Large language models (LLMs) encode vast amounts of pre-trained knowledge in their parameters, but updating them as real-world information evolves remains a challenge. Existing methodologies and benchmarks primarily target entity substitutions, failing to capture the full breadth of complex real-world dynamics. In this paper, we introduce Knowledge Update Playground (KUP), an automatic pipeline for simulating realistic knowledge updates reflected in an evidence corpora. KUP's evaluation framework includes direct and indirect probes to both test memorization of updated facts and reasoning over them, for any update learning methods. Next, we present a lightweight method called memory conditioned training (MCT), which conditions tokens in the update corpus on self-generated "memory" tokens during training. Our strategy encourages LLMs to surface and reason over newly memorized knowledge at inference. Our results on two strong LLMs show that (1) KUP benchmark is highly challenging, with the best CPT models achieving $<2\%$ in indirect probing setting (reasoning) and (2) MCT training significantly outperforms prior continued pre-training (CPT) baselines, improving direct probing (memorization) results by up to $25.4\%$.

Recent work suggests that preference-tuning techniques--including Reinforcement Learning from Human Preferences (RLHF) methods like PPO and GRPO, as well as alternatives like DPO--reduce diversity, creating a dilemma given that such models are widely deployed in applications requiring diverse outputs. To address this, we introduce a framework for measuring effective semantic diversity--diversity among outputs that meet quality thresholds--which better reflects the practical utility of large language models (LLMs). Using open-ended tasks that require no human intervention, we find counterintuitive results: although preference-tuned models--especially those trained via RL--exhibit reduced lexical and syntactic diversity, they produce greater effective semantic diversity than SFT or base models, not from increasing diversity among high-quality outputs, but from generating more high-quality outputs overall. We discover that preference tuning reduces syntactic diversity while preserving semantic diversity--revealing a distinction between diversity in form and diversity in content that traditional metrics often overlook. Our analysis further shows that smaller models are consistently more parameter-efficient at generating unique content within a fixed sampling budget, offering insights into the relationship between model scaling and diversity. These findings have important implications for applications that require diverse yet high-quality outputs, from creative assistance to synthetic data generation.

Although the internet has transformed the way we access informa tion, human navigation of the internet to find information is clunky for several reasons: (1) our m emory and world knowledge are limited; (2) our browsing abilities are hindered by distraction and fatig ue; and (3) human brains can only attend to one thing at a time and cannot be parallelized. Machine in telligence, on the other hand, has much more extensive recall and can operate tirelessly without g etting distracted. A sufficiently capable machine intelligence should be able to, in principle, retrieve any well-specified any piece of information from the open web, even if retrieving it would require bro wsing thousands of web pages. As AI progresses from chatbots to reasoners and to agents, th ere has been increased interest in models that can browse the internet beyond simple queries ( Google, 2024; OpenAI, 2025b, a; perplexity.AI, 2025; x.AI, 2025). While past benchmarks have measured the ability to retrieve information ( Joshi et al., 2017; Yang et al., 2018; Thorne et al., 2018; Dinan et al., 2019; Fan et al., 2019; Mialon et al., 2023), most of these benchmarks focus on retrieving information that ca n be found easily, and hence have become saturated by recent language models. Here we introduce a new benchmark called BrowseComp, which stands for "Browsing Competition" and comprises 1,266 challe nging problems that require browsing a large number of websites to solve. Three example questio ns are shown below.

In this paper, we advance the study of AI-augmented reasoning in the context of Human-Computer Interaction (HCI), psychology and cognitive science, focusing on the critical task of visual perception. Specifically, we investigate the applicability of Multimodal Large Language Models (MLLMs) in this domain. To this end, we leverage established principles and explanations from psychology and cognitive science related to complexity in human visual perception. We use them as guiding principles for the MLLMs to compare and interprete visual content. Our study aims to benchmark MLLMs across various explainability principles relevant to visual perception. Unlike recent approaches that primarily employ advanced deep learning models to predict complexity metrics from visual content, our work does not seek to develop a mere new predictive model. Instead, we propose a novel annotation-free analytical framework to assess utility of MLLMs as cognitive assistants for HCI tasks, using visual perception as a case study. The primary goal is to pave the way for principled study in quantifying and evaluating the interpretability of MLLMs for applications in improving human reasoning capability and uncovering biases in existing perception datasets annotated by humans.

Working with documents is a key part of almost any knowledge work, from contextualizing research in a literature review to reviewing legal precedent. Recently, as their capabilities have expanded, primarily text-based NLP systems have often been billed as able to assist or even automate this kind of work. But to what extent are these systems able to model these tasks as experts conceptualize and perform them now? In this study, we interview sixteen domain experts across two domains to understand their processes of document research, and compare it to the current state of NLP systems. We find that our participants processes are idiosyncratic, iterative, and rely extensively on the social context of a document in addition its content; existing approaches in NLP and adjacent fields that explicitly center the document as an object, rather than as merely a container for text, tend to better reflect our participants' priorities, though they are often less accessible outside their research communities. We call on the NLP community to more carefully consider the role of the document in building useful tools that are accessible, personalizable, iterative, and socially aware.

