Vision-language models (VLMs) can respond to queries about images in many languages. However, beyond language, culture affects how we see things. For example, individuals from Western cultures focus more on the central figure in an image while individuals from Eastern cultures attend more to scene context. In this work, we present a novel investigation that demonstrates and localizes VLMs' Western bias in image understanding. We evaluate large VLMs across subjective and objective visual tasks with culturally diverse images and annotations. We find that VLMs perform better on the Western subset than the Eastern subset of each task. Controlled experimentation tracing the source of this bias highlights the importance of a diverse language mix in text-only pre-training for building equitable VLMs, even when inference is performed in English. Moreover, while prompting in the language of a target culture can lead to reductions in bias, it is not a substitute for building AI more representative of the world's languages.

Despite the recent observation that large language models (LLMs) can store substantial factual knowledge, there is a limited understanding of the mechanisms of how they acquire factual knowledge through pretraining. This work addresses this gap by studying how LLMs acquire factual knowledge during pretraining. The findings reveal several important insights into the dynamics of factual knowledge acquisition during pretraining. First, counterintuitively, we observe that pretraining on more data shows no significant improvement in the model's capability to acquire and maintain factual knowledge. Next, there is a power-law relationship between training steps and forgetting of memorization and generalization of factual knowledge, and LLMs trained with duplicated training data exhibit faster forgetting. Third, training LLMs with larger batch sizes can enhance the models' robustness to forgetting. Overall, our observations suggest that factual knowledge acquisition in LLM pretraining occurs by progressively increasing the probability of factual knowledge presented in the pretraining data at each step. However, this increase is diluted by subsequent forgetting. Based on this interpretation, we demonstrate that we can provide plausible explanations for recently observed behaviors of LLMs, such as the poor performance of LLMs on long-tail knowledge and the benefits of deduplicating the pretraining corpus.

For modern artificial intelligence (AI) applications such as large language models (LLMs), the training paradigm has recently shifted to pre-training followed by fine-tuning. Furthermore, owing to dwindling open repositories of data and thanks to efforts to democratize access to AI models, pre-training is expected to increasingly migrate from the current centralized deployments to federated learning (FL) implementations. Meta-learning provides a general framework in which pre-training and fine-tuning can be formalized. Meta-learning-based personalized FL (meta-pFL) moves beyond basic personalization by targeting generalization to new agents and tasks. This paper studies the generalization performance of meta-pFL for a wireless setting in which the agents participating in the pre-training phase, i.e., meta-learning, are connected via a shared wireless channel to the server. Adopting over-the-air computing, we study the trade-off between generalization to new agents and tasks, on the one hand, and convergence, on the other hand. The trade-off arises from the fact that channel impairments may enhance generalization, while degrading convergence. Extensive numerical results validate the theory.

"Jailbreak" is a major safety concern of Large Language Models (LLMs), which occurs when malicious prompts lead LLMs to produce harmful outputs, raising issues about the reliability and safety of LLMs. Therefore, an effective evaluation of jailbreaks is very crucial to develop its mitigation strategies. However, our research reveals that many jailbreaks identified by current evaluations may actually be hallucinations-erroneous outputs that are mistaken for genuine safety breaches. This finding suggests that some perceived vulnerabilities might not represent actual threats, indicating a need for more precise red teaming benchmarks. To address this problem, we propose the $\textbf{B}$enchmark for reli$\textbf{AB}$ilit$\textbf{Y}$ and jail$\textbf{B}$reak ha$\textbf{L}$l$\textbf{U}$cination $\textbf{E}$valuation (BabyBLUE). BabyBLUE introduces a specialized validation framework including various evaluators to enhance existing jailbreak benchmarks, ensuring outputs are useful malicious instructions. Additionally, BabyBLUE presents a new dataset as an augmentation to the existing red teaming benchmarks, specifically addressing hallucinations in jailbreaks, aiming to evaluate the true potential of jailbroken LLM outputs to cause harm to human society.

Nowadays, the fields of code and natural language processing are evolving rapidly. In particular, models become better at processing long context windows -- supported context sizes have increased by orders of magnitude over the last few years. However, there is a shortage of benchmarks for code processing that go beyond a single file of context, while the most popular ones are limited to a single method. With this work, we aim to close this gap by introducing Long Code Arena, a suite of six benchmarks for code processing tasks that require project-wide context. These tasks cover different aspects of code processing: library-based code generation, CI builds repair, project-level code completion, commit message generation, bug localization, and module summarization. For each task, we provide a manually verified dataset for testing, an evaluation suite, and open-source baseline solutions based on popular LLMs to showcase the usage of the dataset and to simplify adoption by other researchers.

