Large language models (LLMs) demonstrate strong capabilities across a wide range of complex tasks and are increasingly deployed at scale, placing significant demands on inference efficiency. Prior work typically decomposes inference into prefill and decode stages, with the decode stage dominating total latency. To reduce time and memory complexity in the decode stage, a line of work introduces sparse-attention algorithms. In this paper, we show, both empirically and theoretically, that sparse attention can paradoxically increase end-to-end complexity: information loss often induces significantly longer sequences, a phenomenon we term “Less is Less” (Lil). To mitigate the Lil problem, we propose an early-stopping algorithm that detects the threshold where information loss exceeds information gain during sparse decoding. Our early-stopping algorithm reduces token consumption by up to 90% with a marginal accuracy degradation of less than 2% across reasoning-intensive benchmarks.

Large Language Model (LLM) agents deployed for real-world tasks face a fundamental dilemma: user requests are underspecified, yet agents must decide whether to act on incomplete information or interrupt users for clarification. Existing approaches either rely on brittle confidence thresholds that require task-specific tuning, or fail to account for the varying stakes of different decisions. We introduce a decision-theoretic framework that resolves this trade-off through the Value of Information (VoI), enabling agents to dynamically weigh the expected utility gain from asking questions against the cognitive cost imposed on users. Our inference-time method requires no hyperparameter tuning and adapts seamlessly across contexts—from casual games to medical diagnosis. Experiments across four diverse domains (20 Questions, medical diagnosis, flight booking, and e-commerce) show that VoI consistently matches or exceeds the best manually-tuned baselines, achieving up to 1.36 utility points higher in high-cost settings. This work provides a parameter-free framework for adaptive agent communication that explicitly balances task risk, query ambiguity, and user effort.

The "alignment tax" of post-training is typically framed as a drop in task accuracy. We show it also involves a severe loss of calibration, making models overconfident, less reliable, and model outputs less diverse. We demonstrate that this trade-off can be navigated effectively via a simple post-hoc intervention: interpolating between a model’s weights before and after alignment. Crucially, this is not a strict trade-off. We find that the process consistently reveals Pareto-optimal interpolations—models that improve accuracy beyond both parents while substantially recovering the calibration lost during alignment. Our work demonstrates that simple model merging provides a computationally efficient method for mitigating the full scope of the alignment tax, yielding models that are more capable and more reliable.

Understanding how discussion dynamics shape team creativity has been limited by the difficulty of measuring process at scale. We introduce Trace, a corpus of 309 group discussions from 103 teams (460 participants) across six creative problem-solving tasks. The dataset follows an input-process-output framework, integrating team composition (demographics, personalities), full discussion transcripts, and creativity outcomes. Using sentence embeddings and factor analysis, we identify four interpretable discussion dimensions: Coherence, Exploration, Convergence, and Participation. Analysis reveals a depth-breadth trade-off: coherent idea development inversely relates to semantic exploration. Larger teams explore more broadly but converge less effectively while team diversity shapes participation patterns more than discussion content. Novelty and usefulness in the creativity outcomes follow distinct pathways: Exploration and Convergence predict novelty, whereas Coherence predicts usefulness. These findings ground our understanding of how teams talk their way to creative solutions and provide guidance for designing multiagent systems.

While confidence estimation is a promising direction for mitigating hallucinations in Large Language Models (LLMs), current research overwhelmingly focuses on single-turn settings. The dynamics of model confidence in multi-turn conversations, where context accumulates and ambiguity is progressively resolved, remain largely unexplored. This work presents the first systematic study of confidence estimation in multi-turn interactions, establishing a formal evaluation framework grounded in two key desiderata: per-turn calibration and monotonicity of confidence as more information becomes available. To facilitate this, we introduce novel metrics, including a length-normalized Expected Calibration Error (InfoECE), and a new "Hinter-Guesser" paradigm for generating controlled evaluation datasets. Our experiments reveal that widely-used confidence techniques struggle with calibration and monotonicity in multi-turn dialogues. In contrast, a novel logit-based probe we introduce, P(Sufficient), proves comparatively more effective, robustly tracking evidence accumulation and distinguishing it from conversational filler. Our work provides a foundational methodology for developing more reliable and trustworthy conversational agents.

