Reproduce “Quark: Controllable Text Generation with Reinforced Unlearning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Ximing Lu, Sean Welleck, Jack Hessel, Liwei Jiang, Lianhui Qin, Peter West, Prithviraj Ammanabrolu, Yejin Choi
Quark introduces a stable and efficient reinforcement learning algorithm for post-hoc modification of large language models to "unlearn" undesirable behaviors like toxicity and repetition. The method effectively reduces these unwanted properties while preserving the model's core generative capabilities and outperforming existing baselines across multiple tasks.
Key insights
- Quantizing reward signals into learned reward tokens and conditioning generation on them provides a stable and efficient way to steer LLMs without complex RL setups or auxiliary models.
- A token-level KL-divergence penalty is crucial for preserving the core language quality and preventing catastrophic forgetting during the unlearning process.
- Targeted exploration, by conditioning on the highest reward token, effectively accelerates the discovery of desirable samples for the data pool, improving learning efficiency.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.