Personalized Soups: LLM Alignment Via Parameter Merging - Conclusion & References

5 CONCLUSION

Previous work has shown that adapting LLMs with RLHF helps them generate outputs that are preferred by humans over the supervised fine-tuned counterpart. However, recent work has also pointed out that simply training LLMs to abide by the preference of the general may result in ignoring individual preferences and values. In this work, we provide the first steps to tackle this issue by proposing Reinforcement Learning from Personalized Human Feedback as a multi-objective problem so that LLMs can be aligned to follow conflicting preferences. We propose a promising method called P-SOUPS that is able to composite models trained on single objectives on the fly during inference. We also highlight the scalability of P-SOUPS by showing that it scales linearly, instead of exponentially like the MORL baseline, with regards to the number of new preferences, which is required to provide true personalization to individual users.

ACKNOWLEDGMENTS

Thanks to Minyoung Hwang, Sungdong Kim, Tim Dettmers, Yoonjoo Lee, and Margaret Li for helpful feedback and discussion.

REFERENCES

Xiang Ao, Xiting Wang, Ling Luo, Ying Qiao, Qing He, and Xing Xie. PENS: A dataset and generic framework for personalized news headline generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pp. 82–92, Online, August 2021. Association for Computational Linguistics. doi: 10.18653/v1/2021.acl-long.7. URL https: //aclanthology.org/2021.acl-long.7.

Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, T. J. Henighan, Nicholas Joseph, Saurav Kadavath, John Kernion, Tom Conerly, Sheer El-Showk, Nelson Elhage, Zac Hatfield-Dodds, Danny Hernandez, Tristan Hume, Scott Johnston, Shauna Kravec, Liane Lovitt, Neel Nanda, Catherine Olsson, Dario Amodei, Tom B. Brown, Jack Clark, Sam McCandlish, Christopher Olah, Benjamin Mann, and Jared Kaplan. Training a helpful and harmless assistant with reinforcement learning from human feedback. ArXiv, abs/2204.05862, 2022a. URL https://api.semanticscholar.org/ CorpusID:248118878.

Yuntao Bai, Saurav Kadavath, Sandipan Kundu, Amanda Askell, Jackson Kernion, Andy Jones, Anna Chen, Anna Goldie, Azalia Mirhoseini, Cameron McKinnon, et al. Constitutional ai: Harmlessness from ai feedback. arXiv preprint arXiv:2212.08073, 2022b.

Stephen Casper, Xander Davies, Claudia Shi, Thomas Krendl Gilbert, Jer´ emy Scheurer, Javier ´ Rando, Rachel Freedman, Tomasz Korbak, David Lindner, Pedro Freire, et al. Open problems and fundamental limitations of reinforcement learning from human feedback. arXiv preprint arXiv:2307.15217, 2023.

Deborah Cohen, Moonkyung Ryu, Yinlam Chow, Orgad Keller, Ido Greenberg, Avinatan Hassidim, Michael Fink, Yossi Matias, Idan Szpektor, Craig Boutilier, et al. Dynamic planning in openended dialogue using reinforcement learning. arXiv preprint arXiv:2208.02294, 2022.

Shachar Don-Yehiya, Elad Venezian, Colin Raffel, Noam Slonim, Yoav Katz, and Leshem Choshen. Cold fusion: Collaborative descent for distributed multitask finetuning. arXiv preprint arXiv:2212.01378, 2022.

Yann Dubois, Xuechen Li, Rohan Taori, Tianyi Zhang, Ishaan Gulrajani, Jimmy Ba, Carlos Guestrin, Percy Liang, and Tatsunori B Hashimoto. Alpacafarm: A simulation framework for methods that learn from human feedback. arXiv preprint arXiv:2305.14387, 2023.

Shangbin Feng, Chan Young Park, Yuhan Liu, and Yulia Tsvetkov. From pretraining data to language models to downstream tasks: Tracking the trails of political biases leading to unfair nlp models. arXiv preprint arXiv:2305.08283, 2023.

Xinyang Geng, Arnav Gudibande, Hao Liu, Eric Wallace, Pieter Abbeel, Sergey Levine, and Dawn Song. Koala: A dialogue model for academic research. Blog post, April, 1, 2023.

Conor F Hayes, Roxana Radulescu, Eugenio Bargiacchi, Johan K ˘ allstr ¨ om, Matthew Macfarlane, ¨ Mathieu Reymond, Timothy Verstraeten, Luisa M Zintgraf, Richard Dazeley, Fredrik Heintz, et al. A practical guide to multi-objective reinforcement learning and planning. Autonomous Agents and Multi-Agent Systems, 36(1):26, 2022.

Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. URL https://openreview.net/forum? id=nZeVKeeFYf9.

Chengsong Huang, Qian Liu, Bill Yuchen Lin, Tianyu Pang, Chao Du, and Min Lin. Lorahub: Efficient cross-task generalization via dynamic lora composition. ArXiv, abs/2307.13269, 2023. URL https://api.semanticscholar.org/CorpusID:260155012.

Gabriel Ilharco, Marco Tulio Ribeiro, Mitchell Wortsman, Suchin Gururangan, Ludwig Schmidt, Hannaneh Hajishirzi, and Ali Farhadi. Editing models with task arithmetic. ICLR, 2022.

Joel Jang, Seungone Kim, Seonghyeon Ye, Doyoung Kim, Lajanugen Logeswaran, Moontae Lee, Kyungjae Lee, and Minjoon Seo. Exploring the benefits of training expert language models over instruction tuning. In Andreas Krause, Emma Brunskill, Kyunghyun Cho, Barbara Engelhardt, Sivan Sabato, and Jonathan Scarlett (eds.), Proceedings of the 40th International Conference on Machine Learning, volume 202 of Proceedings of Machine Learning Research, pp. 14702–14729. PMLR, 23–29 Jul 2023. URL https://proceedings.mlr.press/v202/jang23a. html.

Sungdong Kim, Sanghwan Bae, Jamin Shin, Soyoung Kang, Donghyun Kwak, Kang Min Yoo, and Minjoon Seo. Aligning large language models through synthetic feedback. arXiv preprint arXiv:2305.13735, 2023.

Hannah Rose Kirk, Bertie Vidgen, Paul Rottger, and Scott A Hale. Personalisation within bounds: A ¨ risk taxonomy and policy framework for the alignment of large language models with personalised feedback. arXiv preprint arXiv:2303.05453, 2023.

Noriyuki Kojima, Alane Suhr, and Yoav Artzi. Continual learning for grounded instruction generation by observing human following behavior. Transactions of the Association for Computational Linguistics, 9:1303–1319, 2021.

Margaret Li, Suchin Gururangan, Tim Dettmers, Mike Lewis, Tim Althoff, Noah A Smith, and Luke Zettlemoyer. Branch-train-merge: Embarrassingly parallel training of expert language models. arXiv preprint arXiv:2208.03306, 2022.

Pan Li and Alexander Tuzhilin. Towards controllable and personalized review generation. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pp. 3237–3245, Hong Kong, China, November 2019. Association for Computational Linguistics. doi: 10.18653/v1/D19-1319. URL https://aclanthology.org/D19-1319.

Bodhisattwa Prasad Majumder, Shuyang Li, Jianmo Ni, and Julian McAuley. Generating personalized recipes from historical user preferences. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pp. 5976–5982, Hong Kong, China, November 2019. Association for Computational Linguistics. doi: 10.18653/v1/D19-1613. URL https://aclanthology.org/D19-1613.

Pierre-Emmanuel Mazare, Samuel Humeau, Martin Raison, and Antoine Bordes. Training mil- ´ lions of personalized dialogue agents. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pp. 2775–2779, Brussels, Belgium, OctoberNovember 2018. Association for Computational Linguistics. doi: 10.18653/v1/D18-1298. URL https://aclanthology.org/D18-1298.

Jacob Menick, Maja Trebacz, Vladimir Mikulik, John Aslanides, Francis Song, Martin Chadwick, Mia Glaese, Susannah Young, Lucy Campbell-Gillingham, Geoffrey Irving, et al. Teaching language models to support answers with verified quotes. arXiv preprint arXiv:2203.11147, 2022.

Reiichiro Nakano, Jacob Hilton, S. Arun Balaji, Jeff Wu, Ouyang Long, Christina Kim, Christopher Hesse, Shantanu Jain, Vineet Kosaraju, William Saunders, Xu Jiang, Karl Cobbe, Tyna Eloundou, Gretchen Krueger, Kevin Button, Matthew Knight, Benjamin Chess, and John Schulman. Webgpt: Browser-assisted question-answering with human feedback. ArXiv, abs/2112.09332, 2021a. URL https://api.semanticscholar.org/CorpusID:245329531.

