I am currently participating in Track 1 of the Orak Challenge.
I would like to ask if there are any specific constraints regarding the agent implementation. Aside from the requirement to use an LLM with 8B parameters or fewer, are there any restrictions on the internal operations of the agent?
Specifically, I would like to confirm if we are permitted to perform tasks such as pre-processing data, post-processing outputs, or enabling tool usage. Furthermore, could you please confirm if we have full freedom to implement the agent in any other way beyond the examples listed above?
Additionally, I would like to ask if there is a strict rule that one act call must correspond to exactly one LLM call, or if it is permissible to make multiple calls or no calls at all within a single act.
And Do I have to play 4 games with just one 8B llm?
I want to make sure there are no limitations on the internal logic or processes.
i have the same question, what sort of additional frameworks are allowed to use, because you can use deep learning model for these tasks, and if you do, is it going to be considered violation or not? Can you be more concrete, what is allowed and what is not allowed to use?
Like many others, I have a question regarding the eligible models for Track 1. According to the rules, the requirements are as follows:
Parameter Limit: Maximum 8B total parameters.
Release Date: Before November 1, 2025.
Fine-tuning: Allowed using public or team-generated data.
Based on these criteria, the following approach appears to meet the requirements: Taking a very simple model released before November 1st (e.g., a 1-layer model) and fine-tuning it specifically for the “2048” game task using Reinforcement Learning.
It seems that, technically, even optimizing an LLM with a single parameter using Reinforcement Learning would not violate the rules. Could you please confirm if this understanding is correct?