We introduce a Riemannian optimistic online learning algorithm for Hadamard manifolds based on inexact implicit updates. Unlike prior work, our method can handle in-manifold constraints, and matches the best known regret bounds in the Euclidean setting with no dependence on geometric constants, like the minimum curvature. Building on this, we develop algorithms for g-convex, g-concave smooth min-max problems on Hadamard manifolds. Notably, one method nearly matches the gradient oracle complexity of the lower bound for Euclidean problems, for the first time.

We address the problem of best policy identification in preference-based reinforcement learning (PbRL), where learning occurs from noisy binary preferences over trajectory pairs rather than explicit numerical rewards. This approach is useful for post-training optimization of generative AI models during multi-turn user interactions, where preference feedback is more robust than handcrafted reward models. In this setting, learning is driven by both an offline preference dataset -- collected from a rater of unknown 'competence' -- and online data collected with pure exploration. Since offline datasets may exhibit out-of-distribution (OOD) biases, principled online data collection is necessary. To address this, we propose Posterior Sampling for Preference Learning ($\mathsf{PSPL}$), a novel algorithm inspired by Top-Two Thompson Sampling, that maintains independent posteriors over the true reward model and transition dynamics. We provide the first theoretical guarantees for PbRL in this setting, establishing an upper bound on the simple Bayesian regret of $\mathsf{PSPL}$. Since the exact algorithm can be computationally impractical, we also provide an approximate version that outperforms existing baselines.

The International Semantic Web Research School (ISWS) is a week-long intensive program designed to immerse participants in the field. This document reports a collaborative effort performed by ten teams of students, each guided by a senior researcher as their mentor, attending ISWS 2023. Each team provided a different perspective to the topic of creative AI, substantiated by a set of research questions as the main subject of their investigation. The 2023 edition of ISWS focuses on the intersection of Semantic Web technologies and Creative AI. ISWS 2023 explored various intersections between Semantic Web technologies and creative AI. A key area of focus was the potential of LLMs as support tools for knowledge engineering. Participants also delved into the multifaceted applications of LLMs, including legal aspects of creative content production, humans in the loop, decentralised approaches to multimodal generative AI models, nanopublications and AI for personal scientific knowledge graphs, commonsense knowledge in automatic story and narrative completion, generative AI for art critique, prompt engineering, automatic music composition, commonsense prototyping and conceptual blending, and elicitation of tacit knowledge. As Large Language Models and semantic technologies continue to evolve, new exciting prospects are emerging: a future where the boundaries between creative expression and factual knowledge become increasingly permeable and porous, leading to a world of knowledge that is both informative and inspiring.

Aligning Large Language Models (LLMs) with human values and away from undesirable behaviors (such as hallucination) has become increasingly important. Recently, steering LLMs towards a desired behavior via activation editing has emerged as an effective method to mitigate harmful generations at inference-time. Activation editing modifies LLM representations by preserving information from positive demonstrations (e.g., truthful) and minimising information from negative demonstrations (e.g., hallucinations). When these demonstrations come from a private dataset, the aligned LLM may leak private information contained in those private samples. In this work, we present the first study of aligning LLM behavior with private datasets. Our work proposes the Private Steering for LLM Alignment (PSA) algorithm to edit LLM activations with differential privacy (DP) guarantees. We conduct extensive experiments on seven different benchmarks with opensource LLMs of different sizes (0.5B to 7B) and model families (LlaMa, Qwen, Mistral and Gemma). Our results show that PSA achieves DP guarantees for LLM alignment with minimal loss in performance, including alignment metrics, openended text generation quality, and general-purpose reasoning. We also develop the first Membership Inference Attack (MIA) for evaluating and auditing the empirical privacy for the problem of LLM steering via activation editing. Our attack is tailored for activation editing and relies solely on the generated texts without their associated probabilities. Our experiments support the theoretical guarantees by showing improved guarantees for our PSA algorithm compared to several existing non-private techniques. LLMs often generate inaccurate, biased or even harmful information that violates human values and preferences (Rawte et al., 2023). In response, recent research has increasingly focused on aligning LLMs towards certain desired behaviors (Konen et al., 2024) while preventing potentially harmful and unsafe outcomes. This has led to the development of several techniques for aligning LLMs, such as Reinforcement Learning from Human Feedback (RLHF) (Ouyang et al., 2022), instruction tuning (Wei et al., 2022), In-Context Learning (ICL) (Dong et al., 2022), and prompt engineering (Cheng et al., 2024). Nevertheless, several challenges remain, including the lack of diverse and representative datasets for alignment (Liu et al., 2024c), difficulties in addressing out-of-distribution issues (Liu et al., 2024a), the choice of alignment strategy (Ivison et al., 2024) and the lack of interpretability in traditional alignment methods (Lee et al., 2024). The linear representation hypothesis (Park et al., 2024b) suggests that high-level concepts are linearly represented as directions in the representation space of LLMs.

