Generating Adversarial Texts for Machine Translation via GRPO
Applied Group Relative Policy Optimization to fine-tune Llama-3.1-8B, producing a scalable RL pipeline for adversarial NLP examples. A multi-objective reward system (COMET, semantic similarity, CoLA) yields rewrites that degrade state-of-the-art translation models while keeping 87%+ naturalness, generalizing across WMT19–WMT24.
Result: adversarial behaviour that transfers to unseen benchmarks — a concrete step toward stress-testing the robustness of production translation systems.