Millions of players engage daily in competitive online games, communicating through in-game chat. Prior research has focused on detecting relatively small volumes of toxic content using various Natural Language Processing (NLP) techniques for the purpose of moderation. However, recent studies emphasize the importance of detecting prosocial communication, which can be as crucial as identifying toxic interactions. Recognizing prosocial behavior allows for its analysis, rewarding, and promotion. Unlike toxicity, there are limited datasets, models, and resources for identifying prosocial behaviors in game-chat text. In this work, we employed unsupervised discovery combined with game domain expert collaboration to identify and categorize prosocial player behaviors from game chat. We further propose a novel Self-Anchored Attention Model (SAAM) which gives 7.9% improvement compared to the best existing technique. The approach utilizes the entire training set as "anchors" to help improve model performance under the scarcity of training data. This approach led to the development of the first automated system for classifying prosocial behaviors in in-game chats, particularly given the low-resource settings where large-scale labeled data is not available. Our methodology was applied to one of the most popular online gaming titles - Call of Duty(R): Modern Warfare(R)II, showcasing its effectiveness. This research is novel in applying NLP techniques to discover and classify prosocial behaviors in player in-game chat communication. It can help shift the focus of moderation from solely penalizing toxicity to actively encouraging positive interactions on online platforms.

Large language model (LLM) unlearning has become a critical topic in machine learning, aiming to eliminate the influence of specific training data or knowledge without retraining the model from scratch. A variety of techniques have been proposed, including Gradient Ascent, model editing, and re-steering hidden representations. While existing surveys often organize these methods by their technical characteristics, such classifications tend to overlook a more fundamental dimension: the underlying intention of unlearning--whether it seeks to truly remove internal knowledge or merely suppress its behavioral effects. In this SoK paper, we propose a new taxonomy based on this intention-oriented perspective. Building on this taxonomy, we make three key contributions. First, we revisit recent findings suggesting that many removal methods may functionally behave like suppression, and explore whether true removal is necessary or achievable. Second, we survey existing evaluation strategies, identify limitations in current metrics and benchmarks, and suggest directions for developing more reliable and intention-aligned evaluations. Third, we highlight practical challenges--such as scalability and support for sequential unlearning--that currently hinder the broader deployment of unlearning methods. In summary, this work offers a comprehensive framework for understanding and advancing unlearning in generative AI, aiming to support future research and guide policy decisions around data removal and privacy.

AI models have become active decision makers, often acting without human supervision. The rapid advancement of AI technology has already caused harmful incidents that have hurt individuals and societies and AI unfairness in heavily criticized. It is urgent to disrupt AI pipelines which largely neglect human principles and focus on computational biases exploration at the data (pre), model(in), and deployment (post) processing stages. We claim that by exploiting the advances of agents technology, we will introduce cautious, prompt, and ongoing fairness watch schemes, under realistic, systematic, and human-centric fairness expectations. We envision agents as fairness guardians, since agents learn from their environment, adapt to new information, and solve complex problems by interacting with external tools and other systems. To set the proper fairness guardrails in the overall AI pipeline, we introduce a fairness-by-design approach which embeds multi-role agents in an end-to-end (human to AI) synergetic scheme. Our position is that we may design adaptive and realistic AI fairness frameworks, and we introduce a generalized algorithm which can be customized to the requirements and goals of each AI decision making scenario. Our proposed, so called FAIRTOPIA framework, is structured over a three-layered architecture, which encapsulates the AI pipeline inside an agentic guardian and a knowledge-based, self-refining layered scheme. Based on our proposition, we enact fairness watch in all of the AI pipeline stages, under robust multi-agent workflows, which will inspire new fairness research hypothesis, heuristics, and methods grounded in human-centric, systematic, interdisciplinary, socio-technical principles.