We introduce the State Stream Transformer (SST), a novel LLM architecture that reveals emergent reasoning behaviours and capabilities latent in pretrained weights through addressing a fundamental limitation in traditional transformer models: the lack of latent computational continuity across autoregressive generations in the state space. SST introduces a sliding window latent state (FFN) cache with weighted decay that maintains and evolves persistent latent processes throughout autoregressive generations. Through controlled experiments comparing base and SST architectures using the same frozen weights, we demonstrate that this architectural modification alone enables enhanced reasoning capabilities which appear best explained by some form of potential higher-order processing, as evidenced by emergent metacognitive behaviours. These behaviours persist under controlled conditions designed to eliminate confounding factors such as stochastic variation or learned response patterns. Analysis of latent state distributions and processing dynamics provides evidence that it is solely the 'state stream' that is responsible for these phenomena. In quantitative evaluations, the SST achieves substantial performance improvements over the base model on two reasoning benchmarks, reaching 89.01\% accuracy on GSM-8K (0-shot) and 91.04\% on ARC Challenge (0-shot CoT). These findings indicate that persistent computation in the latent state space enables fundamentally different information processing and internal reasoning strategies, with implications for our understanding of artificial intelligence systems.

With the degradation of guardrails against mis- and disinformation online, it is more critical than ever to be able to effectively combat it. In this paper, we explore the efficiency and effectiveness of using crowd-sourced truthfulness assessments based on condensed, large language model (LLM) generated summaries of online sources. We compare the use of generated summaries to the use of original web pages in an A/B testing setting, where we employ a large and diverse pool of crowd-workers to perform the truthfulness assessment. We evaluate the quality of assessments, the efficiency with which assessments are performed, and the behavior and engagement of participants. Our results demonstrate that the Summary modality, which relies on summarized evidence, offers no significant change in assessment accuracy over the Standard modality, while significantly increasing the speed with which assessments are performed. Workers using summarized evidence produce a significantly higher number of assessments in the same time frame, reducing the cost needed to acquire truthfulness assessments. Additionally, the Summary modality maximizes both the inter-annotator agreements as well as the reliance on and perceived usefulness of evidence, demonstrating the utility of summarized evidence without sacrificing the quality of assessments.

With the advancement of large language models (LLMs), an increasing number of student models have leveraged LLMs to analyze textual artifacts generated by students to understand and evaluate their learning. These student models typically employ pre-trained LLMs to vectorize text inputs into embeddings and then use the embeddings to train models to detect the presence or absence of a construct of interest. However, how reliable and robust are these models at processing language with different levels of complexity? In the context of learning where students may have different language backgrounds with various levels of writing skills, it is critical to examine the robustness of such models to ensure that these models work equally well for text with varying levels of language complexity. Coincidentally, a few (but limited) research studies show that the use of language can indeed impact the performance of LLMs. As such, in the current study, we examined the robustness of several LLM-based student models that detect student self-regulated learning (SRL) in math problem-solving. Specifically, we compared how the performance of these models vary using texts with high and low lexical, syntactic, and semantic complexity measured by three linguistic measures.