Reasoning-tuned LLMs utilizing long Chain-of-Thought (CoT) excel at single-answer tasks, yet their ability to model Human Label Variation—which requires capturing probabilistic ambiguity rather than resolving it—remains underexplored. We investigate this through systematic disentanglement experiments on distribution-based tasks, employing Cross-CoT experiments to isolate the effect of reasoning text from intrinsic model priors. We observe a distinct "decoupled mechanism": while CoT improves distributional alignment, final accuracy is dictated by CoT content (99% variance contribution), whereas distributional ranking is governed by model priors (over 80%). Step-wise analysis further shows that while CoT’s influence on accuracy grows monotonically during the reasoning process, distributional structure is largely determined by LLM’s intrinsic priors. These findings suggest that long CoT serves as a decisive LLM decision-maker for the top option but fails to function as a granular distribution calibrator for ambiguous tasks.

Personalized content moderation can protect users from harm while facilitating free expression by tailoring moderation decisions to individual preferences rather than enforcing universal rules. However, content moderation that is fully personalized to individual preferences, no matter what these preferences are, may lead to even the most hazardous types of content being propagated on social media. In this paper, we explore this risk using hate speech as a case study. Certain types of hate speech are illegal in many countries. We show that, while fully personalized hate speech detection models increase overall user welfare (as measured by user-level classification performance), they also make predictions that violate such legal hate speech boundaries, especially when tailored to users who tolerate highly hateful content. To address this problem, we enforce legal boundaries in personalized hate speech detection by overriding predictions from personalized models with those from a boundary classifier. This approach significantly reduces legal violations while minimally affecting overall user welfare. Our findings highlight both the promise and the risks of personalized moderation, and offer a practical solution to balance user preferences with legal and ethical obligations.

Understanding how individuals perceive and react to information is fundamental for advancing social and behavioral sciences and developing human-centered AI systems. Current approaches often lack the granular data needed to model these personalized responses, relying instead on aggregated labels that obscure the rich variability driven by individual differences. We introduce iNews, a novel large-scale dataset specifically designed to facilitate the modeling of personalized affective responses to news content. Our dataset comprises annotations from 291 demographically diverse UK participants across 2,899 multimodal Facebook news posts from major UK outlets, with an average of 5.18 annotators per sample. For each post, annotators provide multifaceted labels including valence, arousal, dominance, discrete emotions, content relevance judgments, sharing likelihood, and modality importance ratings. Crucially, we collect comprehensive annotator persona information covering demographics, personality, media trust, and consumption patterns, which explain 15.2% of annotation variance - substantially higher than existing NLP datasets. Incorporating this information yields a 7% accuracy gain in zero-shot prediction and remains beneficial even with 32-shot in-context learning.

Large Language Models (LLMs) have demonstrated strong capabilities across various domains, with recent advancements in challenging reasoning tasks such as mathematics and programming. However, solving reasoning tasks often requires an LLM to generate long sequences, incurring O(N) time and memory complexities per token, where N is the current sequence length. To reduce complexities, existing sparsity-based algorithms propose to retain Key-Value (KV) vectors, the intermediate representations of only the most critical tokens. However, these algorithms struggle with the “impossible trinity” of accuracy, time, and memory. For example, the state-of-the-art algorithm, Quest, achieves high accuracy with O(L) time but O(N) memory (L is the cache budget, L ≪ N). To address the “impossible trinity”, in this paper, we identify a new attention pattern during the decode stage of reasoning tasks, where milestone tokens (analogous to lemmas in mathematical proofs) emerge, are utilized, and then become unimportant afterward. Based on this pattern, we propose a new algorithm RaaS that identifies milestone tokens and retains their KV vectors until they are no longer needed, achieving high accuracy with O(L) time and O(L) memory complexities.

