Are PPO-ed Language Models Hackable?
Abstract
Numerous algorithms have been proposed to $\textit{align}$ language models to remove undesirable behaviors. However, the challenges associated with a very large state space and creating a proper reward function often result in various jailbreaks. Our paper aims to examine this effect of reward in the controlled setting of positive sentiment language generation. Instead of online training of a reward model based on human feedback, we employ a statically learned sentiment classifier. We also consider a setting where our model's weights and activations are exposed to an end-user after training. We examine a pretrained GPT-2 through the lens of mechanistic interpretability before and after proximal policy optimization (PPO) has been applied to promote positive sentiment responses. Using these insights, we (1) attempt to "hack" the PPO-ed model to generate negative sentiment responses and (2) add a term to the reward function to try and alter `negative' weights.
Read the paper
Related studies
- The Matrisome: In Silico Definition and In Vivo Characterization by Proteomics of Normal and Tumor Extracellular Matrices
- The Escherichia coli MG1655 in silico metabolic genotype: Its definition, characteristics, and capabilities
- In silico pharmacology for drug discovery: methods for virtual ligand screening and profiling
- In silico prediction of splice-altering single nucleotide variants in the human genome
- In silico clinical trials: how computer simulation will transform the biomedical industry
- In silico optimization of a guava antimicrobial peptide enables combinatorial exploration for peptide design
Go deeper in Evidence Hub
Evidence Hub adds a study data sheet (design, sample size, evidence level, key numbers), an AI summary and a chat grounded in the paper's text.