In-Context Demonstration Selection Using Difficulty-Based Repetition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional in-context learning (ICL) strategies for long-context generative neural networks (LLMs) are ineffective when utilizing a large number of demonstrations, as the effectiveness of increasing the number of examples varies greatly depending on how the demonstrations are selected, and simply increasing the number does not guarantee improved performance.
Innovation Solution
The system strategically repeats difficult demonstration examples within the context input and incorporates zero-shot predictions as error signals to enhance LLM attention, improving performance without further training.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If the number of demonstration examples is increased, then the model can learn more from in-context learning, but the effectiveness decreases when using traditional selection strategies
Solution Approach 1:
The patent applies local quality by differentiating the treatment of demonstration examples based on their difficulty level. Instead of uniform treatment, difficult examples are repeated multiple times while easy examples are included once, creating localized emphasis on critical learning patterns that the model struggles with, thereby improving overall in-context learning effectiveness
Solution Approach 2:
The patent segments the demonstration examples into distinct categories based on difficulty classification (easy, medium, difficult). This segmentation allows the system to apply different repetition strategies to different segments, with difficult examples receiving multiple repetitions and easy examples receiving single inclusion, optimizing the learning process
2Adaptability or versatility
If traditional ICL selection strategies are used, then similarity and diversity are balanced, but performance does not improve when using a large number of demonstrations
Solution Approach 1:
The patent changes the selection parameter from traditional similarity-based scoring to difficulty-based classification. By evaluating examples through difficulty metrics rather than solely relying on similarity scores, the system identifies which examples provide the most learning value and repeats those, thereby improving task performance while maintaining diversity
3Loss of information
If more demonstration examples are included in the context input, then the model has more information to process, but the model cannot effectively focus on challenging examples
Solution Approach 1:
The patent implements feedback by using zero-shot predictions as error signals. The model generates predictions for difficult examples without training data, and these errors are fed back into the context input where they are repeated. This feedback mechanism ensures the model pays attention to challenging examples by highlighting where it fails, thereby improving detection and measurement of difficult patterns
Data Source
AI summary
Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for performing in-context learning using a generative neural network. In one aspect, a method comprises obtaining a plurality of demonstration examples for a task; obtaining a respective difficulty classification for each of the demonstration examples; generating a context input that includes one or more instances of at least a subset of the demonstration examples, the generating comprising, for each of the demonstration examples, determining how many instances of the demonstration example to include in the context input based on the respective difficulty classification for the demonstration example; receiving a new input for the task; and processing an input that includes the context input and the new input using the first generative neural network to generate a new output.