🤖LLM Agents Battle in Pokémon Showdown, Talking Wins 64% of Battles
Talking LLMs Crush Silent Ones in Pokémon Showdown
TL;DR
LLM agents in a Pokémon Showdown 2v2 environment show that talking helps win 64% of battles. Free talk uses 978k prompt tokens per battle, 3.8x the 259k of no talk.
LLM agents in a Pokémon Showdown 2v2 environment show that talking helps win 64% of battles. This matters for developers working on multi-agent systems, as it highlights the importance of communication in achieving strategic goals. The environment uses a Gym-style setup with four players controlling one Pokémon each from a team of three, two per side. Agents read a plain-English summary of the battle and a numbered list of their legal moves each turn, with the ability to send private messages to their partner or choose a move.
Key Points
LLM agents in Pokémon Showdown 2v2 win 64% of battles when allowed to talk freely.
Free talk uses 978k prompt tokens per battle, 3.8x the 259k of no talk.
The environment uses a Gym-style setup with four players controlling one Pokémon each.
Agents read a plain-English summary of the battle and a numbered list of moves each turn.
The environment can tell models apart, with Qwen3.8-27B winning 96% against 2B models.
Why It Matters
If you're working on multi-agent systems, this highlights the importance of communication in achieving strategic goals. The environment uses a Gym-style setup with four players controlling one Pokémon each from a team of three, two per side. Agents read a plain-English summary of the battle and a numbered list of their legal moves each turn, with the ability to send private messages to their partner or choose a move.
Comments
Be the first to comment
Enjoyed this article?
Get it daily. 7am. Free. Reads in 5 minutes.
Join 3,569 builders reading daily.