
🎉 Excited to share that RECAP has been accepted to #EMNLP2026 Main! The key idea: Flawed thinking can actually help reasoning models learn better alignment. 🧠🛡️ We find that injecting just a small amount of flawed reasoning can reduce model safety by up to 36%. 📉 Instead of simply avoiding these failures, we train reasoning models to recognize, override, and recover from them. 💪 Introducing RECAP 🔄 — an RL post-training method that teaches reasoning models to overcome unsafe reasoning trajectories and reroute toward safe & helpful answers, without additional training cost. ✨ Safer reasoning 🔒 Stronger jailbreak resistance 🙅 Lower overrefusal 🧠 Preserved reasoning ability A huge thank you to all of our amazing collaborators who made this work possible! 🙌 I’m especially grateful to @PoloChau , @jianfengchi , and @pinyuchenTW for their invaluable support, feedback, and guidance throughout this project. #LLM #AISafety #AIAlignment
















