Token Aggregation of Generative Models for Memory-Efficient Joint Output
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing web ecosystem auctions for digital content display do not effectively account for the flexibility of generative models, leading to increased memory usage and processing requirements due to the need for additional web space to cover various concepts from different digital content providers.
Innovation Solution
A token aggregator operates on a token-by-token basis to aggregate outputs from multiple generative models, using weight inputs to determine the relative contribution of each model's output, reducing memory usage and processing by generating a joint output.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If additional web space is allocated to cover various concepts from each digital content provider, then content diversity and provider flexibility are improved, but memory usage and processing requirements increase
Solution Approach 1:
The patent merges multiple generative model outputs into a single unified web space by aggregating token-level representations. Instead of allocating separate web spaces for each content provider, the system combines their outputs through weighted aggregation at the token distribution level, reducing memory usage while preserving content diversity from multiple providers.
Solution Approach 2:
The patent introduces a new dimension of aggregation by operating at the token distribution level rather than at the full content level. This dimensional shift allows the system to combine multiple provider outputs in a compressed representation space, achieving diverse content output with reduced memory requirements compared to storing complete separate content instances.
2Adaptability or versatility
If additional web space is allocated to cover various concepts from each digital content provider, then content diversity and provider flexibility are improved, but processing requirements increase
Solution Approach 1:
The patent segments the content generation process into token-level operations, where each generative model contributes weighted token distributions rather than complete content instances. This segmentation allows efficient processing by operating on discrete token units that can be aggregated mathematically, reducing the computational burden compared to processing and merging full content structures from multiple providers.
Solution Approach 2:
The patent changes the parameter space from full content representations to token probability distributions. By transforming the output of multiple generative models into weighted token distributions and aggregating these parameters, the system achieves diverse content generation with lower processing requirements, as parameter aggregation is computationally more efficient than content-level merging.
3Device complexity
If a simple auction mechanism is used to select the highest bidder's content, then implementation complexity is reduced, but the ability to leverage generative model flexibility and combine multiple content concepts is lost
Solution Approach 1:
The patent introduces a token aggregator as an intermediary mechanism between the auction system and content delivery. This intermediary receives weighted outputs from multiple generative models (representing different bidders) and combines them through aggregation, preserving the simplicity of the auction mechanism while enabling flexible combination of multiple content concepts through the intermediary's aggregation function.
Data Source
AI summary
Aspects of the disclosure are directed to a token aggregator for aggregating outputs from various generative models. The token aggregator can operate on a token-by-token basis, serving to aggregate several weighted generative model outputs to generate a joint output. By providing weights to the token aggregator as to what the preferred distribution may be, the weights can be used to tradeoff between generative model outputs to help determine the relative weight of the generative model outputs for creating the joint output as well as determining contribution amounts, e.g., bid payments, credits, or points, from respective model outputs.