Tsinghua's AAArena: Claude Opus 5.5 + Claude Code Wins Half of 12 Adversarial Game Golds
THU-CST-SAST · hf · 2026-10-12
Tsinghua researchers formalize Adversarial Heuristic Learning (AHL) and launch AAArena, a benchmark of 12 real adversarial games plus 1,920 archived human programs where agents interpret rules, pick opponents, study replays, and revise game agents under fixed budgets. Key findings:
- Claude Opus 5.5 with Claude Code earns 6 of 12 gold medals, yet no configuration tops the human ladders in the other 6 games.
- Performance degrades with more complex rule specifications.
- Opponent selection and dense feedback improve policies; agents learn from both on-policy and off-policy replays, while game understanding and long-horizon strategy remain open challenges.
More from coding & agent
- snowpaw: open-source self-hosted messenger where your contacts are AI agents — gnukeith · 2026-10-12
- Simulithic launches Real-World Verification: thousands of parallel VMs simulate user behavior to catch bugs — KlausCodes · 2026-10-12
- Open experiment log pushes integer multiplication bounds down to κ≈0.001467, invites agent collaboration — IgorCarron · 2026-10-12
- Hu Yanbin built a stage light-beam instrument using ByteDance's TRAE coding tool — vista8 · 2026-10-12
- JetBrains paper: coding-agent benchmarks should match real users' task flows — rohanpaul_ai · 2026-10-12
- Dev is building an AGPL-licensed self-hostable Overleaf alternative — AlpinDale · 2026-10-12