the ability to trace error sources and identify authorship deteriorates
Solution Approach 1:
The patent introduces an intermediary evaluation layer that sits between the LLM's black box generation process and the final output. This intermediary component evaluates generated responses against ground truth data, identifies errors, and traces them back to specific training iterations or data sources, thereby maintaining system simplicity while enabling comprehensive traceability.
Solution Approach 2:
The patent segments the training and evaluation process into distinct phases: initial self-supervised pre-training, supervised fine-tuning on labeled data, and iterative evaluation-retraining cycles. Each phase is separately managed and tracked, allowing errors to be traced to specific segments or iterations without complicating the overall system architecture.
3Reliability
If LLM is continuously improved through additional training, then the model accuracy improves, but the time and computational resources required increase
Solution Approach 1:
The patent applies partial action by performing supervised fine-tuning on a strategically selected subset of labeled data rather than retraining on the entire corpus. This approach achieves sufficient accuracy improvements without the prohibitive time and computational costs of complete retraining, balancing model improvement with resource constraints.
Solution Approach 2:
The patent implements periodic evaluation and retraining cycles where the model is assessed at intervals against ground truth data, and retraining is performed only when accuracy degradation is detected or at predetermined intervals. This periodic approach maintains model accuracy over time while minimizing unnecessary training time and computational resource expenditure.
Data Source
AI summary
A computing platform may train an LLM by performing a plurality of clustering iterations. The computing platform may record, using a distributed ledger and for each of the plurality of clustering iterations, information of the corresponding clustering iteration. The computing platform may test the LLM, which may reveal at least one error in the LLM. The computing platform may identify, using the distributed ledger, a clustering iteration corresponding to the at least one error and information associated with the clustering iteration. The computing platform may identify, based on the clustering iteration corresponding to the at least one error and the information associated with the clustering iteration, a remediating action. The computing platform may update the LLM by performing the remediating action. The computing platform may retest the LLM. Based on identifying that an accuracy of the LLM meets or exceeds an accuracy threshold, the computing platform may deploy the LLM.