Targeted Latent Adversarial Training — Reproduction
Laptop-scale reproduction of latent-space adversarial robustness for LLMs.
Reproduction of Targeted Latent Adversarial Training (arXiv:2407.15549). I contributed code fixes grounded in the paper’s future-work section (hidden-size handling, device-agnostic execution) and ran a laptop-scale GPT-2 latent-attack study that reproduces the reported layer × epsilon sensitivity of the defense.
Original implementation: aengusl/latent-adversarial-training.
Stack: PyTorch · Transformers · Jupyter