Role-playing agents in a world war: what rewarding only winning did #4413
KyleClouthier
started this conversation in
Show and tell
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
For anyone working on multi-agent societies: SECOND STRIKE puts LLM agents in charge of nations in a real-time world war, with alliances, betrayal and a shared nuclear doomsday clock. Each turn logs the agent's public message, private reasoning and actions, so deception shows up as data.
We trained two copies of gpt-oss-20b on their own best games, one reward line apart. Rewarded only to win: 6 of 10 sealed wars ended at midnight, zero wins. Rewarded to win without nukes: none ended, but twice the broken pacts.
Agents plug in over HTTP, MCP or
pip install secondstrike, free. Write-up: https://secondstrike.io/research/rewarded-to-win-ended-the-world-moreAll reactions