Research map / Adversarial robustness and security / Attacks on text and graphs
Safety Alignment of LMs via Non-cooperative Games
Anselm Paulus, Ilia Kulikov, Brandon Amos, Rémi Munos, Ivan Evtimov, Kamalika Chaudhuri, Arman Zharmagambetov
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| use non-zero-sum joint Attacker-Defender training (AdvGame-DPO-MD) instead of Self-RedTeam | Qwen2.5-7B Defender · HarmBench adversarial | ASR: 16.8 -> 4.7 | Improved |
| use AdvGame-DPO-MD instead of Self-RedTeam | Llama3.1-8B Defender · AlpacaEval2 utility | AlpacaEval2 score: 13.1 -> 32.8 | Improved |
| train with AdvGame-DPO-MD on Llama | Llama3.1-8B Defender · WJB benign compliance | compliance: 98.8 -> 69.9 | Hurt |
| train with AdvGame-DPO-MD on Llama | Llama3.1-8B Defender · DAN | ASR: 49.3 -> 42.0 (Self-RedTeam 32.3) | Mixed |
| train the Attacker instead of keeping it fixed (ablation) | AdvGame-DPO-MD · WJB adversarial | ASR: 16.6 -> 8.5 | Improved |
| use pairwise preference judge instead of point-wise judge (ablation) | AdvGame-DPO-MD · WJB adversarial | ASR: 30.8 -> 8.5 | Improved |
| use pairwise judge instead of point-wise judge (ablation) | AdvGame-DPO-MD · ArenaHard | ArenaHard score: 56.3 -> 61.3 | Improved |
| use EMA off-policy generator instead of on-policy generator (ablation) | AdvGame-DPO-MD · WJB adversarial | ASR: 49.5 -> 8.5 | Improved |
| use DPO-MD instead of GRPO with scalar rewards (ablation) | AdvGame · HarmBench adversarial | ASR: 14.9 -> 4.7 | Improved |
| use IPO-MD instead of DPO-MD (ablation) | AdvGame · HarmBench / WJB / DAN adversarial | ASR: WJB 8.5 -> 7.8; DAN 10.3 -> 13.0 | No clear effect |
| use trained AdvGame Defender against adaptive PAIR attack | Qwen2.5-7B Defender · HarmBench test, PAIR | ASR: 45.0 -> 7.2 | Improved |
| use AdvGame Defender against white-box GCG attack | Qwen2.5-7B Defender · HarmBench test, GCG | ASR: 61.6 -> 25.3 (Self-RedTeam 23.4) | Mixed |
Closest papers by meaning
- Safety Alignment Can Be Not Superficial With Explicit Safety Signals (ICML 2025 · similarity 0.90)
- The Alignment Waltz: Jointly Training Agents to Collaborate for Safety (ICLR 2026 · similarity 0.89)
- AlphaAlign: Incentivizing Safety Alignment with Extremely Simplified Reinforcement Learning (ICLR 2026 · similarity 0.89)
- Safety Alignment Should be Made More Than Just a Few Tokens Deep (ICLR 2025 · similarity 0.89)
- Safe RLHF: Safe Reinforcement Learning from Human Feedback (ICLR 2024 · similarity 0.88)
- BSO: Safety Alignment Is Density Ratio Matching (NeurIPS 2026 · similarity 0.88)
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.