We investigate the potential of LLM-generated synthetic data for improving low-resource Machine Translation (MT). Focusing on seven diverse target languages, we construct a document-level synthetic corpus from English Europarl, and extend it via pivoting to 147 additional language pairs. Automatic and human evaluation confirm its overall high quality. We study its practical application by (i) identifying effective training regimes, (ii) comparing our data with the HPLT dataset, (iii) studying the effect of varying training data size, and (iiii) testing its utility beyond English-centric MT. Finally, we introduce SynOPUS, a public repository for synthetic parallel datasets. Our findings show that LLM-generated synthetic data, even when noisy, can substantially improve MT performance for low-resource languages.

While Reinforcement Learning from Human Feedback (RLHF) is widely used to align Large Language Models (LLMs) with human preferences, it typically assumes homogeneous preferences across users, overlooking diverse human values and minority viewpoints.Although personalized preference learning addresses this by tailoring separate preferences for individual users, the field lacks standardized methods to assess its effectiveness. We present a multi-faceted evaluation framework that measures not only performance but also fairness, unintended effects, and adaptability across varying levels of preference divergence. Through extensive experiments comparing eight personalization methods across three preference datasets, we demonstrate that performance differences between methods could reach 36% when users strongly disagree, and personalization can introduce up to 20% safety misalignment. These findings highlight the critical need for holistic evaluation approaches to advance the development of more effective and inclusive preference learning systems.

Large language models (LLMs) have shown remarkable promise in simulating human language and behavior. This study investigates how integrating persona variables—demographic, social, and behavioral factors—impacts LLMs’ ability to simulate diverse perspectives. We find that persona variables account for <10% variance in annotations in existing subjective NLP datasets. Nonetheless, incorporating persona variables via prompting in LLMs provides modest but statistically significant improvements. Persona prompting is most effective in samples where many annotators disagree, but their disagreements are relatively minor. Notably, we find a linear relationship in our setting: the stronger the correlation between persona variables and human annotations, the more accurate the LLM predictions are using persona prompting. In a zero-shot setting, a powerful 70b model with persona prompting captures 81% of the annotation variance achievable by linear regression trained on ground truth annotations. However, for most subjective NLP datasets, where persona variables have limited explanatory power, the benefits of persona prompting are limited.

As large language models (LLMs) gain widespread adoption, ensuring they cater to diverse user needs has become increasingly important. While many researchers have studied LLM personalization and role-playing, they primarily use LLM-as-a-Judge for evaluation without thoroughly examining its validity. This paper investigates the reliability of LLM-as-a-Personalized-Judge—asking LLMs to judge user preferences based on persona. Our results suggest that LLM-as-a-Personalized-Judge is less reliable for personalization than previously believed, showing low agreement with human ground truth. We observed that the personas provided to the LLM often have limited predictive power for the tasks, leading us to introduce verbal uncertainty estimation. We find that powerful LLMs are aware of the certainty of their prediction and can achieve high agreement with ground truth on high-certainty samples, indicating a promising approach for building reliable and scalable proxies for evaluating LLM personalization. Our human annotation reveals that third-person crowd worker evaluations of personalized preferences are even worse than LLM predictions, highlighting the challenges of evaluating LLM personalization.

Recent advances in Large Language Models (LLMs) have sparked wide interest in validating and comprehending the human-like cognitive-behavioral traits LLMs may capture and convey. These cognitive-behavioral traits include typically Attitudes, Opinions, Values (AOVs). However, measuring AOVs embedded within LLMs remains opaque, and different evaluation methods may yield different results. This has led to a lack of clarity on how different studies are related to each other and how they can be interpreted. This paper aims to bridge this gap by providing a comprehensive overview of recent works on the evaluation of AOVs in LLMs. Moreover, we survey related approaches in different stages of the evaluation pipeline in these works. By doing so, we address the potential and challenges with respect to understanding the model, human-AI alignment, and downstream application in social sciences. Finally, we provide practical insights into evaluation methods, model enhancement, and interdisciplinary collaboration, thereby contributing to the evolving landscape of evaluating AOVs in LLMs.

