Test-Time Reinforcement Learning Amplifies Adversarial Safety Vulnerabilities in LLMs

arXiv 2603.15417·high signal

Researchers show that test-time reinforcement learning (TTR), a self-consistency-based method for improving LLM reasoning at inference time, is critically vulnerable to harmful prompt injections that get amplified through the learning loop rather than filtered. A model that is safe at baseline can be gradually steered into unsafe outputs as TTR compounds adversarial signals. The paper demonstrates that TTT/TTR methods require explicit adversarial robustness design — inference-time learning without safety gating is dangerous.