Reinforcement Learning with Backtracking Feedback
Bilgehan Sel, Vaishakh Keshava, Phillip Wallis, Lukas Rutishauser, Ming Jin, Dingcheng Li
TL;DR
RLBF addresses robust safety for LLMs by enabling dynamic, in-generation self-correction through a token-efficient backtracking directive. It combines enhanced SFT data (BSAFE+) with reinforcement learning guided by a live safety critic, using a simple [CATEGORY_c] and [BACKTRACK_BY_X] signal to retract $X$ tokens and continue safely. The approach optimizes a trajectory-level reward $R_{final}$ via GRPO, integrating SFT guidance to improve learning efficiency. Empirical results show substantial reductions in attack success rates across benchmarks and model scales with minimal utility loss, illustrating a practical path toward safer, more trustworthy LLMs.
Abstract
Addressing the critical need for robust safety in Large Language Models (LLMs), particularly against adversarial attacks and in-distribution errors, we introduce Reinforcement Learning with Backtracking Feedback (RLBF). This framework advances upon prior methods, such as BSAFE, by primarily leveraging a Reinforcement Learning (RL) stage where models learn to dynamically correct their own generation errors. Through RL with critic feedback on the model's live outputs, LLMs are trained to identify and recover from their actual, emergent safety violations by emitting an efficient "backtrack by x tokens" signal, then continuing generation autoregressively. This RL process is crucial for instilling resilience against sophisticated adversarial strategies, including middle filling, Greedy Coordinate Gradient (GCG) attacks, and decoding parameter manipulations. To further support the acquisition of this backtracking capability, we also propose an enhanced Supervised Fine-Tuning (SFT) data generation strategy (BSAFE+). This method improves upon previous data creation techniques by injecting violations into coherent, originally safe text, providing more effective initial training for the backtracking mechanism. Comprehensive empirical evaluations demonstrate that RLBF significantly reduces attack success rates across diverse benchmarks and model scales, achieving superior safety outcomes while critically preserving foundational model utility.