01.10.26 (inclusive)

Стажировка в Endgame Studio Ltd.

MVP

ICT tasks

Robotics

Neurotechnology and artificial Intelligence

The prototype of the algorithm

When executing the same or similar queries multiple times, language models may prefer a limited set of typical responses. For example, if you are asked to select a number from a given range, certain values will appear much more often than others. In applied problems, the problem manifests itself in the repetition of similar ideas, plots, and formulations. This limits the usefulness of AI in finding alternatives, creating content, and generating synthetic data. We need a way to expand the variety of results while maintaining their meaningfulness, correctness, and compliance with the constraints of the task. Requirements: - Apply a single method to at least three models of different families without separate manual configuration for each test query and without modification of the basic parameters of the model. - Test the solution on three groups of tasks: selection from a given numerical set, selection from a given set of colors, as well as open name generation. - Compare the results with the sample obtained based on regular queries to the model. - Measure the deviation from the required distribution (percentage of errors) - before and after applying the method. - Specify the additional costs required to execute a single request: the number of requests to the model, the cost of tokens, and the time to receive a response. - Provide a prototype, a description of the method, experimental results, and instructions for reproduction. Methods for modifying a request, receiving and selecting multiple responses, as well as combinations thereof, are allowed. Each test request is processed independently, without saving and using the state of previous launches, including the history of requests and responses, counters, accumulated statistics and lists of used values. As part of the processing of a single test query, multiple calls to the model and the use of their intermediate results are allowed. Using an external source of randomness is allowed as part of the method, provided that it does not depend on the results of previous runs, but replacing the model's response with a random value does not solve the problem.

A working prototype demonstrating a reproducible increase in diversity while maintaining quality within a predefined tolerance. The main evaluation criteria are the amount of improvement, quality retention, portability between models and tasks, cost of application, and reproducibility. Part of the verification is carried out on tasks that were not disclosed to the participants in advance.

Sitnikov Ilya Alekseevich