The capabilities of large language models (LLMs) have raised concerns about their potential to create and propagate convincing narratives. Here, we study their performance in detecting convincing arguments to gain insights into LLMs’ persuasive capabilities without directly engaging in experimentation with humans. We extend a dataset by Durmus and Cardie (2018) with debates, votes, and user traits and propose tasks measuring LLMs’ ability to (1) distinguish between strong and weak arguments, (2) predict stances based on beliefs and demographic characteristics, and (3) determine the appeal of an argument to an individual based on their traits. We show that LLMs perform on par with humans in these tasks and that combining predictions from different LLMs yields significant performance gains, surpassing human performance. The data and code released with this paper contribute to the crucial effort of continuously evaluating and monitoring LLMs’ capabilities and potential impact. (https://go.epfl.ch/persuasion-llm)

Despite the importance of understanding causality, corpora addressing causal relations are limited. There is a discrepancy between existing annotation guidelines of event causality and conventional causality corpora that focus more on linguistics. Many guidelines restrict themselves to include only explicit relations or clause-based arguments. Therefore, we propose an annotation schema for event causality that addresses these concerns. We annotated 3,559 event sentences from protest event news with labels on whether it contains causal relations or not. Our corpus is known as the Causal News Corpus (CNC). A neural network built upon a state-of-the-art pre-trained language model performed well with 81.20% F1 score on test set, and 83.46% in 5-folds cross-validation. CNC is transferable across two external corpora: CausalTimeBank (CTB) and Penn Discourse Treebank (PDTB). Leveraging each of these external datasets for training, we achieved up to approximately 64% F1 on the CNC test set without additional fine-tuning. CNC also served as an effective training and pre-training dataset for the two external corpora. Lastly, we demonstrate the difficulty of our task to the layman in a crowd-sourced annotation exercise. Our annotated corpus is publicly available, providing a valuable resource for causal text mining researchers.

An ever-increasing amount of text, in the form of social media posts and news articles, gives rise to new challenges and opportunities for the automatic extraction of socio-political events. In this paper, we present our submission to the Shared Tasks on Socio-Political and Crisis Events Detection, Task 1, Multilingual Protest News Detection, Subtask 2, Event Sentence Classification, of CASE @ ACL-IJCNLP 2021. In our submission, we utilize the RoBERTa model with additional pretraining, and achieve the best F1 score of 0.8532 in event sentence classification in English and the second-best F1 score of 0.8700 in Portuguese via simple translation. We analyze the failure cases of our model. We also conduct an ablation study to show the effect of choosing the right pretrained language model, adding additional training data and data augmentation.

Evaluating the state-of-the-art event detection systems on determining spatio-temporal distribution of the events on the ground is performed unfrequently. But, the ability to both (1) extract events “in the wild” from text and (2) properly evaluate event detection systems has potential to support a wide variety of tasks such as monitoring the activity of socio-political movements, examining media coverage and public support of these movements, and informing policy decisions. Therefore, we study performance of the best event detection systems on detecting Black Lives Matter (BLM) events from tweets and news articles. The murder of George Floyd, an unarmed Black man, at the hands of police officers received global attention throughout the second half of 2020. Protests against police violence emerged worldwide and the BLM movement, which was once mostly regulated to the United States, was now seeing activity globally. This shared task asks participants to identify BLM related events from large unstructured data sources, using systems pretrained to extract socio-political events from text. We evaluate several metrics, accessing each system’s ability to identify protest events both temporally and spatially. Results show that identifying daily protest counts is an easier task than classifying spatial and temporal protest trends simultaneously, with maximum performance of 0.745 and 0.210 (Pearson r), respectively. Additionally, all baselines and participant systems suffered from low recall, with a maximum recall of 5.08.