Chasing Moving Targets with Online Self-Play Reinforcement Learning for Safer Language Models

• URL: http://arxiv.org/abs/2506.07468v1
• Date: Mon, 09 Jun 2025 06:35:12 GMT
• Title: Chasing Moving Targets with Online Self-Play Reinforcement Learning for Safer Language Models
• Authors: Mickel Liu, Liwei Jiang, Yancheng Liang, Simon Shaolei Du, Yejin Choi, Tim Althoff, Natasha Jaques,
• Abstract summary: Conventional language model (LM) safety alignment relies on a reactive, disjoint procedure: attackers exploit a static model, followed by defensive fine-tuning to patch exposed vulnerabilities.<n>This sequential approach creates a mismatch -- attackers overfit to obsolete defenses, while defenders perpetually lag behind emerging threats.<n>We propose Self-RedTeam, an online self-play reinforcement learning algorithm where an attacker and defender agent co-evolve through continuous interaction.
• Score: 55.28518567702213
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Conventional language model (LM) safety alignment relies on a reactive, disjoint procedure: attackers exploit a static model, followed by defensive fine-tuning to patch exposed vulnerabilities. This sequential approach creates a mismatch -- attackers overfit to obsolete defenses, while defenders perpetually lag behind emerging threats. To address this, we propose Self-RedTeam, an online self-play reinforcement learning algorithm where an attacker and defender agent co-evolve through continuous interaction. We cast safety alignment as a two-player zero-sum game, where a single model alternates between attacker and defender roles -- generating adversarial prompts and safeguarding against them -- while a reward LM adjudicates outcomes. This enables dynamic co-adaptation. Grounded in the game-theoretic framework of zero-sum games, we establish a theoretical safety guarantee which motivates the design of our method: if self-play converges to a Nash Equilibrium, the defender will reliably produce safe responses to any adversarial input. Empirically, Self-RedTeam uncovers more diverse attacks (+21.8% SBERT) compared to attackers trained against static defenders and achieves higher robustness on safety benchmarks (e.g., +65.5% on WildJailBreak) than defenders trained against static attackers. We further propose hidden Chain-of-Thought, allowing agents to plan privately, which boosts adversarial diversity and reduces over-refusals. Our results motivate a shift from reactive patching to proactive co-evolution in LM safety training, enabling scalable, autonomous, and robust self-improvement of LMs via multi-agent reinforcement learning (MARL).

Related papers

• A Dynamic Stackelberg Game Framework for Agentic AI Defense Against LLM Jailbreaking [13.343937277604892]
  This paper presents a Stackelberg game framework to model the interactions between attackers and defenders in the context of large language models jailbreaking.<n>We propose a novel agentic AI solution, the "Purple Agent," which integrates adversarial exploration and defensive strategies using Rapidly-exploring Random Trees (RRT)
  arXiv  Detail & Related papers  (2025-07-10T22:37:47Z)
• ShieldLearner: A New Paradigm for Jailbreak Attack Defense in LLMs [4.534938642552179]
  ShieldLearner is a novel paradigm that mimics human learning in defense.<n>Through trial and error, it autonomously distills attack signatures into a Pattern Atlas.<n> Adaptive Adversarial Augmentation generates adversarial variations of successfully defended prompts.
  arXiv  Detail & Related papers  (2025-02-16T18:47:41Z)
• SPIN: Self-Supervised Prompt INjection [16.253558670549697]
  adversarial and jailbreak attacks have been proposed to bypass the safety alignment and cause the model to produce harmful responses. We introduce Self-supervised Prompt INjection (SPIN) which can detect and reverse these various attacks on LLMs. Our system can reduce the attack success rate by up to 87.9%, while maintaining the performance on benign user requests.
  arXiv  Detail & Related papers  (2024-10-17T05:40:54Z)
• Purple-teaming LLMs with Adversarial Defender Training [57.535241000787416]
  We present Purple-teaming LLMs with Adversarial Defender training (PAD) PAD is a pipeline designed to safeguard LLMs by novelly incorporating the red-teaming (attack) and blue-teaming (safety training) techniques. PAD significantly outperforms existing baselines in both finding effective attacks and establishing a robust safe guardrail.
  arXiv  Detail & Related papers  (2024-07-01T23:25:30Z)
• Toward Optimal LLM Alignments Using Two-Player Games [86.39338084862324]
  In this paper, we investigate alignment through the lens of two-agent games, involving iterative interactions between an adversarial and a defensive agent. We theoretically demonstrate that this iterative reinforcement learning optimization converges to a Nash Equilibrium for the game induced by the agents. Experimental results in safety scenarios demonstrate that learning in such a competitive environment not only fully trains agents but also leads to policies with enhanced generalization capabilities for both adversarial and defensive agents.
  arXiv  Detail & Related papers  (2024-06-16T15:24:50Z)
• On the Difficulty of Defending Contrastive Learning against Backdoor Attacks [58.824074124014224]
  We show how contrastive backdoor attacks operate through distinctive mechanisms. Our findings highlight the need for defenses tailored to the specificities of contrastive backdoor attacks.
  arXiv  Detail & Related papers  (2023-12-14T15:54:52Z)
• BadCLIP: Dual-Embedding Guided Backdoor Attack on Multimodal Contrastive Learning [85.2564206440109]
  This paper reveals the threats in this practical scenario that backdoor attacks can remain effective even after defenses. We introduce the emphtoolns attack, which is resistant to backdoor detection and model fine-tuning defenses.
  arXiv  Detail & Related papers  (2023-11-20T02:21:49Z)
• Learning Cyber Defence Tactics from Scratch with Multi-Agent Reinforcement Learning [4.796742432333795]
  Team of intelligent agents in computer network defence roles may reveal promising avenues to safeguard cyber and kinetic assets. Agents are evaluated on their ability to jointly mitigate attacker activity in host-based defence scenarios.
  arXiv  Detail & Related papers  (2023-08-25T14:07:50Z)
• Adversarial Machine Learning and Defense Game for NextG Signal Classification with Deep Learning [1.1726528038065764]
  NextG systems can employ deep neural networks (DNNs) for various tasks such as user equipment identification, physical layer authentication, and detection of incumbent users. This paper presents a game-theoretic framework to study the interactions of attack and defense for deep learning-based NextG signal classification.
  arXiv  Detail & Related papers  (2022-12-22T15:13:03Z)
• Fixed Points in Cyber Space: Rethinking Optimal Evasion Attacks in the Age of AI-NIDS [70.60975663021952]
  We study blackbox adversarial attacks on network classifiers. We argue that attacker-defender fixed points are themselves general-sum games with complex phase transitions. We show that a continual learning approach is required to study attacker-defender dynamics.
  arXiv  Detail & Related papers  (2021-11-23T23:42:16Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.