Large Multimodal Models (e.g., GPT-4, Gemini, Chameleon) have evolved into powerful tools with millions of users. However, they remain generic models and lack personalized knowledge of specific user concepts. Previous work has explored personalization for text generation, yet it remains unclear how these methods can be adapted to new modalities, such as image generation. In this paper, we introduce Yo'Chameleon, the first attempt to study personalization for large multimodal models. Given 3-5 images of a particular concept, Yo'Chameleon leverages soft-prompt tuning to embed subject-specific information to (i) answer questions about the subject and (ii) recreate pixel-level details to produce images of the subject in new contexts. Yo'Chameleon is trained with (i) a self-prompting optimization mechanism to balance performance across multiple modalities, and (ii) a ``soft-positive" image generation approach to enhance image quality in a few-shot setting.

A burgeoning area within reinforcement learning (RL) is the design of sequential decision-making agents centered around large language models (LLMs). While autonomous decision-making agents powered by modern LLMs could facilitate numerous real-world applications, such successes demand agents that are capable of data-efficient RL. One key obstacle to achieving data efficiency in RL is exploration, a challenge that we demonstrate many recent proposals for LLM agent designs struggle to contend with. Meanwhile, classic algorithms from the RL literature known to gracefully address exploration require technical machinery that can be challenging to operationalize in purely natural language settings. In this work, rather than relying on finetuning or in-context learning to coax LLMs into implicitly imitating a RL algorithm, we illustrate how LLMs can be used to explicitly implement an existing RL algorithm (Posterior Sampling for Reinforcement Learning) whose capacity for statistically-efficient exploration is already well-studied. We offer empirical results demonstrating how our LLM-based implementation of a known, data-efficient RL algorithm can be considerably more effective in natural language tasks that demand prudent exploration.

Trust in AI is undermined by the fact that there is no science that predicts -- or that can explain to the public -- when an LLM's output (e.g. ChatGPT) is likely to tip mid-response to become wrong, misleading, irrelevant or dangerous. With deaths and trauma already being blamed on LLMs, this uncertainty is even pushing people to treat their 'pet' LLM more politely to 'dissuade' it (or its future Artificial General Intelligence offspring) from suddenly turning on them. Here we address this acute need by deriving from first principles an exact formula for when a Jekyll-and-Hyde tipping point occurs at LLMs' most basic level. Requiring only secondary school mathematics, it shows the cause to be the AI's attention spreading so thin it suddenly snaps. This exact formula provides quantitative predictions for how the tipping-point can be delayed or prevented by changing the prompt and the AI's training. Tailored generalizations will provide policymakers and the public with a firm platform for discussing any of AI's broader uses and risks, e.g. as a personal counselor, medical advisor, decision-maker for when to use force in a conflict situation. It also meets the need for clear and transparent answers to questions like ''should I be polite to my LLM?''

Large Language Models (LLMs) excel in tasks such as retrieval and question answering but require updates to incorporate new knowledge and reduce inaccuracies and hallucinations. Traditional updating methods, like fine-tuning and incremental learning, face challenges such as overfitting and high computational costs. Knowledge Editing (KE) provides a promising alternative but often overlooks the Knowledge Element Overlap (KEO) phenomenon, where multiple triplets share common elements, leading to editing conflicts. We identify the prevalence of KEO in existing KE datasets and show its significant impact on current KE methods, causing performance degradation in handling such triplets. To address this, we propose a new formulation, Knowledge Set Editing (KSE), and introduce SetKE, a method that edits sets of triplets simultaneously. Experimental results demonstrate that SetKE outperforms existing methods in KEO scenarios on mainstream LLMs. Additionally, we introduce EditSet, a dataset containing KEO triplets, providing a comprehensive benchmark.

Large language models (LLMs) have revolutionized the field of artificial intelligence, demonstrating text understanding and generation capabilities approaching human levels. However, despite impressive results, the internal functioning mechanisms of these models largely remain a "black box." As Amodei [1] notes in his essay "The Urgency of Interpretability," researchers have limited understanding of why LLMs generate specific responses and how they arrive at their conclusions. This lack of transparency becomes increasingly problematic as LLMs begin to play central roles in economics, technology, and national security. Of particular concern are phenomena such as unpredictable hallucinations, extreme sensitivity to query formulations, and puzzling patterns in the probability distributions of generated tokens. These phenomena not only limit the reliability of LLMs in critical applications but also point to fundamental gaps in our understanding of their operation.

We propose Low-Rank Sparse Attention (Lorsa), a sparse replacement model of Transformer attention layers to disentangle original Multi Head Self Attention (MHSA) into individually comprehensible components. Lorsa is designed to address the challenge of attention superposition to understand attention-mediated interaction between features in different token positions. We show that Lorsa heads find cleaner and finer-grained versions of previously discovered MHSA behaviors like induction heads, successor heads and attention sink behavior (i.e., heavily attending to the first token). Lorsa and Sparse Autoencoder (SAE) are both sparse dictionary learning methods applied to different Transformer components, and lead to consistent findings in many ways. For instance, we discover a comprehensive family of arithmetic-specific Lorsa heads, each corresponding to an atomic operation in Llama-3.1-8B. Automated interpretability analysis indicates that Lorsa achieves parity with SAE in interpretability while Lorsa exhibits superior circuit discovery properties, especially for features computed collectively by multiple MHSA heads. We also conduct extensive experiments on architectural design ablation, Lorsa scaling law and error analysis.

Early exits (EEs) offer a promising approach to reducing computational costs and latency by dynamically terminating inference once a satisfactory prediction confidence on a data sample is achieved. Although many works integrate EEs into encoder-only Transformers, their application to decoder-only architectures and, more importantly, Mamba models, a novel family of state-space architectures in the LLM realm, remains insufficiently explored. This work introduces DYNAMAX, the first framework to exploit the unique properties of Mamba architectures for early exit mechanisms. We not only integrate EEs into Mamba but also repurpose Mamba as an efficient EE classifier for both Mamba-based and transformer-based LLMs, showcasing its versatility. Our experiments employ the Mistral 7B transformer compared to the Codestral 7B Mamba model, using data sets such as TruthfulQA, CoQA, and TriviaQA to evaluate computational savings, accuracy, and consistency. The results highlight the adaptability of Mamba as a powerful EE classifier and its efficiency in balancing computational cost and performance quality across NLP tasks. By leveraging Mamba's inherent design for dynamic processing, we open pathways for scalable and efficient inference in embedded applications and resource-constrained environments. This study underscores the transformative potential of Mamba in redefining dynamic computing paradigms for LLMs.

Access to high-quality medical data is often restricted due to privacy concerns, posing significant challenges for training artificial intelligence (AI) algorithms within Electronic Health Record (EHR) applications. In this study, prompt engineering with the GPT-4 API was employed to generate high-quality synthetic datasets aimed at overcoming this limitation. The generated data encompassed a comprehensive array of patient admission information, including healthcare provider details, hospital departments, wards, bed assignments, patient demographics, emergency contacts, vital signs, immunizations, allergies, medical histories, appointments, hospital visits, laboratory tests, diagnoses, treatment plans, medications, clinical notes, visit logs, discharge summaries, and referrals. To ensure data quality and integrity, advanced validation techniques were implemented utilizing models such as BERT's Next Sentence Prediction for sentence coherence, GPT-2 for overall plausibility, RoBERTa for logical consistency, autoencoders for anomaly detection, and conducted diversity analysis. Synthetic data that met all validation criteria were integrated into a comprehensive PostgreSQL database, serving as the data management system for the EHR application. This approach demonstrates that leveraging generative AI models with rigorous validation can effectively produce high-quality synthetic medical data, facilitating the training of AI algorithms while addressing privacy concerns associated with real patient data.

Red-teaming is a core part of the infrastructure that ensures that AI models do not produce harmful content. Unlike past technologies, the black box nature of generative AI systems necessitates a uniquely interactional mode of testing, one in which individuals on red teams actively interact with the system, leveraging natural language to simulate malicious actors and solicit harmful outputs. This interactional labor done by red teams can result in mental health harms that are uniquely tied to the adversarial engagement strategies necessary to effectively red team. The importance of ensuring that generative AI models do not propagate societal or individual harm is widely recognized -- one less visible foundation of end-to-end AI safety is also the protection of the mental health and wellbeing of those who work to keep model outputs safe. In this paper, we argue that the unmet mental health needs of AI red-teamers is a critical workplace safety concern. Through analyzing the unique mental health impacts associated with the labor done by red teams, we propose potential individual and organizational strategies that could be used to meet these needs, and safeguard the mental health of red-teamers. We develop our proposed strategies through drawing parallels between common red-teaming practices and interactional labor common to other professions (including actors, mental health professionals, conflict photographers, and content moderators), describing how individuals and organizations within these professional spaces safeguard their mental health given similar psychological demands. Drawing on these protective practices, we describe how safeguards could be adapted for the distinct mental health challenges experienced by red teaming organizations as they mitigate emerging technological risks on the new digital frontlines.

A person downloading a pre-trained model from the web should be aware of its biases. Existing approaches for bias identification rely on datasets containing labels for the task of interest, something that a non-expert may not have access to, or may not have the necessary resources to collect: this greatly limits the number of tasks where model biases can be identified. In this work, we present Classifier-to-Bias (C2B), the first bias discovery framework that works without access to any labeled data: it only relies on a textual description of the classification task to identify biases in the target classification model. This description is fed to a large language model to generate bias proposals and corresponding captions depicting biases together with task-specific target labels. A retrieval model collects images for those captions, which are then used to assess the accuracy of the model w.r.t. the given biases. C2B is training-free, does not require any annotations, has no constraints on the list of biases, and can be applied to any pre-trained model on any classification task. Experiments on two publicly available datasets show that C2B discovers biases beyond those of the original datasets and outperforms a recent state-of-the-art bias detection baseline that relies on task-specific annotations, being a promising first step toward addressing task-agnostic unsupervised bias detection.