Improving LLMs via Validator-to-Generator Alignment
Department of Computer Science, The University of Texas at Austin(德克萨斯大学奥斯汀分校计算机科学系);Departments of Linguistics and Computer Science, University of Massachusetts Amherst(马萨诸塞大学阿默斯特分校语言学与计算机科学系);Department of Computer Science & Center for Data Science, New York University(纽约大学计算机科学系及数据科学中心)
Juan Diego Rodriguez, Jocelyn Zhang, Katrin Erk, Greg Durrett
Large language models are inconsistent: varying prompts or including unrelated information can lead to unexpected changes in model outputs. The generator-validator (G-V) gap is one manifestation of this phenomenon, where LLMs generate responses that they then deem as invalid if re-queried to validate them. In this work, we introduce a new formulation of G-V consistency that involves a principled correction for utterance frequency. Specifically, generators often assign low likelihood to valid strings simply because those strings are a priori unlikely, which makes naive notions of G-V consistency unworkable. We show that under a natural model of rational agents answering questions with multiple answers, consistency of the validator with a frequency-corrected generator score emerges naturally. Our method, \emph{\FCPAname} (\FCPA), is a training objective implementing frequency-corrected G-V consistency for real-world LLMs. Our experimental results show that training with \FCPA{} substantially improves both G-V consistency and generator performance over prior methods, with gains of up to $+27$pp in Pearson correlation on IFEval and HumanEval, while preserving validator quality across all evaluated tasks.
PA-RAG: RAG Alignment via Multi-Perspective Preference Optimization
School of Data Science and Engineering, East China Normal University(华东师范大学数据科学与工程学院);Chinese Academy of Sciences(中国科学院);Baidu Inc(百度公司);Peking University(北京大学)
LIAR: Leveraging Inference Time Alignment (Best-of-N) to Jailbreak LLMs in Seconds
University of Central Florida(中央佛罗里达大学);University of Maryland, College Park(马里兰大学 College Park 分校);Princeton University(普林斯顿大学);Capital One(Capital One 公司)
GENERator: A Long-Context Generative Genomic Foundation Model
Equal Contribution 12pt Senior Authorship 0.15in 1Alibaba Cloud Computing, Beijing, China 0.05in 2Zhongguancun Academy, Beijing, China 0.05in 3Zhongguancun Institute of Artificial Intelligence, Beijing, China 0.05in 4University of Science;Technology of China, Hefei, China 0.05in 5State Key Laboratory of Genome;Multi-omics Technologies, Shenzhen Branch, Guangdong Laboratory for Lingnan Modern Agriculture, Key Laboratory of Livestock;Poultry Multi-Omics of MARA, Agricultural Genomics Institute at Shenzhen, Chinese Academy of Agricultural Sciences, Shenzhen, China 0.05in 6Innovation Group of Pig Genome Design;Breeding, Research Centre for Animal Genome, Agricultural Genomics Institute at Shenzhen, Chinese Academy of Agricultural Sciences, Shenzhen, China 0.05in 7The Hong Kong University of Science;Technology (Guangzhou), Guangzhou, China 0.05in 8The Hong Kong University of Science;Technology , Hong Kong SAR, China 0.05in 9Mila - Qu\'ebec AI Institute, Montr\'eal, Canada 0.05in 10University of Montr\'eal, Montr\'eal, Canada 0.05in 11HEC Montr\'eal, Montr\'eal, Canada 0.05in 12CIFAR AI Chair, Canada 0.05in 13Carnegie Mellon University, Pittsburgh, USA 0.15in to
Quantifying Retriever-Generator Alignment in RAG with Local Explanations
Department of Computer and Systems Sciences, Stockholm University(斯德哥尔摩大学计算机与系统科学系);BIFOLD, Technische Universität Berlin(柏林技术大学BIFOLD);Department of Informatics, Athens University of Economics and Business(雅典经济与商业大学信息系);Archimedes, Athena Research Centre(雅典研究中心Archimedes)
Internal Safety Collapse in LLMs and Coding Agents
Deakin University(德克萨斯大学);Institute of Trustworthy Embodied AI, Fudan University(复旦大学可信具身人工智能研究院);Shanghai Key Laboratory of Multimodal Embodied AI(上海多模态具身人工智能重点实验室);City University of Hong Kong(香港城市大学);The University of Melbourne(墨尔本大学);Singapore Management University(新加坡管理大学);University of Illinois at Urbana-Champaign(伊利诺伊大学厄巴纳-香槟分校)
NAST: A Non-Autoregressive Generator with Word Alignment for Unsupervised Text Style Transfer
Tsinghua University(清华大学);Institute for Artificial Intelligence(人工智能研究院);State Key Lab of Intelligent Technology and Systems(智能技术与系统国家重点实验室);Beijing National Research Center for Information Science and Technology(北京信息科学与技术国家研究中心)