Recent advances in generative Artificial Intelligence have raised public awareness, shaping expectations and concerns about their societal implications. Central to these debates is the question of AI alignment -- how well AI systems meet public expectations regarding safety, fairness, and social values. However, little is known about what people expect from AI-enabled systems and how these expectations differ across national contexts. We present evidence from two surveys of public preferences for key functional features of AI-enabled systems in Germany (n = 1800) and the United States (n = 1756). We examine support for four types of alignment in AI moderation: accuracy and reliability, safety, bias mitigation, and the promotion of aspirational imaginaries. U.S. respondents report significantly higher AI use and consistently greater support for all alignment features, reflecting broader technological openness and higher societal involvement with AI. In both countries, accuracy and safety enjoy the strongest support, while more normatively charged goals -- like fairness and aspirational imaginaries -- receive more cautious backing, particularly in Germany. We also explore how individual experience with AI, attitudes toward free speech, political ideology, partisan affiliation, and gender shape these preferences. AI use and free speech support explain more variation in Germany. In contrast, U.S. responses show greater attitudinal uniformity, suggesting that higher exposure to AI may consolidate public expectations. These findings contribute to debates on AI governance and cross-national variation in public preferences. More broadly, our study demonstrates the value of empirically grounding AI alignment debates in public attitudes and of explicitly developing normatively grounded expectations into theoretical and policy discussions on the governance of AI-generated content.

Fine-tuning plays a crucial role in adapting models to downstream tasks with minimal training efforts. However, the rapidly increasing size of foundation models poses a daunting challenge for accommodating foundation model fine-tuning in most commercial devices, which often have limited memory bandwidth. Techniques like model sharding and tensor parallelism address this issue by distributing computation across multiple devices to meet memory requirements. Nevertheless, these methods do not fully leverage their foundation nature in facilitating the fine-tuning process, resulting in high computational costs and imbalanced workloads. We introduce a novel Distributed Dynamic Fine-Tuning (D2FT) framework that strategically orchestrates operations across attention modules based on our observation that not all attention modules are necessary for forward and backward propagation in fine-tuning foundation models. Through three innovative selection strategies, D2FT significantly reduces the computational workload required for fine-tuning foundation models. Furthermore, D2FT addresses workload imbalances in distributed computing environments by optimizing these selection strategies via multiple knapsack optimization. Our experimental results demonstrate that the proposed D2FT framework reduces the training computational costs by 40% and training communication costs by 50% with only 1% to 2% accuracy drops on the CIFAR-10, CIFAR-100, and Stanford Cars datasets. Moreover, the results show that D2FT can be effectively extended to recent LoRA, a state-of-the-art parameter-efficient fine-tuning technique. By reducing 40% computational cost or 50% communication cost, D2FT LoRA top-1 accuracy only drops 4% to 6% on Stanford Cars dataset.

In our paper we explore the definition, and extrapolation of fallacies as they pertain to the automatic detection of manipulation on social media. In particular we explore how these logical fallacies might appear in the real world i.e internet forums. We discovered a prevalence of misinformation / misguided intention in discussion boards specifically centered around the Ukrainian Russian Conflict which serves to narrow the domain of our task. Although automatic fallacy detection has gained attention recently, most datasets use unregulated fallacy taxonomies or are limited to formal linguistic domains like political debates or news reports. Online discourse, however, often features non-standardized and diverse language not captured in these domains. We present Shady Linguistic Utterance Replication-Generation (SLURG) to address these limitations, exploring the feasibility of generating synthetic fallacious forum-style comments using large language models (LLMs), specifically DeepHermes-3-Mistral-24B. Our findings indicate that LLMs can replicate the syntactic patterns of real data} and that high-quality few-shot prompts enhance LLMs' ability to mimic the vocabulary diversity of online forums.

Pretraining data has a direct impact on the behaviors and quality of language models (LMs), but we only understand the most basic principles of this relationship. While most work focuses on pretraining data's effect on downstream task behavior, we investigate its relationship to LM representations. Previous work has discovered that, in language models, some concepts are encoded `linearly' in the representations, but what factors cause these representations to form? We study the connection between pretraining data frequency and models' linear representations of factual relations. We find evidence that the formation of linear representations is strongly connected to pretraining term frequencies; specifically for subject-relation-object fact triplets, both subject-object co-occurrence frequency and in-context learning accuracy for the relation are highly correlated with linear representations. This is the case across all phases of pretraining. In OLMo-7B and GPT-J, we discover that a linear representation consistently (but not exclusively) forms when the subjects and objects within a relation co-occur at least 1k and 2k times, respectively, regardless of when these occurrences happen during pretraining. Finally, we train a regression model on measurements of linear representation quality in fully-trained LMs that can predict how often a term was seen in pretraining. Our model achieves low error even on inputs from a different model with a different pretraining dataset, providing a new method for estimating properties of the otherwise-unknown training data of closed-data models. We conclude that the strength of linear representations in LMs contains signal about the models' pretraining corpora that may provide new avenues for controlling and improving model behavior: particularly, manipulating the models' training data to meet specific frequency thresholds.

Training a state-of-the-art Large Language Model (LLM) is an increasingly expensive endeavor due to growing computational, hardware, energy, and engineering demands. Yet, an often-overlooked (and seldom paid) expense is the human labor behind these models' training data. Every LLM is built on an unfathomable amount of human effort: trillions of carefully written words sourced from books, academic papers, codebases, social media, and more. This position paper aims to assign a monetary value to this labor and argues that the most expensive part of producing an LLM should be the compensation provided to training data producers for their work. To support this position, we study 64 LLMs released between 2016 and 2024, estimating what it would cost to pay people to produce their training datasets from scratch. Even under highly conservative estimates of wage rates, the costs of these models' training datasets are 10-1000 times larger than the costs to train the models themselves, representing a significant financial liability for LLM providers. In the face of the massive gap between the value of training data and the lack of compensation for its creation, we highlight and discuss research directions that could enable fairer practices in the future.