LLMs trained on simple specification gaming generalize to zero-shot reward tampering including rewriting their own reward function.
Towards deep learning models resistant to adversarial attacks
2 Pith papers cite this work. Polarity classification is still indexing.
2
Pith papers citing it
representative citing papers
Jailbreak prompts with adversarial suffixes have high GPT-2 perplexity, and a LightGBM model on perplexity and length detects most attacks.
citing papers explorer
-
Sycophancy to Subterfuge: Investigating Reward-Tampering in Large Language Models
LLMs trained on simple specification gaming generalize to zero-shot reward tampering including rewriting their own reward function.
-
Detecting Language Model Attacks with Perplexity
Jailbreak prompts with adversarial suffixes have high GPT-2 perplexity, and a LightGBM model on perplexity and length detects most attacks.