Summary of The Perfect Blend: Redefining Rlhf with Mixture Of Judges, by Tengyu Xu et al.
The Perfect Blend: Redefining RLHF with Mixture of Judges
by Tengyu Xu, Eryk Helenowski, Karthik Abinav Sankararaman, Di Jin, Kaiyan Peng, Eric Han, Shaoliang Nie, Chen Zhu, Hejia Zhang, Wenxuan Zhou, Zhouhao Zeng, Yun He, Karishma Mandyam, Arya Talabzadeh, Madian Khabsa, Gabriel Cohen, Yuandong Tian, Hao Ma, Sinong Wang, Han Fang
First submitted to arxiv on: 30 Sep 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Artificial Intelligence (cs.AI); Computation and Language (cs.CL)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary This paper addresses the limitations of Reinforcement Learning from Human Feedback (RLHF) when applied to multi-task learning (MTL). RLHF is currently the dominant approach for fine-tuning large language models, but it struggles with reward hacking and extreme multi-objective optimization. The authors propose a novel post-training paradigm called Constrained Generative Policy Optimization (CGPO), which uses Mixture of Judges (MoJ) to identify the optimal blend in RLHF. CGPO shows strong empirical results, provides theoretical guarantees, and does not require extensive hyperparameter tuning. This plug-and-play approach can detect and mitigate reward hacking behaviors while achieving a pareto-optimal point across an extremely large number of objectives. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper helps us understand how to make big language models better at doing multiple tasks at once. Right now, we use something called Reinforcement Learning from Human Feedback (RLHF) to fine-tune these models. But RLHF has some problems when trying to do many tasks at the same time. The authors came up with a new way to solve this problem called Constrained Generative Policy Optimization (CGPO). CGPO uses a clever trick to figure out what’s the best combination of tasks for the model, and it works really well! It also helps prevent bad behavior in the model and gets close to being perfect at many tasks. This is an important breakthrough that can help us make language models even more powerful. |
Keywords
» Artificial intelligence » Fine tuning » Hyperparameter » Multi task » Optimization » Reinforcement learning from human feedback » Rlhf