The widespread adoption of generative AI models has raised growing concerns about representational harm and potential discriminatory outcomes. Yet, despite growing literature on this topic, the mechanisms by which bias emerges - especially in unconditional generation - remain disentangled. We define the bias of an attribute as the difference between the probability of its presence in the observed distribution and its expected proportion in an ideal reference distribution. In our analysis, we train a set of unconditional image generative models and adopt a commonly used bias evaluation framework to study bias shift between training and generated distributions. Our experiments reveal that the detected attribute shifts are small. We find that the attribute shifts are sensitive to the attribute classifier used to label generated images in the evaluation framework, particularly when its decision boundaries fall in high-density regions. Our empirical analysis indicates that this classifier sensitivity is often observed in attributes values that lie on a spectrum, as opposed to exhibiting a binary nature. This highlights the need for more representative labeling practices, understanding the shortcomings through greater scrutiny of evaluation frameworks, and recognizing the socially complex nature of attributes when evaluating bias.

Despite advances in deep learning for automatic sleep staging, clinical adoption remains limited due to challenges in fair model evaluation, generalization across diverse datasets, model bias, and variability in human annotations. We present SLEEPYLAND, an open-source sleep staging evaluation framework designed to address these barriers. It includes more than 220'000 hours in-domain (ID) sleep recordings, and more than 84'000 hours out-of-domain (OOD) sleep recordings, spanning a broad range of ages, sleep-wake disorders, and hardware setups. We release pre-trained models based on high-performing SoA architectures and evaluate them under standardized conditions across single- and multi-channel EEG/EOG configurations. We introduce SOMNUS, an ensemble combining models across architectures and channel setups via soft voting. SOMNUS achieves robust performance across twenty-four different datasets, with macro-F1 scores between 68.7% and 87.2%, outperforming individual models in 94.9% of cases. Notably, SOMNUS surpasses previous SoA methods, even including cases where compared models were trained ID while SOMNUS treated the same data as OOD. Using a subset of the BSWR (N=6'633), we quantify model biases linked to age, gender, AHI, and PLMI, showing that while ensemble improves robustness, no model architecture consistently minimizes bias in performance and clinical markers estimation. In evaluations on OOD multi-annotated datasets (DOD-H, DOD-O), SOMNUS exceeds the best human scorer, i.e., MF1 85.2% vs 80.8% on DOD-H, and 80.2% vs 75.9% on DOD-O, better reproducing the scorer consensus than any individual expert (k = 0.89/0.85 and ACS = 0.95/0.94 for healthy/OSA cohorts). Finally, we introduce ensemble disagreement metrics - entropy and inter-model divergence based - predicting regions of scorer disagreement with ROC AUCs up to 0.828, offering a data-driven proxy for human uncertainty.

This chapter explores advancements in decoding strategies for large language models (LLMs), focusing on enhancing the Locally Typical Sampling (LTS) algorithm. Traditional decoding methods, such as top - k and nucleus sampling, often struggle to balance fluency, diversity, and coherence in text generation. To address these challenges, Adaptive Semantic - Aware Typicality Sampling (ASTS) is proposed as an improved version of LTS, incorporating dynamic entropy thresholding, multi - objective scoring, and reward - penalty adjustments. ASTS ensures contextually coherent and diverse text generation while maintaining computational efficiency. Its performance is evaluated across multiple benchmarks, including story generation and abstractive summarization, using metrics such a s perplexity, MAUVE, and diversity scores. Experimental results demonstrate that ASTS outperforms existing sampling techniques by reducing repetition, enhancing semantic alignment, and improving fluency. Keywords: Locally Typical Sampling, Adaptive Semantic - Aware Typicality Sampling (ASTS), Decoding Strategies, Large Language Models (LLMs), Entrop y - Based Sampling, Multi - Objective Scoring.

