Safeguard models help large language models (LLMs) detect and block harmful content, but most evaluations remain English-centric and overlook linguistic and cultural diversity. Existing multilingual safety benchmarks often rely on machine-translated English data, which fails to capture nuances in low-resource languages. Southeast Asian (SEA) languages are underrepresented despite the region's linguistic diversity and unique safety concerns, from culturally sensitive political speech to region-specific misinformation. Addressing these gaps requires benchmarks that are natively authored to reflect local norms and harm scenarios. We introduce SEA-SafeguardBench, the first human-verified safety benchmark for SEA, covering eight languages, 21,640 samples, across three subsets: general, in-the-wild, and content generation. The experimental results from our benchmark demonstrate that even state-of-the-art LLMs and guardrails are challenged by SEA cultural and harm scenarios and underperform when compared to English texts.

As large language models (LLMs) grow more capable, concerns about their safe deployment have also grown. Although alignment mechanisms have been introduced to deter misuse, they remain vulnerable to carefully designed adversarial prompts. In this work, we present a scalable attack strategy: intent-hiding adversarial prompting, which conceals malicious intent through the composition of skills. We develop a game-theoretic framework to model the interaction between such attacks and defense systems that apply both prompt and response filtering. Our analysis identifies equilibrium points and reveals structural advantages for the attacker. To counter these threats, we propose and analyze a defense mechanism tailored to intent-hiding attacks. Empirically, we validate the attack's effectiveness on multiple real-world LLMs across a range of malicious behaviors, demonstrating clear advantages over existing adversarial prompting techniques.

Instruction Tuning has emerged as a pivotal post-training paradigm that enables pre-trained language models to better follow user instructions. Despite its significance, little attention has been given to optimizing the loss function used. A fundamental, yet often overlooked, question is whether the conventional auto-regressive objective - where loss is computed only on response tokens, excluding prompt tokens - is truly optimal for instruction tuning. In this work, we systematically investigate the impact of differentially weighting prompt and response tokens in instruction tuning loss, and propose Weighted Instruction Tuning (WIT) as a better alternative to conventional instruction tuning. Through extensive experiments on five language models of different families and scale, three finetuning datasets of different sizes, and five diverse evaluation benchmarks, we show that the standard instruction tuning loss often yields suboptimal performance and limited robustness to input prompt variations. We find that a low-to-moderate weight for prompt tokens coupled with a moderate-to-high weight for response tokens yields the best-performing models across settings and also serve as better starting points for the subsequent preference alignment training. These findings highlight the need to reconsider instruction tuning loss and offer actionable insights for developing more robust and generalizable models. Our code is open-sourced at https://github.com/kowndinya-renduchintala/WIT.

--We present a novel approach for detecting hallucinations in large language models (LLMs) by analyzing the probabilistic divergence between prompt and response hidden-state distributions. Counterintuitively, we find that hallucinated responses exhibit smaller deviations from their prompts compared to grounded responses, suggesting that hallucinations often arise from superficial rephrasing rather than substantive reasoning. T o enhance sensitivity, we employ deep learn-able kernels that automatically adapt to capture nuanced geometric differences between distributions. Our approach outperforms existing baselines, demonstrating state-of-the-art performance on several benchmarks. The method remains competitive even without kernel training, offering a robust, scalable solution for hallucination detection. In recent years, large language models (LLMs) have been widely adopted in many applications. However, they often generate hallucinations -- incorrect or fabricated content that does not match real-world facts or the provided context [1]. The latter case is of special interest, as it refers to incorrect generations in retrieval-augmented generation (RAG) settings, where LLMs rely on retrieved information to answer user queries.

Peter Kyle, the UK's secretary of state for science, innovation and technology, has said he uses ChatGPT to understand difficult concepts The UK's technology secretary, Peter Kyle, has asked ChatGPT for advice on why the adoption of artificial intelligence is so slow in the UK business community – and which podcasts he should appear on. This week, Prime Minister Keir Starmer said that the UK government should be making far more use of AI in an effort to increase efficiency. "No person's substantive time should be spent on a task where digital or AI can do it better, quicker and to the same high quality and standard," he said. Now, New Scientist has obtained records of Kyle's ChatGPT use under the Freedom of Information (FOI) Act, in what is believed to be a world-first test of whether chatbot interactions are subject to such laws. These records show that Kyle asked ChatGPT to explain why the UK's small and medium business (SMB) community has been so slow to adopt AI.

In addition, researchers have explored how LLMs can be In early 2024, Volkswagen premiered the first vehicle with used in vision-and-language navigation (VLN). In robotics, ChatGPT integrated into its voice assistant (Volkswagen VLN involves giving robots or agents verbal instructions on 2024). Volkswagen claimed that its ChatGPT-enabled voice how to navigate using visual cues and landmarks (Schumann assistant could be used to control the infotainment, navigation, et al. 2024). Challenges of VLN include visual and natural and air conditioning, or to answer general knowledge language understanding, as well as spatial and temporal questions (Volkswagen 2024).

Large Language Models (LLMs) like GPT-4 and LLaMA have shown incredible proficiency at natural language processing tasks and have even begun to excel at tasks across other modalities such as vision and audio. Despite their success, LLMs often struggle to perform well on low-resource languages because there is so little training data available. This shortcoming is especially prevalent with open source models. In this work, we explore training LLaMA-2 to speak Amharic, a language which is spoken by over 50 million people world wide, but has orders of magnitude less data available than languages like English. We employ methods previously used for training LLMs on other languages with data scarcity, and use open source translation models to perform data augmentation and grow our dataset from millions of tokens to billions. We further enhance the capabilities of our model by connecting an image encoder and training on a translated visual instruction tuning dataset in the same manner as LLaVA, resulting in a multimodal Amharic LLM that can understand images along with text. We introduce an Amharic version of a popular benchmarking dataset to evaluate our work. Our models and datasets are open source and available on GitHub.

Red teaming is a common strategy for identifying weaknesses in generative language models (LMs), where adversarial prompts are produced that trigger an LM to generate unsafe responses. Red teaming is instrumental for both model alignment and evaluation, but is labor-intensive and difficult to scale when done by humans. In this paper, we present Gradient-Based Red Teaming (GBRT), a red teaming method for automatically generating diverse prompts that are likely to cause an LM to output unsafe responses. GBRT is a form of prompt learning, trained by scoring an LM response with a safety classifier and then backpropagating through the frozen safety classifier and LM to update the prompt. To improve the coherence of input prompts, we introduce two variants that add a realism loss and fine-tune a pretrained model to generate the prompts instead of learning the prompts directly. Our experiments show that GBRT is more effective at finding prompts that trigger an LM to generate unsafe responses than a strong reinforcement learning-based red teaming approach, and succeeds even when the LM has been fine-tuned to produce safer outputs.

Foundation models, specifically Large Language Models (LLMs), have lately gained wide-spread attention and adoption. Reinforcement Learning with Human Feedback (RLHF) involves training a reward model to capture desired behaviors, which is then used to align LLM's. These reward models are additionally used at inference-time to estimate LLM responses' adherence to those desired behaviors. However, there is little work measuring how robust these reward models are to distribution shifts. In this work, we evaluate how reward model performance - measured via accuracy and calibration (i.e. alignment between accuracy and confidence) - is affected by distribution shift. We show novel calibration patterns and accuracy drops due to OOD prompts and responses, and that the reward model is more sensitive to shifts in responses than prompts. Additionally, we adapt an OOD detection technique commonly used in classification to the reward model setting to detect these distribution shifts in prompts and responses.

The open-source publishing of large language models (LLMs) has created many possibilities for how anyone who understands language and has access to a computer can interact with significant tools of artificial intelligence, particularly in the context of learning and knowledge dissemination. However, the utility of these models in specialized fields like Classics is still largely unexplored. This project is an attempt to merge the knowledge of Classics with the capabilities of artificial intelligence by finetuning an LLM to cater to the specific needs of learners and professionals. The goal of this project is to develop an LLM that not only reproduces contextual knowledge accurately but also exhibits a consistent "personality" - and, indeed, has consistent propriety - to appeal to a diverse audience who possess differing levels of knowledge. A significant portion of this project was dedicated to refining the dataset, following the principle of "garbage in, garbage out," to ensure the model generates relevant, useful, and creative responses when given a prompt (a statement, question, or single word). After training and evaluation, my model's ability to handle a vast array of different types of inputs and prompting exceeded expectations for a 355M parameter model, though its occasional hallucinations (especially when set with a high temperature), particularly in its assertions about historical events or its own identity, make it seem somewhat capricious and more work in the form of continuous finetuning will be undertaken.