The paper introduces BALIGN, a balanced data selection strategy designed to reduce catastrophic forgetting—referred to as the alignment tax—in large language models during preference-based alignment. By analyzing preference optimization gradients, the authors identify three data-centric features that influence parameter drift: the reference model's log-probability margin, token length differences between chosen and rejected responses, and TF‑IDF similarity to general capability corpora. BALIGN aggregates these features into a composite risk score to filter out high-risk preference samples, thereby preserving foundational capabilities while maintaining alignment gains with minimal computational overhead.
Aligning large language models to human preferences is crucial for real-world deployment but frequently incurs an alignment tax, leading to the catastrophic forgetting of pre-trained general capabilit...
arXiv:2609.05899v1 Announce Type: cross Abstract: Aligning large language models with human preferences remains a challenge, primarily due to the critical role of preference data quality in effective...
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.
InfiFPO is a new preference optimization technique for implicit model fusion that replaces the reference model in Direct Preference Optimization with a fused source model, preserving probability information and avoiding complex vocabulary alignment. By incorporating probability clipping and max‑margin fusion, it aligns a pivot model with human preferences while distilling knowledge from multiple source models. Experiments on 11 benchmarks show that InfiFPO consistently outperforms existing fusion and preference optimization methods, raising Phi‑4’s average score from 79.95 to 83.33 and improving performance in mathematics, coding, and reasoning tasks.
arXiv:2504. 06659v2 Announce Type: replace-cross Abstract: Despite advances in Preference Alignment (PA) for Large Language Models (LLMs), mainstream methods like reinforcement learning with human feedback face notable challenges.
arXiv:2606. 09850v1 Announce Type: new Abstract: Post-training alignment algorithms are predominantly evaluated as black boxes, obscuring how they reshape language models' internal computations.
arXiv:2609.06893v1 Announce Type: cross Abstract: Direct preference optimization DPO is a promising offline approach for aligning large language models (LLMs) due to its simplicity, computational eff...
arXiv:2509. 08022v3 Announce Type: replace-cross Abstract: Aligning large language models (LLMs) with diverse human values is essential for safe and effective deployment, yet existing benchmarks often overlook cultural and demographic variation.
The paper introduces Comparison-based Preference Optimization (ComPO), a zeroth-order method that aligns large language models with human preferences using comparison oracles instead of direct differentiable loss optimization. It provides theoretical convergence guarantees for both offline and online variants under smoothness, gradient sparsity, and oracle compatibility assumptions, and establishes performance bounds under local coverage and in-distribution reward accuracy. Experiments on several LLMs (Mistral, Llama, Gemma-2, Qwen3, Gemma-3) show that ComPO outperforms existing direct alignment methods, achieving higher length-controlled win rates and diagnostics that suggest mitigation of likelihood displacement.
arXiv:2607. 03248v1 Announce Type: cross Abstract: The alignment of large language models with human preferences is commonly achieved through Reinforcement Learning from Human Feedback or Direct Preference Optimization.
The paper investigates how preference tuning—optimizing language models with explicit preference signals—behaves when applied to new domains. It systematically compares five alignment objectives and several adaptation strategies, such as target‑domain supervised fine‑tuning and pseudo‑labeling, across summarization, question‑answering helpfulness, and safety tasks. Results show that while pseudo‑labeling reduces domain‑shift degradation, it also causes mode collapse, highlighting a trade‑off between generalization and diversity.
arXiv:2608.23149v1 Announce Type: cross Abstract: The alignment of Large Language Models heavily relies on English-centric high-quality preference data, which often leads to suboptimal performance in...