LLM Worker Assignment With Token-Length and External-Data Profiles
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing task assignment techniques for large-scale language models (LLMs) do not adequately consider the unique characteristics such as token length and the use of external data, often requiring multiple attempts to achieve a desired result, leading to inefficiencies.
Innovation Solution
A computer system that manages worker systems, including LLMs, by using worker management information to identify candidate workers based on characteristics like token length, external data use, and performance history, selecting the most suitable worker for task execution.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If existing task assignment techniques are used for LLMs, then task assignment can be performed using general worker management methods, but the unique characteristics of LLMs such as token length and external data use are not considered, leading to multiple attempts being required to achieve desired results
Solution Approach 1:
The patent applies local quality by creating worker management information that is specific to each LLM's characteristics (token length, external data usage, model architecture) rather than using generic worker management. This allows the task assignment system to match tasks to LLMs based on their specific strengths and limitations, improving first-attempt success rates without requiring multiple retries
Solution Approach 2:
The patent changes the parameters used for worker evaluation from generic metrics to LLM-specific parameters including token length capabilities, external data usage requirements, and model architecture characteristics. By optimizing task assignment based on these specific parameters, the system achieves better matching between task requirements and LLM capabilities, reducing the need for multiple attempts
2Productivity
If LLMs are assigned tasks without considering their characteristics, then task assignment process is simple, but multiple attempts are required to achieve desired results, reducing execution efficiency
Solution Approach 1:
The patent implements preliminary action by pre-collecting and organizing worker management information that documents each LLM's characteristics (token length, external data usage, model architecture) before task assignment occurs. This preparatory work enables rapid matching during task assignment without adding complexity to the actual execution process, thereby improving productivity while maintaining manageable system complexity
Solution Approach 2:
The patent adds new dimensions to the task assignment system by incorporating LLM-specific characteristics (token length, external data usage, model architecture) as additional selection criteria beyond traditional worker metrics. This multi-dimensional approach enables more precise matching and improves execution efficiency, with the added complexity confined to the assignment decision layer rather than the execution layer
3Measurement precision
If worker management information includes detailed LLM characteristics, then accurate worker identification is possible, but data storage and processing requirements increase
Solution Approach 1:
The patent extracts only the most critical LLM characteristics (token length, external data usage, model architecture) into the worker management information, rather than storing all possible model parameters and metadata. This selective extraction maintains high worker identification accuracy while minimizing data storage requirements by focusing only on the attributes most relevant to task matching
Data Source
AI summary
A computer system is connected with a plurality of worker systems where workers that execute tasks are operated, and holds worker management information for managing the workers. In the worker management information, data that is constituted of items indicating identification information of the workers and characteristics of the workers is stored. In at least one worker system, a large-scale language model is operated as the worker. The computer system receives a task execution request, generates the task execution information relating to the plurality of tasks that are executed until a desired result is obtained with respect to the respective workers, identifies candidate workers based on the worker management information and the task execution information, and selects the worker to which the task is assigned out of the candidate workers.