Reward models (RMs) have driven the state-of-the-art performance of LLMs today by enabling the integration of human feedback into the language modeling process. However, RMs are primarily trained and evaluated in English, and their capabilities in multilingual settings remain largely understudied. In this work, we conduct a systematic evaluation of several reward models in multilingual settings. We first construct the first-of-its-kind multilingual RM evaluation benchmark, M-RewardBench, consisting of 2.87k preference instances for 23 typologically diverse languages, that tests the chat, safety, reasoning, and translation capabilities of RMs. We then rigorously evaluate a wide range of reward models on M-RewardBench, offering fresh insights into their performance across diverse languages. We identify a significant gap in RMs' performances between English and non-English languages and show that RM preferences can change substantially from one language to another. We also present several findings on how different multilingual aspects impact RM performance. Specifically, we show that the performance of RMs is improved with improved translation quality. Similarly, we demonstrate that the models exhibit better performance for high-resource languages. We release M-RewardBench dataset and the codebase in this study to facilitate a better understanding of RM evaluation in multilingual settings.

Recent advancements in large language models (LLMs) have led to their increased application across various tasks, with reinforcement learning from human feedback (RLHF) being a crucial part of their training to align responses with user intentions. In the RLHF process, a reward model is trained using responses preferences determined by human labelers or AI systems, which then refines the LLM through reinforcement learning. This work introduces weak supervision as a strategy to extend RLHF datasets and enhance reward model performance. Weak supervision employs noisy or imprecise data labeling, reducing reliance on expensive manually labeled data. By analyzing RLHF datasets to identify heuristics that correlate with response preference, we wrote simple labeling functions and then calibrated a label model to weakly annotate unlabeled data. Our evaluation show that while weak supervision significantly benefits smaller datasets by improving reward model performance, its effectiveness decreases with larger, originally labeled datasets. Additionally, using an LLM to generate and then weakly label responses offers a promising method for extending preference data.

Knowledge Graph (KG) is playing an increasingly important role in various AI systems. For e-commerce, an efficient and low-cost automated knowledge graph construction method is the foundation of enabling various successful downstream applications. In this paper, we propose a novel method for constructing structured product knowledge graphs from raw product images. The method cooperatively leverages recent advances in the vision-language model (VLM) and large language model (LLM), fully automating the process and allowing timely graph updates. We also present a human-annotated e-commerce product dataset for benchmarking product property extraction in knowledge graph construction. Our method outperforms our baseline in all metrics and evaluated properties, demonstrating its effectiveness and bright usage potential.

Large language models (LLMs) have shown great promise in machine translation, but they still struggle with contextually dependent terms, such as new or domain-specific words. This leads to inconsistencies and errors that are difficult to address. Existing solutions often depend on manual identification of such terms, which is impractical given the complexity and evolving nature of language. While Retrieval-Augmented Generation (RAG) could provide some assistance, its application to translation is limited by issues such as hallucinations from information overload. In this paper, we propose CRAT, a novel multi-agent translation framework that leverages RAG and causality-enhanced self-reflection to address these challenges. This framework consists of several specialized agents: the Unknown Terms Identification agent detects unknown terms within the context, the Knowledge Graph (KG) Constructor agent extracts relevant internal knowledge about these terms and retrieves bilingual information from external sources, the Causality-enhanced Judge agent validates the accuracy of the information, and the Translator agent incorporates the refined information into the final output. This automated process allows for more precise and consistent handling of key terms during translation. Our results show that CRAT significantly improves translation accuracy, particularly in handling context-sensitive terms and emerging vocabulary.

Deep learning-based expert models have reached superhuman performance in decision-making domains such as chess and Go. However, it is under-explored to explain or comment on given decisions although it is important for human education and model explainability. The outputs of expert models are accurate, but yet difficult to interpret for humans. On the other hand, large language models (LLMs) produce fluent commentary but are prone to hallucinations due to their limited decision-making capabilities. To bridge this gap between expert models and LLMs, we focus on chess commentary as a representative case of explaining complex decision-making processes through language and address both the generation and evaluation of commentary. We introduce Concept-guided Chess Commentary generation (CCC) for producing commentary and GPT-based Chess Commentary Evaluation (GCC-Eval) for assessing it. CCC integrates the decision-making strengths of expert models with the linguistic fluency of LLMs through prioritized, concept-based explanations. GCC-Eval leverages expert knowledge to evaluate chess commentary based on informativeness and linguistic quality. Experimental results, validated by both human judges and GCC-Eval, demonstrate that CCC generates commentary that is accurate, informative, and fluent.

