Since CodeClash's release, our top priority has been enabling practitioners to improve models as CodeClash competitors and ultimately, long-running, autonomous software developers.

As an initial step, we're releasing an initial set of 9 arenas that we're designating as the official "train" split of CodeClash (CC:Train).

• Bridge
• Chess
• Figgie
• Gomoku
• Halite II
• Halite III
• MIT BattleCode 2023
• MIT BattleCode 2024
• MIT BattleCode 2025

CC:Train arenas span a range of properties, including:

• Perfect (Chess) vs. Imperfect Information (Bridge)
• Classical board games (Gomoku) vs. Custom competition formats (BattleCode)
• Head-to-head (Chess, Gomoku, BattleCode) vs. Multi-player (Figgie, Gomoku, Halite)

Today's models are mainly trained with tasks that use unit tests as verification (e.g., SWE-bench, SWE-smith).

We are curious if coding capabilities could improve by post-training on open-ended, competitive objectives. Some ideas:

• Self play RL with competition outcomes as rewards
• Transferability.
  • Training on Halite II/III may likely lead to better performance on Halite I, but what about Gomoku?
  • Does training on open-ended code tasks (e.g., CodeClash, improving runtime [1, 2, 3]) improve performance on in/out-of-distribution coding benchmarks?
• Mitigating code slop and bad development practices (e.g., single use scripts, redundant code, poor organization)