Research map / Adversarial robustness and security / Attacks on text and graphs
Adversarial Policies Beat Superhuman Go AIs
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add pass-alive defense to victim (prohibit passing unless only pass-alive territory moves remain) (ablation) | KataGo Latest pass-adversary · Go vs KataGo Latest, no search | adversary win rate: pass-adversary falls from 99.9% to 0 / 1000 games | Helped |
| Adversarial training on 0.08% cyclic-exploit self-play games (ablation) | KataGo b60 victim vs cyclic-adversary 600 visits · Go vs b60-s7702m 32 visits | adversary win rate: adversary drops to 0/50 won games | Helped |
| Fine-tune cyclic adversary against adversarially trained victim for 1154.9 V100 GPU days (ablation) | cyclic-adversary · Go vs b60-s7702m 4096 visits | adversary win rate: recovers to 47% over 400 games | Helped |
| Use A-MCTS-R (recursive victim simulation) instead of A-MCTS-S (sampling victim policy head) (ablation) | pass-adversary · Go vs KataGo Latest with 8 visits | adversary win rate: 88% with A-MCTS-R vs 15% with A-MCTS-S | Helped |
| Use A-MCTS-R instead of A-MCTS-S for the cyclic adversary (ablation) | cyclic-adversary · Go vs KataGo Latest with up to 128 victim visits | adversary win rate | No clear effect |
| Symmetry-averaged victim model (A-MCTS-S++) instead of A-MCTS-S (ablation) | cyclic-adversary · Go vs KataGo Latest with search | adversary win rate | No clear effect |
| Increase adversary visits beyond 600 (ablation) | cyclic-adversary · Go vs KataGo Latest with search | adversary win rate | Hurt |
| Increase victim search budget | cyclic-adversary vs KataGo Latest · Go vs KataGo Latest, 4096 to 10^7 visits | adversary win rate: 95.7% at 4096 visits, 82% at 10^6, 72% at 10^7 (win rate declines with more victim search) | Hurt |
| Pass-adversary trained against no-search victim transfers to low-search victim with A-MCTS-R | pass-adversary · Go vs KataGo Latest with 8 visits | adversary win rate: 88% with A-MCTS-R, 15% with A-MCTS-S | Mixed |
| Cyclic adversary zero-shot transfer to Leela Zero and ELF OpenGo | cyclic-adversary · Leela Zero and ELF OpenGo | adversary win rate: 6.1% vs Leela Zero, 3.5% vs ELF OpenGo | Mixed |
| Pass-alive defense added to victim (ablation) | KataGo Latest victim (cyclic adversary) · Go vs KataGo Latest, no search | adversary win rate: cyclic adversary achieves 100.0% vs hardened and 100.0% vs undefended | No clear effect |
Closest papers by meaning
- Can Go AIs Be Adversarially Robust? (AAAI 2025 · similarity 0.93)
- Rethinking Adversarial Policies: A Generalized Attack Formulation and Provable Defense in RL (ICLR 2024 · similarity 0.82)
- The Attacker in the Mirror: Breaking Self-Consistency in Safety via Anchored Bipolicy Self-Play (NeurIPS 2026 · similarity 0.81)
- RAT: Adversarial Attacks on Deep Reinforcement Agents for Targeted Behaviors (AAAI 2025 · similarity 0.81)
- Wolfpack Adversarial Attack for Robust Multi-Agent Reinforcement Learning (ICML 2025 · similarity 0.80)
- Robust and Diverse Multi-Agent Learning via Rational Policy Gradient (NeurIPS 2025 · similarity 0.80)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.