Adaptive Helpfulness-Harmlessness Alignment with Preference Vectors
Read the original on arXiv Machine Learning →The Flow has not summarised this story yet — read it at arXiv Machine Learning.
The Flow has not summarised this story yet — read it at arXiv Machine Learning.
arXiv:2509. 22851v4 Announce Type: replace-cross Abstract: Margin-based optimization is fundamental to improving generalization and robustness in classification tasks.
LambdaMix is a parameter‑efficient method that learns a family of policies conditioned on user preferences by jointly training a low‑rank adapter for each objective over a frozen base model. A single preference vector selects an adapter mixture to generate responses and weight normalized rewards, enabling on‑line reinforcement learning without further optimization at inference. Experiments on four models show LambdaMix outperforms baselines in helpfulness, harmlessness, math problem solving, and comment rewriting, while using fewer training completions and storing fewer adapter parameters.
arXiv:2410. 15595v4 Announce Type: replace Abstract: With the rapid advancement of large language models (LLMs), aligning policy models with human preferences has become increasingly critical.
The paper introduces Collaborative Reward Modeling (CRM), an online framework that uses cross-model peer review and curriculum learning to identify reliable preference pairs for training reward models (RMs). By having two RMs evaluate each other's sample selections and synchronizing their learning progress, CRM reduces the impact of noisy feedback, improving RM generalization by up to 9.94 points on RewardBench under 40% preference noise. The approach also extends to implicit-reward alignment, demonstrating robustness and broad applicability for learning from imperfect feedback.
The paper introduces GAP-DPO, a method for personalizing large language models by selecting preference pairs based on gradient alignment with user utility. It formalizes personalized preference learning as a geometry‑aligned optimization problem, showing that off‑policy sampling can shift DPO updates from error correction to reinforcement when preference margins align with utility gradients. Experiments demonstrate that GAP‑DPO improves stylistic fidelity, preference alignment, and overall generation quality over standard DPO variants.
The paper introduces a doubly robust preference optimization algorithm for reinforcement learning from human feedback (RLHF) in large language models. It addresses sensitivity to misspecifications in the preference model, reference policy, or reward function by ensuring consistency when either the preference model or the reference policy is correctly specified. Experiments show that the method outperforms state‑of‑the‑art algorithms both theoretically and empirically.