Dynamic Token Prediction for Variable Sequence Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current token prediction methods rely on fixed numbers of previously observed tokens, which may not be optimal for all instances, as they assume a static snapshot of history and fail to account for variable sequences that could provide more accurate predictions.
Innovation Solution
A system that dynamically adjusts the number of previously observed tokens to predict the next token by building a dictionary of observed n-tuples up to a predetermined point and recursively adding tokens until a unique or target number of predictions is achieved, allowing for variable-length sequences to determine the next token.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a fixed number of previously observed tokens is used for prediction, then the prediction method is simple and fast, but the prediction accuracy is limited because it cannot adapt to variable historical sequences
Solution Approach 1:
The patent applies dynamics by transitioning from a fixed-size prediction window to a variable-size window that adapts based on the uniqueness of historical sequences. The system dynamically adjusts the number of previously observed tokens to consider, expanding the window when needed to find unique sequences and contracting it when simpler patterns suffice. This dynamic adaptation resolves the contradiction by allowing the system to achieve higher prediction accuracy through variable sequence lengths while maintaining computational efficiency when fixed patterns are sufficient.
Solution Approach 2:
The patent changes the parameter of sequence length from a fixed value to a variable that can be adjusted based on the uniqueness of observed token sequences. By modifying this parameter dynamically - expanding it when historical sequences need to be more extensive to achieve uniqueness, and contracting it when shorter sequences suffice - the system resolves the contradiction between accuracy and complexity. The parameter change allows the prediction method to adapt its complexity to the specific requirements of each prediction task.
2Measurement precision
If more previously observed tokens are considered, then prediction accuracy improves, but the processing time and computational resources increase
Solution Approach 1:
The system dynamically adjusts the number of observed tokens based on the uniqueness requirement rather than always using the maximum available history. When a unique prediction can be achieved with fewer tokens, the system contracts the observation window, thereby reducing processing time. This dynamic adjustment resolves the contradiction by making the processing time variable - longer only when necessary to achieve uniqueness, and shorter when simpler patterns suffice.
Solution Approach 2:
The patent applies preliminary action by building and storing a dictionary of observed n-tuples in advance, organizing historical sequences for efficient retrieval. This pre-processing allows the system to quickly check for unique sequences without repeatedly analyzing raw historical data during prediction. The preliminary organization of data resolves the contradiction by reducing the computational burden during actual prediction, allowing the system to consider more historical tokens when needed without proportionally increasing processing time.
3Productivity
If a static snapshot of recent history is used, then the prediction method is computationally efficient, but it fails to account for variable sequences that could provide more accurate predictions
Solution Approach 1:
The patent transforms the static snapshot approach into a dynamic one by allowing the observation window to expand and contract based on the uniqueness of historical sequences. The system starts with a smaller window for efficiency and expands it only when necessary to achieve unique predictions, thereby maintaining productivity while improving accuracy when variable sequences are needed. This dynamic approach resolves the contradiction by making the system adaptable to different sequence complexities.
Solution Approach 2:
The system performs preliminary organization of historical data into a dictionary structure that enables efficient querying of sequence uniqueness. This pre-processed organization allows the system to quickly determine whether a given sequence is unique without extensive computation, thereby maintaining prediction efficiency while enabling the use of variable-length sequences for improved accuracy when needed.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
An approach is provided for next token prediction based on previously observed tokens. The approach involves receiving an observed time series of tokens, wherein each of the tokens represents an observed data pattern. The approach also involves adding a most recent token from the observed time series of tokens into a variable token set. The approach further involves processing a historical token set to determine a historical token sequence comprising the variable token set followed by a next token. The approach further involves recursively adding a next most recent token from the observed time series of tokens into the variable token set for processing until the next token following the variable token set in the determined historical token sequence is unique or meets a target number of possible predictions. The approach further involves presenting the next token as a predicted next token of the observed time series of tokens.