Reiichiro Nakano, Jacob Hilton, Suchir Balaji, Jeff Wu, Long Ouyang, Christina Kim, Christopher Hesse, Shantanu Jain, Vineet Kosaraju, William Saunders, et al. Webgpt: Browser-assisted question-answering with human feedback. arXiv preprint arXiv:2112.09332, 2021b.

Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke E. Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul Francis Christiano, Jan Leike, and Ryan J. Lowe. Training language models to follow instructions with human feedback. ArXiv, abs/2203.02155, 2022a. URL https://api.semanticscholar.org/ CorpusID:246426909.

Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems, 35: 27730–27744, 2022b.

Baolin Peng, Chunyuan Li, Pengcheng He, Michel Galley, and Jianfeng Gao. Instruction tuning with gpt-4. arXiv preprint arXiv:2304.03277, 2023.

Jonas Pfeiffer, Naman Goyal, Xi Victoria Lin, Xian Li, James Cross, Sebastian Riedel, and Mikel Artetxe. Lifting the curse of multilinguality by pre-training modular transformers. arXiv preprint arXiv:2205.06266, 2022.

Alexandre Rame, Guillaume Couairon, Mustafa Shukor, Corentin Dancette, Jean-Baptiste Gaya, Laure Soulier, and Matthieu Cord. Rewarded soups: towards pareto-optimal alignment by interpolating weights fine-tuned on diverse rewards. arXiv preprint arXiv:2306.04488, 2023.

Alireza Salemi, Sheshera Mysore, Michael Bendersky, and Hamed Zamani. Lamp: When large language models meet personalization. arXiv preprint arXiv:2304.11406, 2023.

Shibani Santurkar, Esin Durmus, Faisal Ladhak, Cinoo Lee, Percy Liang, and Tatsunori Hashimoto. Whose opinions do language models reflect? arXiv preprint arXiv:2303.17548, 2023.

John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017.

Umer Siddique, Paul Weng, and Matthieu Zimmer. Learning fair policies in multi-objective (deep) reinforcement learning with average and discounted rewards. In International Conference on Machine Learning, pp. 8905–8915. PMLR, 2020.

David Silver, Satinder Singh, Doina Precup, and Richard S Sutton. Reward is enough. Artificial Intelligence, 299:103535, 2021.

Prasann Singhal, Tanya Goyal, Jiacheng Xu, and Greg Durrett. A long way to go: Investigating length correlations in rlhf. arXiv preprint arXiv:2310.03716, 2023.

Feifan Song, Bowen Yu, Minghao Li, Haiyang Yu, Fei Huang, Yongbin Li, and Houfeng Wang. Preference ranking optimization for human alignment. arXiv preprint arXiv:2306.17492, 2023.

Nisan Stiennon, Long Ouyang, Jeffrey Wu, Daniel Ziegler, Ryan Lowe, Chelsea Voss, Alec Radford, Dario Amodei, and Paul F Christiano. Learning to summarize with human feedback. Advances in Neural Information Processing Systems, 33:3008–3021, 2020.

Alane Suhr and Yoav Artzi. Continual learning for instruction following from realtime feedback. arXiv preprint arXiv:2212.09710, 2022.

Richard S Sutton and Andrew G Barto. Reinforcement learning: An introduction. MIT press, 2018.

Romal Thoppilan, Daniel De Freitas, Jamie Hall, Noam Shazeer, Apoorv Kulshreshtha, Heng-Tze Cheng, Alicia Jin, Taylor Bos, Leslie Baker, Yu Du, et al. Lamda: Language models for dialog applications. arXiv preprint arXiv:2201.08239, 2022.

Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee´ Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and ` efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023.

Peter Vamplew, Richard Dazeley, Cameron Foale, Sally Firmin, and Jane Mummery. Humanaligned artificial intelligence is a multiobjective problem. Ethics and Information Technology, 20:27–40, 2018.

Kristof Van Moffaert and Ann Nowe. Multi-objective reinforcement learning using sets of pareto ´ dominating policies. The Journal of Machine Learning Research, 15(1):3483–3512, 2014.

Kristof Van Moffaert, Madalina M Drugan, and Ann Nowe. Scalarized multi-objective reinforce- ´ ment learning: Novel design techniques. In 2013 IEEE symposium on adaptive dynamic programming and reinforcement learning (ADPRL), pp. 191–199. IEEE, 2013.

Yizhong Wang, Hamish Ivison, Pradeep Dasigi, Jack Hessel, Tushar Khot, Khyathi Raghavi Chandu, David Wadden, Kelsey MacMillan, Noah A Smith, Iz Beltagy, et al. How far can camels go? exploring the state of instruction tuning on open resources. arXiv preprint arXiv:2306.04751, 2023.

