Reward Templates on GRF Environment

[lexvvcy]

Hi, thanks for the clarification and for releasing the DB-Football codebase.

We understand that the GRF-related code cannot be released due to privacy constraints, and we fully respect this limitation.

For reproduction and fair comparison purposes, we would like to ask whether it is possible to share some high-level details about the intrinsic reward template design used for GRF in your experiments.

To be clear, we are not requesting any code or exact coefficients. Even a conceptual or abstract description would be extremely helpful. For example:

• The general form of the reward templates (e.g., linear combination of handcrafted signals, rule-based components, etc.)
• What types of environment signals are involved (e.g., distance to ball, possession, progress toward goal, scoring-related events)
• Whether the template parameters are fixed or evolved/learned (e.g., agent-level mutation, population-level evolution)
• Any high-level pseudo-code or mathematical formulation that describes the structure of the reward template

This information would greatly help us ensure a faithful reproduction of the GRF setup when applying M3HF on the public DB-Football environment.

Thanks again for your work and for maintaining the repository.