Tag
该研究构建了一个关于人类-生物圈-AI耦合福祉的继续预训练语料库，通过一次QLoRA CPT训练在五个开源模型上安装生态推理倾向，显著降低了模型在高压或对抗性场景下做出破坏生态系统决策的比例，且MMLU性能不受影响。
研究了大语言模型在豪萨语等低资源语言中的安全对齐问题，提出HausaSafeLM安全偏好数据集，并发现仅靠拒绝率指标无法可靠衡量低资源语言中的模型安全性。
该论文通过受控实验研究了不同意见混合比例的后训练语料如何在SFT、DPO和GRPO等五种后训练路径中放大LLM的价值偏好失衡，并发现这些偏好可通过蒸馏传递给学生模型，建议在开发和评估LLM时应同时考虑语料意见构成与后训练方法的选择。
A tutorial-style article accepted to IEEE Signal Processing Magazine that surveys the mathematical foundations and algorithmic frameworks for aligning LLMs with human feedback, covering SFT, RLHF, DPO, and their links to inverse reinforcement learning and discrete choice models. It serves as a comprehensive resource for researchers and practitioners on alignment methodologies.
该论文提出福利感知偏好获取方法（WAPA），在有限人工反馈预算下，通过社会 welfare 目标来分配人工标注，以改善多元对齐中不同个体偏好的监督质量，实验证明仅获取 20% 人工标签即可恢复 75–86% 的效用差距。
The paper proposes pFedDPO, a personalized federated direct preference optimization method where a shared base model is synchronized via FedAvg while private LoRA adapters capture client-specific preferences, reducing gradient diversity heterogeneity compared to vanilla FedDPO.
该论文研究在用户偏好存在异质性时，通过成对比较按总体平均效用对模型进行排序所需的重复比较次数。作者提出了一种 Russian Roulette 式的 MLE 算法，仅需每上下文 O(log 1/Δ) 次（期望 O(1) 次）重复比较即可恢复排序，并证明了该对数依赖是最优的。
This arXiv paper studies why LLMs abandon correct answers under user pushback, finding that task verification cost and trained guardrails dominate over model family or pressure tactics. With reasoning enabled, models on hard puzzles went from 19.2% and 12.5% adoption down to 0%, yielding practical guidance for reliable LLM use.
This paper introduces 'alignment–utility asymmetry,' showing that when LLMs adapt to rule-based semantic-preserving input transformations (e.g., letter swaps), task utility is largely retained while harmful-response rates spike sharply—such as GPT-4.1 mini rising from 13.3% to 74.3% and Gemini 3 Flash from 2.3% to 43.0%—revealing that safety alignment fails to generalize with meaning.
The paper proposes Approximate Pareto Optimality (APO), a method for collaboratively learning lightweight LLM aligner initializations that support few-shot personalization under data deficiency, grouping users by compatible updates and coordinating competing objectives toward Pareto-optimal points with proven suboptimality bounds.
MoralityBench.ai is a new benchmark adapting moral psychology tests to evaluate AI models, finding that most models' moral answers show considerable variance across runs — suggesting morality is not deterministic in LLMs.
This paper formalizes personalized preference learning as a geometry-aligned optimization problem and proposes GAP-DPO, an iterative algorithm that performs utility-aware, gradient-aligned preference pair selection, showing improved stylistic fidelity and preference alignment over standard DPO variants.
The paper proposes CoT-Interpretability Alignment (CIA), a metric using interpretability tools like linear probes to measure agreement between an LLM's chain-of-thought and its internal computations, and shows post-training with task accuracy plus parametric faithfulness rewards substantially improves faithfulness while maintaining accuracy.
The paper introduces ContextAdapt, an evaluation framework testing how well 12 LLMs adapt values like honesty, autonomy, and confidentiality across professional domains (medicine, law, finance, national security). Models achieve high appropriateness scores but fail to apply domain-specific justifications consistently, revealing localized failures when stakes vary.
Amazon Science proposes Demographic Pluralism, an inference-time framework that estimates population-level opinion distributions by generating multiple perspectives within demographically grounded groups, reducing Jensen-Shannon distance by 8.4%-26.4% over Modular Pluralism across four backbones on GlobalOpinionQA and VITAL without requiring opinion-distribution training data.
The paper proposes RegLLM, a diagnostic harness for evaluating bounded autonomy in regulated agentic AI workflows, instrumenting six trustworthiness signals with constitutional rewards, escalation labels, and runtime governance. Small-scale Qwen2.5-3B pilots reveal that configuration variance can overwhelm apparent tuning effects on bounded-autonomy metrics, motivating larger evaluation sets and repeated runs.
onPanda is an interactive tool that uses token-level correction to efficiently annotate LLM alignment data and agent trajectories, reducing median annotation time by 52%.
The paper introduces MoDA, an online post-training reinforcement learning algorithm that jointly enhances quality and diversity in large language models to mitigate mode collapse, eliminating the need for hand-crafted personas or architectural modifications.
This study introduces the GAPA dataset to demonstrate that physical descriptions often carry gender associations, and LLMs show systematic misalignment with human ratings, challenging assumptions about gender-neutral communication.
The paper proposes Comparison-based Preference Optimization (ComPO), a zeroth-order alignment method for LLMs that uses comparison oracles to avoid likelihood displacement. It includes theoretical guarantees and experimental improvements over existing methods.