Reinforcement learning from human feedback (RLHF) has become a cornerstone for aligning large language models with human preferences. However, the heterogeneity of human feedback, driven by diverse individual contexts and preferences, poses significant challenges for reward learning. To address this, we propose a Low-rank Contextual RLHF (LoCo-RLHF) framework that integrates contextual information to better model heterogeneous feedback while maintaining computational efficiency. Our approach builds on a contextual preference model, leveraging the intrinsic low-rank structure of the interaction between user contexts and query-answer pairs to mitigate the high dimensionality of feature representations. Furthermore, we address the challenge of distributional shifts in feedback through our Pessimism in Reduced Subspace (PRS) policy, inspired by pessimistic offline reinforcement learning techniques. We theoretically demonstrate that our policy achieves a tighter sub-optimality gap compared to existing methods. Extensive experiments validate the effectiveness of LoCo-RLHF, showcasing its superior performance in personalized RLHF settings and its robustness to distribution shifts.

The proliferation of Internet of Things (IoT) has increased interest in federated learning (FL) for privacy-preserving distributed data utilization. However, traditional two-tier FL architectures inadequately adapt to multi-tier IoT environments. While Hierarchical Federated Learning (HFL) improves practicality in multi-tier IoT environments by multi-layer aggregation, it still faces challenges in communication efficiency and accuracy due to high data transfer volumes, data heterogeneity, and imbalanced device distribution, struggling to meet the low-latency and high-accuracy model training requirements of practical IoT scenarios. To overcome these limitations, we propose H-FedSN, an innovative approach for practical IoT environments. H-FedSN introduces a binary mask mechanism with shared and personalized layers to reduce communication overhead by creating a sparse network while keeping original weights frozen. To address data heterogeneity and imbalanced device distribution, we integrate personalized layers for local data adaptation and apply Bayesian aggregation with cumulative Beta distribution updates at edge and cloud levels, effectively balancing contributions from diverse client groups. Evaluations on three real-world IoT datasets and MNIST under non-IID settings demonstrate that H-FedSN significantly reduces communication costs by 58 to 238 times compared to HierFAVG while achieving high accuracy, making it highly effective for practical IoT applications in hierarchical federated learning scenarios.

Knowledge distillation effectively reduces model complexity while improving performance, yet the underlying knowledge transfer mechanisms remain poorly understood. We propose novel spectral analysis methods and guidelines to optimize distillation, making the knowledge transfer process more interpretable. Our analysis reveals that CaiT models concentrate information in their first and last few layers, informing optimal layer selection for feature map distillation. Surprisingly, we discover that Swin Transformer and CaiT exhibit similar spectral encoding patterns despite their architectural differences, enhancing our understanding of transformer architectures and leading to improved feature map alignment strategies. Based on these insights, we introduce a simple yet effective spectral alignment method named SpectralKD. Experimental results demonstrate that following our guidelines enables SpectralKD to achieve state-of-the-art performance (DeiT-Tiny: $+5.2\%$, Swin-Tiny: $+1.4\%$ in ImageNet-1k Top-1 accuracy). Furthermore, through spectral analysis of student models trained with and without distillation, we show that distilled models mirror spectral patterns of their teachers, providing a new lens for interpreting knowledge distillation dynamics. Our code, pre-trained models, and experimental logs will be made publicly available.

What is appropriateness? Humans navigate a multi-scale mosaic of interlocking notions of what is appropriate for different situations. We act one way with our friends, another with our family, and yet another in the office. Likewise for AI, appropriate behavior for a comedy-writing assistant is not the same as appropriate behavior for a customer-service representative. What determines which actions are appropriate in which contexts? And what causes these standards to change over time? Since all judgments of AI appropriateness are ultimately made by humans, we need to understand how appropriateness guides human decision making in order to properly evaluate AI decision making and improve it. This paper presents a theory of appropriateness: how it functions in human society, how it may be implemented in the brain, and what it means for responsible deployment of generative AI technology.

We introduce a bootstrapping approach to train long-context language models by exploiting their short-context capabilities only. Our method utilizes a simple agent workflow to synthesize diverse long-context instruction tuning data, thereby eliminating the necessity for manual data collection and annotation. The proposed data synthesis workflow requires only a short-context language model, a text retriever, and a document collection, all of which are readily accessible within the open-source ecosystem. Subsequently, language models are fine-tuned using the synthesized data to extend their context lengths. In this manner, we effectively transfer the short-context capabilities of language models to long-context scenarios through a bootstrapping process. We conduct experiments with the open-source Llama-3 family of models and demonstrate that our method can successfully extend the context length to up to 1M tokens, achieving superior performance across various benchmarks.

