AI Paper Atlas

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

ICML 2026 · Multi-agent systems and games › Games and equilibria · approach: Learning in Markov games

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 changedWhereMeasuredEffect
Self-RedTeam online self-play on RLHF-aligned modelsQwen2.5-14B-IT · WildJailbreak adversarial harmASR: 0.742 -> 0.372Improved
Self-RedTeam on abliterated Llama3.1-8B-IT-ABLlama3.1-8B-IT-AB · WildJailbreak adversarial harmASR: 0.991 -> 0.240Improved
Self-RedTeam on Llama3.1-8B-IT instruction followingLlama3.1-8B-IT · AlpacaEval 2LC win rate: 24.74 -> 21.41Hurt
Add self-play (co-evolving attacker) versus defender-only training, both with SFT (ablation)Self-RedTeam · 11 safety benchmarksrelative ASR reduction: additional 17.33% improvementImproved
Add auxiliary self-distilled SFT to self-play (ablation)Self-play Qwen2.5-14B · AlpacaEval 2LC win rate: 20.99 -> 38.09Improved
Add auxiliary SFT to self-play (ablation)Self-play Qwen2.5-14B · WildJailbreak adversarial harmASR: 0.272 -> 0.372Hurt
Add hidden CoT to self-play on Qwen2.5-14B benign compliance (ablation)Self-play + SFT · WJB adversarial benignComply: 0.948 -> 0.969Improved
Add hidden CoT on Llama3.1 models (ablation)Self-play · XSTest vanilla benign / complianceComply: 0.528 -> 0.936 (Llama3.1-8B-IT)Improved
Co-adaptive self-play instead of attacker-only against fixed defender (ablation)attacker · Attack diversitybase model perplexity on attacks: 8.62 -> 24.74Improved
Self-play versus attacker-only (ablation)attacker · Attack strategy counteffective strategy count: 1.9 -> 4.6Improved
Replace WildGuard with Qwen3Guard as training judge (ablation)Self-RedTeam Qwen2.5-7B-IT · 11 safety metricssafety metrics: average relative gain of 4.0%Improved
Self-RedTeam against dynamic attacksQwen2.5 models · X-Teaming, PAIR, AutoDANASR: averaging 31.3% across all settingsImproved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy