# Online Distributionally Robust LLM Alignment via Regression to Relative Reward

arXiv: https://arxiv.org/abs/2509.19104 · HTML: https://mlsanity.com/paper/2509.19104/

## Key facts

- Cited by 2 papers in the map (data through 14 Oct 2026).
- Placed in LLM preference optimization methods, one of 10 directions in LLM alignment / RLHF / preference optimisation.
- Builds on 18 of its 70 references that are in the map.
- First posted on arXiv 23 Sep 2025.

## Placed in

- LLM alignment / RLHF / preference optimisation / [LLM preference optimization methods](https://mlsanity.com/alignment/llm-preference-optimization-methods/)

## Builds on (most cited first, 18 of 18)

- [Proximal Policy Optimization Algorithms](https://mlsanity.com/paper/1707.06347/) (2017)
- [Deep reinforcement learning from human preferences](https://mlsanity.com/paper/1706.03741/) (2017)
- [Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback](https://mlsanity.com/paper/2204.05862/) (2022)
- [Fine-Tuning Language Models from Human Preferences](https://mlsanity.com/paper/1909.08593/) (2019)
- [Open Problems and Fundamental Limitations of Reinforcement Learning from Human Feedback](https://mlsanity.com/paper/2307.15217/) (2023)
- [Understanding the Effects of RLHF on LLM Generalisation and Diversity](https://mlsanity.com/paper/2310.06452/) (2023)
- [Is DPO Superior to PPO for LLM Alignment? A Comprehensive Study](https://mlsanity.com/paper/2404.10719/) (2024)
- [What are the Statistical Limits of Offline RL with Linear Function Approximation?](https://mlsanity.com/paper/2010.11895/) (2020)
- [Provably Robust DPO: Aligning Language Models with Noisy Feedback](https://mlsanity.com/paper/2403.00409/) (2024)
- [MaxMin-RLHF: Alignment with Diverse Human Preferences](https://mlsanity.com/paper/2402.08925/) (2024)
- [Correcting the Mythos of KL-Regularization: Direct Alignment without Overoptimization via Chi-Squared Preference Optimization](https://mlsanity.com/paper/2407.13399/) (2024)
- [Towards Robust Alignment of Language Models: Distributionally Robustifying Direct Preference Optimization](https://mlsanity.com/paper/2407.07880/) (2024)
- [Overcoming Reward Overoptimization via Adversarial Policy Optimization with Lightweight Uncertainty Estimation](https://mlsanity.com/paper/2403.05171/) (2024)
- [Reward-Robust RLHF in LLMs](https://mlsanity.com/paper/2409.15360/) (2024)
- [Reward Model Learning vs. Direct Policy Optimization: A Comparative Analysis of Learning from Human Preferences](https://mlsanity.com/paper/2403.01857/) (2024)
- [A Baseline Analysis of Reward Models' Ability To Accurately Analyze Foundation Models Under Distribution Shift](https://mlsanity.com/paper/2311.14743/) (2023)
- [Distributionally Robust Reinforcement Learning with Human Feedback](https://mlsanity.com/paper/2503.00539/) (2025)
- [Beyond the Binary: Capturing Diverse Preferences With Reward Regularization](https://mlsanity.com/paper/2412.03822/) (2024)

## Built on by (most cited first, 2 of 2)

- [Injecting Distributional Awareness into MLLMs via Reinforcement Learning for Deep Imbalanced Regression](https://mlsanity.com/paper/2605.01402/) (2026)
- [Optimal Transport Meets Reinforcement Learning: A Survey](https://mlsanity.com/paper/2610.01413/) (2026)

## Abstract (arXiv)

Reinforcement Learning with Human Feedback (RLHF) has become crucial for aligning Large Language Models (LLMs) with human intent. However, existing offline RLHF approaches suffer from overoptimization, where language models degrade by overfitting inaccuracies and drifting from preferred behaviors observed during training. Distributionally robust optimization (DRO) is a natural solution, but existing DRO-DPO methods are sample-inefficient, ignore heterogeneous preferences, and lean on brittle heuristics. We introduce DRO-REBEL, a family of robust online REBEL updates built on type-p Wasserstein, Kullback-Leibler (KL), and chi2 ambiguity sets. Strong duality reduces each update to a relative-reward regression, retaining REBEL's scalability without PPO-style clipping or value networks. Under linear rewards, log-linear policies, and a standard coverage condition, we prove widetildeO(√d/n) bounds on squared parameter error, with sharper constants than prior DRO-DPO analyses, and give the first parametric widetildeO(d/n) rate for DRO-based alignment under preference shift, matching non-robust RLHF in benign regimes. Each divergence yields a tractable SGD-based algorithm: gradient regularization for Wasserstein, importance weighting for KL, and a 1-D dual solve for chi2. On Emotion Alignment, the ArmoRM multi-objective benchmark, and HH-Alignment, DRO-REBEL outperforms prior robust and non-robust baselines across unseen preference mixtures, model sizes, and dataset scales.