Offline-to-online (O2O) reinforcement learning (RL) provides an effective means of leveraging an offline pre-trained policy as initialization to improve performance rapidly with limited online interactions. Recent studies often design fine-tuning strategies for a specific offline RL method and cannot perform general O2O learning from any offline method. To deal with this problem, we disclose that there are evaluation and improvement mismatches between the offline dataset and the online environment, which hinders the direct application of pre-trained policies to online fine-tuning. In this paper, we propose to handle these two mismatches simultaneously, which aims to achieve general O2O learning from any offline method to any online method. Before online fine-tuning, we re-evaluate the pessimistic critic trained on the offline dataset in an optimistic way and then calibrate the misaligned critic with the reliable offline actor to avoid erroneous update. After obtaining an optimistic and and aligned critic, we perform constrained fine-tuning to combat distribution shift during online learning. We show empirically that the proposed method can achieve stable and efficient performance improvement on multiple simulated tasks when compared to the state-of-the-art methods.

Our society increasingly benefits from Artificial Intelligence (AI). Unfortunately, more and more evidence shows that AI is also used for offensive purposes. Prior works have revealed various examples of use cases in which the deployment of AI can lead to violation of security and privacy objectives. No extant work, however, has been able to draw a holistic picture of the offensive potential of AI. In this SoK paper we seek to lay the ground for a systematic analysis of the heterogeneous capabilities of offensive AI. In particular we (i) account for AI risks to both humans and systems while (ii) consolidating and distilling knowledge from academic literature, expert opinions, industrial venues, as well as laypeople -- all of which being valuable sources of information on offensive AI. To enable alignment of such diverse sources of knowledge, we devise a common set of criteria reflecting essential technological factors related to offensive AI. With the help of such criteria, we systematically analyze: 95 research papers; 38 InfoSec briefings (from, e.g., BlackHat); the responses of a user study (N=549) entailing individuals with diverse backgrounds and expertise; and the opinion of 12 experts. Our contributions not only reveal concerning ways (some of which overlooked by prior work) in which AI can be offensively used today, but also represent a foothold to address this threat in the years to come.

Emergency response services are vital for enhancing public safety by safeguarding the environment, property, and human lives. As frontline members of these services, 9-1-1 dispatchers have a direct impact on response times and the overall effectiveness of emergency operations. However, traditional dispatcher training methods, which rely on role-playing by experienced personnel, are labor-intensive, time-consuming, and often neglect the specific needs of underserved communities. To address these challenges, we introduce Sim911, the first training simulation for 9-1-1 dispatchers powered by Large Language Models (LLMs). Sim911 enhances training through three key technical innovations: (1) knowledge construction, which utilizes archived 9-1-1 call data to generate simulations that closely mirror real-world scenarios; (2) context-aware controlled generation, which employs dynamic prompts and vector bases to ensure that LLM behavior aligns with training objectives; and (3) validation with looped correction, which filters out low-quality responses and refines the system performance.

Today's generative AI systems are tuned to present information by default rather than engage users in service of learning as a human tutor would. To address the wide range of potential education use cases for these systems, we reframe the challenge of injecting pedagogical behavior as one of \textit{pedagogical instruction following}, where training and evaluation examples include system-level instructions describing the specific pedagogy attributes present or desired in subsequent model turns. This framing avoids committing our models to any particular definition of pedagogy, and instead allows teachers or developers to specify desired model behavior. It also clears a path to improving Gemini models for learning -- by enabling the addition of our pedagogical data to post-training mixtures -- alongside their rapidly expanding set of capabilities. Both represent important changes from our initial tech report. We show how training with pedagogical instruction following produces a LearnLM model (available on Google AI Studio) that is preferred substantially by expert raters across a diverse set of learning scenarios, with average preference strengths of 31\% over GPT-4o, 11\% over Claude 3.5, and 13\% over the Gemini 1.5 Pro model LearnLM was based on.

Despite their expertise, AI developers don't always know what their most advanced systems are capable of--at least, not at first. To find out, systems are subjected to a range of tests--often called evaluations, or'evals'--designed to tease out their limits. But due to rapid progress in the field, today's systems regularly achieve top scores on many popular tests, including SATs and the U.S. bar exam, making it harder to judge just how quickly they are improving. A new set of much more challenging evals has emerged in response, created by companies, nonprofits, and governments. Yet even on the most advanced evals, AI systems are making astonishing progress.