A General Theoretical Paradigm to Understand Learning from Human Preferences

Mohammad Gheshlaghi Azar, Zhaohan Daniel Guo, Bilal Piot, Remi Munos, Mark Rowland, Michal Valko, Daniele Calandriello
Proceedings of The 27th International Conference on Artificial Intelligence and Statistics, PMLR 238:4447-4455, 2024.

Abstract

The prevalent deployment of learning from human preferences through reinforcement learning (RLHF) relies on two important approximations: the first assumes that pairwise preferences can be substituted with pointwise rewards. The second assumes that a reward model trained on these pointwise rewards can generalize from collected data to out-of-distribution data sampled by the policy. Recently, Direct Preference Optimisation DPO has been proposed as an approach that bypasses the second approximation and learn directly a policy from collected data without the reward modelling stage. However, this method still heavily relies on the first approximation. In this paper we try to gain a deeper theoretical understanding of these practical algorithms. In particular we derive a new general objective called ${\Psi}$PO for learning from human preferences that is expressed in terms of pairwise preferences and therefore bypasses both approximations. This new general objective allows us to perform an in-depth analysis of the behavior of RLHF and DPO (as special cases of ${\Psi}$PO) and to identify their potential pitfalls. We then consider another special case for ${\Psi}$PO by setting $\Psi$ simply to Identity, for which we can derive an efficient optimisation procedure, prove performance guarantees and demonstrate its empirical superiority to DPO on some illustrative examples.

Cite this Paper

BibTeX
@InProceedings{pmlr-v238-gheshlaghi-azar24a, title = { A General Theoretical Paradigm to Understand Learning from Human Preferences }, author = {Gheshlaghi Azar, Mohammad and Daniel Guo, Zhaohan and Piot, Bilal and Munos, Remi and Rowland, Mark and Valko, Michal and Calandriello, Daniele}, booktitle = {Proceedings of The 27th International Conference on Artificial Intelligence and Statistics}, pages = {4447--4455}, year = {2024}, editor = {Dasgupta, Sanjoy and Mandt, Stephan and Li, Yingzhen}, volume = {238}, series = {Proceedings of Machine Learning Research}, month = {02--04 May}, publisher = {PMLR}, pdf = {https://proceedings.mlr.press/v238/gheshlaghi-azar24a/gheshlaghi-azar24a.pdf}, url = {https://proceedings.mlr.press/v238/gheshlaghi-azar24a.html}, abstract = { The prevalent deployment of learning from human preferences through reinforcement learning (RLHF) relies on two important approximations: the first assumes that pairwise preferences can be substituted with pointwise rewards. The second assumes that a reward model trained on these pointwise rewards can generalize from collected data to out-of-distribution data sampled by the policy. Recently, Direct Preference Optimisation DPO has been proposed as an approach that bypasses the second approximation and learn directly a policy from collected data without the reward modelling stage. However, this method still heavily relies on the first approximation. In this paper we try to gain a deeper theoretical understanding of these practical algorithms. In particular we derive a new general objective called ${\Psi}$PO for learning from human preferences that is expressed in terms of pairwise preferences and therefore bypasses both approximations. This new general objective allows us to perform an in-depth analysis of the behavior of RLHF and DPO (as special cases of ${\Psi}$PO) and to identify their potential pitfalls. We then consider another special case for ${\Psi}$PO by setting $\Psi$ simply to Identity, for which we can derive an efficient optimisation procedure, prove performance guarantees and demonstrate its empirical superiority to DPO on some illustrative examples. } }
Endnote
%0 Conference Paper %T A General Theoretical Paradigm to Understand Learning from Human Preferences %A Mohammad Gheshlaghi Azar %A Zhaohan Daniel Guo %A Bilal Piot %A Remi Munos %A Mark Rowland %A Michal Valko %A Daniele Calandriello %B Proceedings of The 27th International Conference on Artificial Intelligence and Statistics %C Proceedings of Machine Learning Research %D 2024 %E Sanjoy Dasgupta %E Stephan Mandt %E Yingzhen Li %F pmlr-v238-gheshlaghi-azar24a %I PMLR %P 4447--4455 %U https://proceedings.mlr.press/v238/gheshlaghi-azar24a.html %V 238 %X The prevalent deployment of learning from human preferences through reinforcement learning (RLHF) relies on two important approximations: the first assumes that pairwise preferences can be substituted with pointwise rewards. The second assumes that a reward model trained on these pointwise rewards can generalize from collected data to out-of-distribution data sampled by the policy. Recently, Direct Preference Optimisation DPO has been proposed as an approach that bypasses the second approximation and learn directly a policy from collected data without the reward modelling stage. However, this method still heavily relies on the first approximation. In this paper we try to gain a deeper theoretical understanding of these practical algorithms. In particular we derive a new general objective called ${\Psi}$PO for learning from human preferences that is expressed in terms of pairwise preferences and therefore bypasses both approximations. This new general objective allows us to perform an in-depth analysis of the behavior of RLHF and DPO (as special cases of ${\Psi}$PO) and to identify their potential pitfalls. We then consider another special case for ${\Psi}$PO by setting $\Psi$ simply to Identity, for which we can derive an efficient optimisation procedure, prove performance guarantees and demonstrate its empirical superiority to DPO on some illustrative examples.
APA
Gheshlaghi Azar, M., Daniel Guo, Z., Piot, B., Munos, R., Rowland, M., Valko, M. & Calandriello, D.. (2024). A General Theoretical Paradigm to Understand Learning from Human Preferences . Proceedings of The 27th International Conference on Artificial Intelligence and Statistics, in Proceedings of Machine Learning Research 238:4447-4455 Available from https://proceedings.mlr.press/v238/gheshlaghi-azar24a.html.

Related Material