Investigating Regularization of Self-Play Language Models

• URL: http://arxiv.org/abs/2404.04291v1
• Date: Thu, 4 Apr 2024 05:38:44 GMT
• Title: Investigating Regularization of Self-Play Language Models
• Authors: Reda Alami, Abdalgader Abubaker, Mastane Achab, Mohamed El Amine Seddik, Salem Lahlou,
• Abstract summary: This paper explores the effects of various forms of regularization in the context of language model alignment via self-play. We show that the KL-based regularizer boils down to replacing the previous policy by its geometric mixture with the base policy inside of the SPIN loss function.
• Score: 5.162978497363913
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: This paper explores the effects of various forms of regularization in the context of language model alignment via self-play. While both reinforcement learning from human feedback (RLHF) and direct preference optimization (DPO) require to collect costly human-annotated pairwise preferences, the self-play fine-tuning (SPIN) approach replaces the rejected answers by data generated from the previous iterate. However, the SPIN method presents a performance instability issue in the learning phase, which can be mitigated by playing against a mixture of the two previous iterates. In the same vein, we propose in this work to address this issue from two perspectives: first, by incorporating an additional Kullback-Leibler (KL) regularization to stay at the proximity of the reference policy; second, by using the idea of fictitious play which smoothens the opponent policy across all previous iterations. In particular, we show that the KL-based regularizer boils down to replacing the previous policy by its geometric mixture with the base policy inside of the SPIN loss function. We finally discuss empirical results on MT-Bench as well as on the Hugging Face Open LLM Leaderboard.

Related papers

• Achieving Logarithmic Regret in KL-Regularized Zero-Sum Markov Games [53.447182734351]
  We develop and analyze algorithms that provably achieve improved sample efficiency under Reverse Kullback-Leibler (KL) regularization.<n>We study both two-player zero-sum Matrix games and Markov games: for Matrix games, we propose OMG, an algorithm based on best response sampling with optimistic bonuses, and extend this idea to Markov games through the algorithm SOMG.<n>Both algorithms achieve a logarithmic regret in $T$ that scales inversely with the KL regularization strength $beta$ in addition to the standard $widetildemathcalO(sqrtT)
  arXiv  Detail & Related papers  (2025-10-15T01:00:54Z)
• Game-Theoretic Regularized Self-Play Alignment of Large Language Models [54.593523736962]
  We show that our regularization method can improve the unregularized self-play significantly. We show that our regularization method can improve the unregularized self-play significantly.
  arXiv  Detail & Related papers  (2025-02-24T22:43:21Z)
• SePPO: Semi-Policy Preference Optimization for Diffusion Alignment [67.8738082040299]
  We propose a preference optimization method that aligns DMs with preferences without relying on reward models or paired human-annotated data. We validate SePPO across both text-to-image and text-to-video benchmarks.
  arXiv  Detail & Related papers  (2024-10-07T17:56:53Z)
• Bridging and Modeling Correlations in Pairwise Data for Direct Preference Optimization [75.1240295759264]
  We propose an effective framework for Bridging and Modeling Correlations in pairwise data, named BMC. We increase the consistency and informativeness of the pairwise preference signals through targeted modifications. We identify that DPO alone is insufficient to model these correlations and capture nuanced variations.
  arXiv  Detail & Related papers  (2024-08-14T11:29:47Z)
• WARP: On the Benefits of Weight Averaged Rewarded Policies [66.95013068137115]
  We introduce a novel alignment strategy named Weight Averaged Rewarded Policies (WARP) WARP merges policies in the weight space at three distinct stages. Experiments with GEMMA policies validate that WARP improves their quality and alignment, outperforming other open-source LLMs.
  arXiv  Detail & Related papers  (2024-06-24T16:24:34Z)
• From $r$ to $Q^*$: Your Language Model is Secretly a Q-Function [50.812404038684505]
  We show that we can derive DPO in the token-level MDP as a general inverse Q-learning algorithm, which satisfies the Bellman equation. We discuss applications of our work, including information elicitation in multi-turn dialogue, reasoning, agentic applications and end-to-end training of multi-model systems.
  arXiv  Detail & Related papers  (2024-04-18T17:37:02Z)
• Nash Learning from Human Feedback [86.09617990412941]
  We introduce an alternative pipeline for the fine-tuning of large language models using pairwise human feedback. We term this approach Nash learning from human feedback (NLHF) We present a novel algorithmic solution, Nash-MD, founded on the principles of mirror descent.
  arXiv  Detail & Related papers  (2023-12-01T19:26:23Z)
• Provable Offline Preference-Based Reinforcement Learning [95.00042541409901]
  We investigate the problem of offline Preference-based Reinforcement Learning (PbRL) with human feedback. We consider the general reward setting where the reward can be defined over the whole trajectory. We introduce a new single-policy concentrability coefficient, which can be upper bounded by the per-trajectory concentrability.
  arXiv  Detail & Related papers  (2023-05-24T07:11:26Z)
• Hypothesis Transfer in Bandits by Weighted Models [8.759884299087835]
  We consider the problem of contextual multi-armed bandits in the setting of hypothesis transfer learning. We show a re-weighting scheme for which we show a reduction in the regret over the classic Linear UCB when transfer is desired. We further extend this method to an arbitrary amount of source models, where the algorithm decides which model is preferred at each time step.
  arXiv  Detail & Related papers  (2022-11-14T14:13:02Z)
• Improving Variational Autoencoders with Density Gap-based Regularization [16.770753948524167]
  Variational autoencoders (VAEs) are one of the powerful unsupervised learning frameworks in NLP for latent representation learning and latent-directed generation. In practice, optimizing ELBo often leads the posterior distribution of all samples converge to the same degenerated local optimum, namely posterior collapse or KL vanishing. We introduce new training objectives to tackle both problems through a novel regularization based on the probabilistic density gap between the aggregated posterior distribution and the prior distribution.
  arXiv  Detail & Related papers  (2022-11-01T08:17:10Z)
• A Batch Normalized Inference Network Keeps the KL Vanishing Away [35.40781000297285]
  Variational Autoencoder (VAE) is widely used to approximate a model's posterior on latent variables. VAE often converges to a degenerated local optimum known as "posterior collapse"
  arXiv  Detail & Related papers  (2020-04-27T05:20:01Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.