Research map / Multi-agent systems and games / Games and equilibria
Chasing Moving Targets with Online Self-Play Reinforcement Learning for Safer Language Models
Mickel Liu, Liwei Jiang, Yancheng Liang, Simon Shaolei Du, Yejin Choi, Tim Althoff, Natasha Jaques
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 |
|---|---|---|---|
| Self-RedTeam online self-play on RLHF-aligned models | Qwen2.5-14B-IT · WildJailbreak adversarial harm | ASR: 0.742 -> 0.372 | Improved |
| Self-RedTeam on abliterated Llama3.1-8B-IT-AB | Llama3.1-8B-IT-AB · WildJailbreak adversarial harm | ASR: 0.991 -> 0.240 | Improved |
| Self-RedTeam on Llama3.1-8B-IT instruction following | Llama3.1-8B-IT · AlpacaEval 2 | LC win rate: 24.74 -> 21.41 | Hurt |
| Add self-play (co-evolving attacker) versus defender-only training, both with SFT (ablation) | Self-RedTeam · 11 safety benchmarks | relative ASR reduction: additional 17.33% improvement | Improved |
| Add auxiliary self-distilled SFT to self-play (ablation) | Self-play Qwen2.5-14B · AlpacaEval 2 | LC win rate: 20.99 -> 38.09 | Improved |
| Add auxiliary SFT to self-play (ablation) | Self-play Qwen2.5-14B · WildJailbreak adversarial harm | ASR: 0.272 -> 0.372 | Hurt |
| Add hidden CoT to self-play on Qwen2.5-14B benign compliance (ablation) | Self-play + SFT · WJB adversarial benign | Comply: 0.948 -> 0.969 | Improved |
| Add hidden CoT on Llama3.1 models (ablation) | Self-play · XSTest vanilla benign / compliance | Comply: 0.528 -> 0.936 (Llama3.1-8B-IT) | Improved |
| Co-adaptive self-play instead of attacker-only against fixed defender (ablation) | attacker · Attack diversity | base model perplexity on attacks: 8.62 -> 24.74 | Improved |
| Self-play versus attacker-only (ablation) | attacker · Attack strategy count | effective strategy count: 1.9 -> 4.6 | Improved |
| Replace WildGuard with Qwen3Guard as training judge (ablation) | Self-RedTeam Qwen2.5-7B-IT · 11 safety metrics | safety metrics: average relative gain of 4.0% | Improved |
| Self-RedTeam against dynamic attacks | Qwen2.5 models · X-Teaming, PAIR, AutoDAN | ASR: averaging 31.3% across all settings | Improved |
Closest papers by meaning
- The Attacker in the Mirror: Breaking Self-Consistency in Safety via Anchored Bipolicy Self-Play (NeurIPS 2026 · similarity 0.92)
- AdvEvo-MARL: Shaping Internalized Safety through Adversarial Co-Evolution in Multi-Agent Reinforcement Learning (ICML 2026 · similarity 0.90)
- There are Levels to It: Red Teaming LLMs with Hierarchical Reinforcement Learning (NeurIPS 2026 · similarity 0.89)
- Self-playing Adversarial Language Game Enhances LLM Reasoning (NeurIPS 2024 · similarity 0.87)
- Learning Diverse Attacks on Large Language Models for Robust Red-Teaming and Safety Tuning (ICLR 2025 · similarity 0.87)
- The Alignment Waltz: Jointly Training Agents to Collaborate for Safety (ICLR 2026 · similarity 0.87)
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.