Data science plays a critical role in clinical research, but it requires professionals with expertise in coding and medical data analysis. Large language models (LLMs) have shown great potential in supporting medical tasks and performing well in general coding tests. However, these tests do not assess LLMs' ability to handle data science tasks in medicine, nor do they explore their practical utility in clinical research. To address this, we developed a dataset consisting of 293 real-world data science coding tasks, based on 39 published clinical studies, covering 128 tasks in Python and 165 tasks in R. This dataset simulates realistic clinical research scenarios using patient data. Our findings reveal that cutting-edge LLMs struggle to generate perfect solutions, frequently failing to follow input instructions, understand target data, and adhere to standard analysis practices. Consequently, LLMs are not yet ready to fully automate data science tasks. We benchmarked advanced adaptation methods and found two to be particularly effective: chain-of-thought prompting, which provides a step-by-step plan for data analysis, which led to a 60% improvement in code accuracy; and self-reflection, enabling LLMs to iteratively refine their code, yielding a 38% accuracy improvement. Building on these insights, we developed a platform that integrates LLMs into the data science workflow for medical professionals. In a user study with five medical doctors, we found that while LLMs cannot fully automate coding tasks, they significantly streamline the programming process. We found that 80% of their submitted code solutions were incorporated from LLM-generated code, with up to 96% reuse in some cases. Our analysis highlights the potential of LLMs, when integrated into expert workflows, to enhance data science efficiency in clinical research.

Fine-tuning is a crucial paradigm for adapting pre-trained large language models to downstream tasks. Recently, methods like Low-Rank Adaptation (LoRA) have been shown to match the performance of fully fine-tuned models on various tasks with an extreme reduction in the number of trainable parameters. Even in settings where both methods learn similarly accurate models, \emph{are their learned solutions really equivalent?} We study how different fine-tuning methods change pre-trained models by analyzing the model's weight matrices through the lens of their spectral properties. We find that full fine-tuning and LoRA yield weight matrices whose singular value decompositions exhibit very different structure; moreover, the fine-tuned models themselves show distinct generalization behaviors when tested outside the adaptation task's distribution. More specifically, we first show that the weight matrices trained with LoRA have new, high-ranking singular vectors, which we call \emph{intruder dimensions}. Intruder dimensions do not appear during full fine-tuning. Second, we show that LoRA models with intruder dimensions, despite achieving similar performance to full fine-tuning on the target task, become worse models of the pre-training distribution and adapt less robustly to multiple tasks sequentially. Higher-rank, rank-stabilized LoRA models closely mirror full fine-tuning, even when performing on par with lower-rank LoRA models on the same tasks. These results suggest that models updated with LoRA and full fine-tuning access different parts of parameter space, even when they perform equally on the fine-tuned distribution. We conclude by examining why intruder dimensions appear in LoRA fine-tuned models, why they are undesirable, and how their effects can be minimized.

Supervised fine-tuning (SFT) and alignment of large language models (LLMs) are key steps in providing a good user experience. However, the concept of an appropriate alignment is inherently application-dependent, and current methods often rely on heuristic choices to drive the optimization. In this work, we formulate SFT and alignment as a constrained optimization problem, where the LLM is trained on a task while being required to meet application-specific requirements, without resorting to heuristics. To solve this, we propose Lagrange Large Language Models (L3Ms), which employ logarithmic barriers to enforce the constraints. This approach allows for the customization of L3Ms across diverse applications while avoiding heuristic-driven processes. We demonstrate experimentally the versatility and efficacy of L3Ms in achieving tailored alignments for various applications. Large language models (LLMs) are used for a wide range of tasks: as chatbots (Brown et al., 2020; OpenAI, 2024), for code generation (Ahmad et al., 2021; Wang et al., 2021; Rozière et al., 2024), for medical assistance (Yang et al., 2022; Moor et al., 2023), and more. The key ingredients for their impressive downstream performance are supervised fine-tuning (SFT) and alignment; the former fine-tunes the LLM to a task of interest, while the latter instills it with preferential properties. Arguably, the right combination of preferential properties is highly application/task-dependent. For instance, a scholar might want a chatbot to be honest and factual to assist with their work, whereas a fiction writer might prefer the opposite behavior to help create fantastical imaginary worlds.

Fine-tuning large language models (LLMs) to align with human preferences has become a critical challenge in artificial intelligence to ensure the safety of their deployment. Reinforcement Learning from Human Feedback (RLHF) has emerged as a dominant approach, significantly improving LLM performance as demonstrated by InstructGPT [Ouyang et al., 2022] and subsequent works. RLHF combines reward modeling to quantify human preferences and RL fine-tuning to adjust the LLM's output distribution, enhancing desired responses while suppressing unfavorable ones. While RLHF has shown promising results, it comes with significant extra post-training cost, and the aligned LLM may exhibit performance degeneration due to the alignment tax [Askell et al., 2021, OpenAI, 2023]. Alternatively, best-of-N (BoN) sampling has emerged as a simple and surprisingly effective technique to obtain high-quality outputs from an LLM [Stiennon et al., 2020]. In BoN sampling, multiple samples are drawn from an LLM, ranked according to a specific attribute, and the best one is selected.

We propose a novel approach that uses large language models (LLMs) to generate persona-driven conversations between Players and Non-Player Characters (NPC) in games. Showcasing the application of our methodology, we introduce the Minecraft Persona-driven Dialogue dataset (MCPDial). Starting with a small seed of expert-written conversations, we employ our method to generate hundreds of additional conversations. Each conversation in the dataset includes rich character descriptions of the player and NPC. The conversations are long, allowing for in-depth and extensive interactions between the player and NPC. MCPDial extends beyond basic conversations by incorporating canonical function calls (e.g. "Call find a resource on iron ore") between the utterances. Finally, we conduct a qualitative analysis of the dataset to assess its quality and characteristics.