Text embedding is becoming an increasingly popular AI methodology, especially among businesses, yet the potential of text embedding models to be biased is not well understood. This paper examines the degree to which a selection of popular text embedding models are biased, particularly along gendered dimensions. More specifically, this paper studies the degree to which these models associate a list of given professions with gendered terms. The analysis reveals that text embedding models are prone to gendered biases but in varying ways. Although there are certain inter-model commonalities, for instance, greater association of professions like nurse, homemaker, and socialite with female identifiers, and greater association of professions like CEO, manager, and boss with male identifiers, not all models make the same gendered associations for each occupation. Furthermore, the magnitude and directionality of bias can also vary on a model-by-model basis and depend on the particular words models are prompted with. This paper demonstrates that gender bias afflicts text embedding models and suggests that businesses using this technology need to be mindful of the specific dimensions of this problem.

Understanding alignment techniques begins with comprehending zero-shot generalization brought by instruction tuning, but little of the mechanism has been understood. Existing work has largely been confined to the task level, without considering that tasks are artificially defined and, to LLMs, merely consist of tokens and representations. This line of research has been limited to examining transfer between tasks from a task-pair perspective, with few studies focusing on understanding zero-shot generalization from the perspective of the data itself. To bridge this gap, we first demonstrate through multiple metrics that zero-shot generalization during instruction tuning happens very early. Next, we investigate the facilitation of zero-shot generalization from both data similarity and granularity perspectives, confirming that encountering highly similar and fine-grained training data earlier during instruction tuning, without the constraints of defined "tasks", enables better generalization. Finally, we propose a more grounded training data arrangement method, Test-centric Multi-turn Arrangement, and show its effectiveness in promoting continual learning and further loss reduction. For the first time, we show that zero-shot generalization during instruction tuning is a form of similarity-based generalization between training and test data at the instance level. We hope our analysis will advance the understanding of zero-shot generalization during instruction tuning and contribute to the development of more aligned LLMs.

Large language models have achieved remarkable success on general NLP tasks, but they may fall short for domain-specific problems. Recently, various Retrieval-Augmented Large Language Models (RALLMs) are proposed to address this shortcoming. However, existing evaluation tools only provide a few baselines and evaluate them on various domains without mining the depth of domain knowledge. In this paper, we address the challenges of evaluating RALLMs by introducing the R-Eval toolkit, a Python toolkit designed to streamline the evaluation of different RAG workflows in conjunction with LLMs. Our toolkit, which supports popular built-in RAG workflows and allows for the incorporation of customized testing data on the specific domain, is designed to be user-friendly, modular, and extensible. We conduct an evaluation of 21 RALLMs across three task levels and two representative domains, revealing significant variations in the effectiveness of RALLMs across different tasks and domains. Our analysis emphasizes the importance of considering both task and domain requirements when choosing a RAG workflow and LLM combination. We are committed to continuously maintaining our platform at https://github.com/THU-KEG/R-Eval to facilitate both the industry and the researchers.

Reinforcement learning from human feedback (RLHF) is a promising solution to align large language models (LLMs) more closely with human values. Off-policy preference optimization, where the preference data is obtained from other models, is widely adopted due to its cost efficiency and scalability. However, off-policy preference optimization often suffers from a distributional gap between the policy used for data collection and the target policy, leading to suboptimal optimization. In this paper, we propose a novel strategy to mitigate this problem by simulating on-policy learning with off-policy preference data. Our Weighted Preference Optimization (WPO) method adapts off-policy data to resemble on-policy data more closely by reweighting preference pairs according to their probability under the current policy. This method not only addresses the distributional gap problem but also enhances the optimization process without incurring additional costs. We validate our method on instruction following benchmarks including Alpaca Eval 2 and MT-bench. WPO not only outperforms Direct Preference Optimization (DPO) by up to 5.6% on Alpaca Eval 2 but also establishes a remarkable length-controlled winning rate against GPT-4-turbo of 48.6% based on Llama-3-8B-Instruct, making it the strongest 8B model on the leaderboard. We will release the code and models at https://github.com/wzhouad/WPO.

Recent studies have demonstrated that In-Context Learning (ICL), through the use of specific demonstrations, can align Large Language Models (LLMs) with human preferences known as In-Context Alignment (ICA), indicating that models can comprehend human instructions without requiring parameter adjustments. However, the exploration of the mechanism and applicability of ICA remains limited. In this paper, we begin by dividing the context text used in ICA into three categories: format, system prompt, and example. Through ablation experiments, we investigate the effectiveness of each part in enabling ICA to function effectively. We then examine how variants in these parts impact the model's alignment performance. Our findings indicate that the example part is crucial for enhancing the model's alignment capabilities, with changes in examples significantly affecting alignment performance. We also conduct a comprehensive evaluation of ICA's zero-shot capabilities in various alignment tasks. The results indicate that compared to parameter fine-tuning methods, ICA demonstrates superior performance in knowledge-based tasks and tool-use tasks. However, it still exhibits certain limitations in areas such as multi-turn dialogues and instruction following.