While large language models (LLMs) can solve PhD-level reasoning problems over long context inputs, they still struggle with a seemingly simpler task: following explicit length instructions-e.g., write a 10,000-word novel. Additionally, models often generate far too short outputs, terminate prematurely, or even refuse the request. Existing benchmarks focus primarily on evaluating generations quality, but often overlook whether the generations meet length constraints. To this end, we introduce Length Instruction Following Evaluation Benchmark (LIFEBench) to comprehensively evaluate LLMs' ability to follow length instructions across diverse tasks and a wide range of specified lengths. LIFEBench consists of 10,800 instances across 4 task categories in both English and Chinese, covering length constraints ranging from 16 to 8192 words. We evaluate 26 widely-used LLMs and find that most models reasonably follow short-length instructions but deteriorate sharply beyond a certain threshold. Surprisingly, almost all models fail to reach the vendor-claimed maximum output lengths in practice, as further confirmed by our evaluations extending up to 32K words. Even long-context LLMs, despite their extended input-output windows, counterintuitively fail to improve length-instructions following. Notably, Reasoning LLMs outperform even specialized long-text generation models, achieving state-of-the-art length following. Overall, LIFEBench uncovers fundamental limitations in current LLMs' length instructions following ability, offering critical insights for future progress.

Large language models (LLMs) exhibit remarkable capabilities across diverse tasks, yet aligning them efficiently and effectively with human expectations remains a critical challenge. This thesis advances LLM alignment by introducing novel methodologies in data collection, training, and evaluation. We first address alignment data collection. Existing approaches rely heavily on manually curated datasets or proprietary models. To overcome these limitations, we propose Lion, an adversarial distillation framework that iteratively refines training data by identifying and generating challenging instructions, enabling state-of-the-art zero-shot reasoning. Additionally, we introduce Web Reconstruction (WebR), a fully automated framework that synthesizes instruction-tuning data directly from raw web documents, significantly improving data diversity and scalability over existing synthetic data methods. Next, we enhance alignment training through novel optimization techniques. We develop Learning to Edit (LTE), a framework that enables LLMs to efficiently integrate new knowledge while preserving existing information. LTE leverages meta-learning to improve both real-time and batch knowledge updates. Furthermore, we introduce Bridging and Modeling Correlations (BMC), a refinement of Direct Preference Optimization (DPO) that explicitly captures token-level correlations in preference data, leading to superior alignment across QA and mathematical reasoning tasks. Finally, we tackle the challenge of evaluating alignment. Existing benchmarks emphasize response quality but overlook adherence to specific constraints. To bridge this gap, we introduce FollowBench, a multi-level, fine-grained benchmark assessing LLMs' ability to follow complex constraints across diverse instruction types. Our results expose key weaknesses in current models' constraint adherence, offering insights for future improvements.

Complementary product recommendation is a powerful strategy to improve customer experience and retail sales. However, recommending the right product is not a simple task because of the noisy and sparse nature of user-item interactions. In this work, we propose a simple yet effective method to predict a list of complementary products given a query item, based on the structure of a directed weighted graph projected from the user-item bipartite graph. We revisit bipartite graph projections for recommender systems and propose a novel approach for inferring complementarity relationships from historical user-item interactions. We compare our model with recent methods from the literature and show, despite the simplicity of our approach, an average improvement of +43% and +38% over sequential and graph-based recommenders, respectively, over different benchmarks.

In this work, we provide an extended discussion of a new approach to explainable Reinforcement Learning called Diverse Near-Optimal Alternatives (DNA), first proposed at L4DC 2025. DNA seeks a set of reasonable "options" for trajectory-planning agents, optimizing policies to produce qualitatively diverse trajectories in Euclidean space. In the spirit of explainability, these distinct policies are used to "explain" an agent's options in terms of available trajectory shapes from which a human user may choose. In particular, DNA applies to value function-based policies on Markov decision processes where agents are limited to continuous trajectories. Here, we describe DNA, which uses reward shaping in local, modified Q-learning problems to solve for distinct policies with guaranteed epsilon-optimality. We show that it successfully returns qualitatively different policies that constitute meaningfully different "options" in simulation, including a brief comparison to related approaches in the stochastic optimization field of Quality Diversity. Beyond the explanatory motivation, this work opens new possibilities for exploration and adaptive planning in RL.