Federated Learning (FL) enables clients to collaboratively train machine learning models without sharing local data, preserving privacy in diverse environments. While traditional FL approaches preserve privacy, they often struggle with high computational and communication overhead. To address these issues, model pruning is introduced as a strategy to streamline computations. However, existing pruning methods, when applied solely based on local data, often produce sub-models that inadequately reflect clients' specific tasks due to data insufficiency. To overcome these challenges, this paper introduces SAFL (Structure-Aware Federated Learning), a novel framework that enhances personalized federated learning through client-specific clustering and Similar Client Structure Information (SCSI)-guided model pruning. SAFL employs a two-stage process: initially, it groups clients based on data similarities and uses aggregated pruning criteria to guide the pruning process, facilitating the identification of optimal sub-models. Subsequently, clients train these pruned models and engage in server-based aggregation, ensuring tailored and efficient models for each client. This method significantly reduces computational overhead while improving inference accuracy. Extensive experiments demonstrate that SAFL markedly diminishes model size and improves performance, making it highly effective in federated environments characterized by heterogeneous data.

The logistic loss (a.k.a. cross-entropy loss) is one of the most popular loss functions used for multiclass classification. It is also the loss function of choice for next-token prediction in language modeling. It is associated with the Kullback--Leibler (KL) divergence and the softargmax operator. In this work, we propose to construct new convex loss functions based on $f$-divergences. Our loss functions generalize the logistic loss in two directions: i) by replacing the KL divergence with $f$-divergences and ii) by allowing non-uniform reference measures. We instantiate our framework for numerous $f$-divergences, recovering existing losses and creating new ones. By analogy with the logistic loss, the loss function generated by an $f$-divergence is associated with an operator, that we dub $f$-softargmax. We derive a novel parallelizable bisection algorithm for computing the $f$-softargmax associated with any $f$-divergence. On the empirical side, one of the goals of this paper is to determine the effectiveness of loss functions beyond the classical cross-entropy in a language model setting, including on pre-training, post-training (SFT) and distillation. We show that the loss function generated by the $\alpha$-divergence (which is equivalent to Tsallis $\alpha$-negentropy in the case of unit reference measures) with $\alpha=1.5$ performs well across several tasks.

Large Language Models (LLMs) demand significant computational resources, making it essential to enhance their capabilities without retraining from scratch. A key challenge in this domain is \textit{catastrophic forgetting} (CF), which hampers performance during Continuous Pre-training (CPT) and Continuous Supervised Fine-Tuning (CSFT). We propose \textbf{Control LLM}, a novel approach that leverages parallel pre-trained and expanded transformer blocks, aligning their hidden-states through interpolation strategies This method effectively preserves performance on existing tasks while seamlessly integrating new knowledge. Extensive experiments demonstrate the effectiveness of Control LLM in both CPT and CSFT. On Llama3.1-8B-Instruct, it achieves significant improvements in mathematical reasoning ($+14.4\%$ on Math-Hard) and coding performance ($+10\%$ on MBPP-PLUS). On Llama3.1-8B, it enhances multilingual capabilities ($+10.6\%$ on C-Eval, $+6.8\%$ on CMMLU, and $+30.2\%$ on CMMLU-0shot-CoT). It surpasses existing methods and achieves SOTA among open-source models tuned from the same base model, using substantially less data and compute. Crucially, these gains are realized while preserving strong original capabilities, with minimal degradation ($<4.3\% \text{on MMLU}$) compared to $>35\%$ in open-source Math and Coding models. This approach has been successfully deployed in LinkedIn's GenAI-powered job seeker and Ads unit products. To support further research, we release the training and evaluation code (https://github.com/linkedin/ControlLLM) along with models trained on public datasets (https://huggingface.co/ControlLLM) to the community.