Mitchell Wortsman, Suchin Gururangan, Shen Li, Ali Farhadi, Ludwig Schmidt, Michael Rabbat, and Ari S Morcos. lo-fi: distributed fine-tuning without communication. arXiv preprint arXiv:2210.11948, 2022a.

Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, et al. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In International Conference on Machine Learning, pp. 23965–23998. PMLR, 2022b.

Jeff Wu, Long Ouyang, Daniel M Ziegler, Nisan Stiennon, Ryan Lowe, Jan Leike, and Paul Christiano. Recursively summarizing books with human feedback. arXiv preprint arXiv:2109.10862, 2021a.

Yuwei Wu, Xuezhe Ma, and Diyi Yang. Personalized response generation via generative split memory network. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pp. 1956–1970, 2021b.

Zeqiu Wu, Yushi Hu, Weijia Shi, Nouha Dziri, Alane Suhr, Prithviraj Ammanabrolu, Noah A Smith, Mari Ostendorf, and Hannaneh Hajishirzi. Fine-grained human feedback gives better rewards for language model training. arXiv preprint arXiv:2306.01693, 2023.

Jie Xu, Yunsheng Tian, Pingchuan Ma, Daniela Rus, Shinjiro Sueda, and Wojciech Matusik. Prediction-guided multi-objective reinforcement learning for continuous robot control. In International conference on machine learning, pp. 10607–10616. PMLR, 2020.

Jing Xu, Arthur Szlam, and Jason Weston. Beyond goldfish memory: Long-term open-domain conversation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 5180–5197, Dublin, Ireland, May 2022. Association for Computational Linguistics. doi: 10.18653/v1/2022.acl-long.356. URL https: //aclanthology.org/2022.acl-long.356.

Runzhe Yang, Xingyuan Sun, and Karthik Narasimhan. A generalized algorithm for multi-objective reinforcement learning and policy adaptation. Advances in neural information processing systems, 32, 2019.

Saizheng Zhang, Emily Dinan, Jack Urbanek, Arthur Szlam, Douwe Kiela, and Jason Weston. Personalizing dialogue agents: I have a dog, do you have pets too? arXiv preprint arXiv:1801.07243, 2018.

Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. arXiv preprint arXiv:2306.05685, 2023.

Yinhe Zheng, Guanyi Chen, Minlie Huang, Song Liu, and Xuan Zhu. Personalized dialogue generation with diversified traits. arXiv preprint arXiv:1901.09672, 2019.

Daniel M Ziegler, Nisan Stiennon, Jeffrey Wu, Tom B Brown, Alec Radford, Dario Amodei, Paul Christiano, and Geoffrey Irving. Fine-tuning language models from human preferences. arXiv preprint arXiv:1909.08593, 2019.

A DETAILS OF EVALUATION SETUP

Simulated pairwise evaluation We use a modified version of the GPT4 annotation prompt used by Dubois et al. (2023). We modify the criteria to perform the pairwise evaluation from general to a single preference dimension. We also provide 3 demonstrations: one scenario where there is a tie because both responses do not contain any notion of the preference (e.g. both responses do not show any signs of friendliness), one scenario where there is a clear winner, and one scenario where they are both good, but one is better than the other.

We recruited 24 crowd workers for our human evaluation. Figure 5 shows the interface used for human evaluation. We consider both the ‘Tie’ and ‘Both are bad’ options to be Ties.

B CRITERIA-WISE EVALUATION

The criteria-wise win rate (%) across all the methods are shown in Figure 6. The criteria-wise win rate was calculated by getting the average win-rate of the preference combinations that contained the specific preference dimension. For example, when calculating the criteria-wise win rate of ‘Elementary’, we got the average win-rate of the preference combinations that contained the ‘Elementary’ preference, which includes AAA, AAB, ABA, and ABB.

C THE FULL LIST OF EVALUATION PROMPTS

The full list of evaluation prompts used in our experiments are provided in Table 5

D DETAILED RESULTS FOR HUMAN EVALUATION AND GPT-4 EVALUATION

We provide detailed results (win / loss / tie for each of the preference combinations for our main experimental results. Table 6 shows the GPT4 evaluation and Table 7 shows the human evaluation results.

E EXAMPLES OF P-SOUPS TEXT GENERATIONS

Table 8 shows empirical examples of the text generated from each preference combination of the 16 preference combination experiments for the same prompt.