As language models have a greater impact on society, it is important to ensure they are aligned to a diverse range of perspectives and are able to reflect nuance in human values. However, the most popular training paradigms for modern language models often assume there is one optimal answer for every query, leading to generic responses and poor alignment. In this work, we aim to enhance pluralistic alignment of language models in a low-resource setting with two methods: pluralistic decoding and model steering. We empirically demonstrate that model steering offers consistent improvement over zero-shot and few-shot baselines with only 50 annotated samples. Our proposed methods decrease false positives in several high-stakes tasks such as hate speech detection and misinformation detection, and improves the distributional alignment to human values in GlobalOpinionQA. We hope our work highlights the importance of diversity and how language models can be adapted to consider nuanced perspectives.

Large Language Model (LLM) providers expose fine-tuning APIs that let end users fine-tune their frontier LLMs. Unfortunately, it has been shown that an adversary with fine-tuning access to an LLM can bypass safeguards. Particularly concerning, such attacks may avoid detection with datasets that are only implicitly harmful. Our work studies robust detection mechanisms for adversarial use of fine-tuning APIs. We introduce the concept of a fine-tuning auditing agent and show it can detect harmful fine-tuning prior to model deployment. We provide our auditing agent with access to the fine-tuning dataset, as well as the fine-tuned and pre-fine-tuned models, and request the agent assigns a risk score for the fine-tuning job. We evaluate our detection approach on a diverse set of eight strong fine-tuning attacks from the literature, along with five benign fine-tuned models, totaling over 1400 independent audits. These attacks are undetectable with basic content moderation on the dataset, highlighting the challenge of the task. With the best set of affordances, our auditing agent achieves a 56.2% detection rate of adversarial fine-tuning at a 1% false positive rate. Most promising, the auditor is able to detect covert cipher attacks that evade safety evaluations and content moderation of the dataset. While benign fine-tuning with unintentional subtle safety degradation remains a challenge, we establish a baseline configuration for further work in this area. We release our auditing agent at https://github.com/safety-research/finetuning-auditor.

Training and evaluation of Reinforcement Learning (RL) web agents have gained increasing attention, yet a scalable and efficient environment that couples realistic and robust browser-side interaction with controllable server-side state at scale is still missing. Existing environments tend to have one or more of the following issues: they overwhelm policy models with excessive and noisy context; they perform actions non-deterministically without waiting for the UI or network to stabilize; or they cannot scale isolated client-server containers effectively for parallel RL rollouts. We propose WEBSERV, an environment that includes 1) a compact, site-agnostic browser environment that balances context and action complexity, and 2) a scalable RL environment via efficient launching and resetting web-servers to enable scalable RL training and evaluation. We evaluate WEBSERV on the shopping CMS and Gitlab tasks in WebArena, achieving state-of-the-art single-prompt success rates while cutting launch latency by ~5x and storage need by ~240x, with a comparable memory footprint, enabling 200+ concurrent containers on a single host.

As AI tools become increasingly common for research ideation, robust evaluation is critical to ensure the validity and usefulness of generated ideas. We introduce ScholarEval, a retrieval augmented evaluation framework that assesses research ideas based on two fundamental criteria: soundness - the empirical validity of proposed methods based on existing literature, and contribution - the degree of advancement made by the idea across different dimensions relative to prior research. To evaluate ScholarEval, we introduce ScholarIdeas, the first expert-annotated dataset of multi-domain research ideas and reviews, comprised of 117 ideas across four disciplines: artificial intelligence, neuroscience, biochemistry, and ecology. Our evaluation shows that ScholarEval achieves significantly higher coverage of points mentioned in the human expert annotated rubrics in ScholarIdeas compared to all baselines. Furthermore, ScholarEval is consistently preferred over our strongest baseline o4-mini-deep-research, a reasoning and search-enabled agentic system by OpenAI, in terms of evaluation actionability, depth, and evidence support. Our large-scale user study also shows that ScholarEval significantly outperforms deep research in literature engagement, idea refinement, and usefulness. We openly release our code, dataset, and ScholarEval tool for the community to use and build on.

Despite recent advances in AI, multimodal culturally diverse datasets are still limited, particularly for regions in the Middle East and Africa. In this paper, we introduce EgMM-Corpus, a multimodal dataset dedicated to Egyptian culture. By designing and running a new data collection pipeline, we collected over 3,000 images, covering 313 concepts across landmarks, food, and folklore. Each entry in the dataset is manually validated for cultural authenticity and multimodal coherence. EgMM-Corpus aims to provide a reliable resource for evaluating and training vision-language models in an Egyptian cultural context. We further evaluate the zero-shot performance of Contrastive Language-Image Pre-training CLIP on EgMM-Corpus, on which it achieves 21.2% Top-1 accuracy and 36.4% Top-5 accuracy in classification. These results underscore the existing cultural bias in large-scale vision-language models and demonstrate the importance of EgMM-Corpus as a benchmark for developing culturally aware models.

The rise of generative AI (GenAI) has impacted many aspects of human life. As these systems become embedded in everyday practices, understanding public trust in them also becomes essential for responsible adoption and governance. Prior work on trust in AI has largely drawn from psychology and human-computer interaction, but there is a lack of computational, large-scale, and longitudinal approaches to measuring trust and distrust in GenAI and large language models (LLMs). This paper presents the first computational study of Trust and Distrust in GenAI, using a multi-year Reddit dataset (2022--2025) spanning 39 subreddits and 197,618 posts. Crowd-sourced annotations of a representative sample were combined with classification models to scale analysis. We find that Trust and Distrust are nearly balanced over time, with shifts around major model releases. Technical performance and usability dominate as dimensions, while personal experience is the most frequent reason shaping attitudes. Distinct patterns also emerge across trustors (e.g., experts, ethicists, general users). Our results provide a methodological framework for large-scale Trust analysis and insights into evolving public perceptions of GenAI.

Effective human-AI collaboration on complex reasoning tasks requires that users understand and interact with the model's process, not just receive an output. However, the monolithic text from methods like Chain-of-Thought (CoT) prevents this, as current interfaces lack real-time verbalization and robust user barge-in. We present AsyncVoice Agent, a system whose asynchronous architecture decouples a streaming LLM backend from a conversational voice frontend. This design allows narration and inference to run in parallel, empowering users to interrupt, query, and steer the model's reasoning process at any time. Objective benchmarks show this approach reduces interaction latency by more than 600x compared to monolithic baselines while ensuring high fidelity and competitive task accuracy. By enabling a two-way dialogue with a model's thought process, AsyncVoice Agent offers a new paradigm for building more effective, steerable, and trustworthy human-AI systems for high-stakes tasks.

Scientific literature is increasingly siloed by complex language, static disciplinary structures, and potentially sparse keyword systems, making it cumbersome to capture the dynamic nature of modern science. This study addresses these challenges by introducing an adaptable large language model (LLM)-driven framework to quantify thematic trends and map the evolving landscape of scientific knowledge. The approach is demonstrated over a 20-year collection of more than 1,500 engineering articles published by the Proceedings of the National Academy of Sciences (PNAS), marked for their breadth and depth of research focus. A two-stage classification pipeline first establishes a primary thematic category for each article based on its abstract. The subsequent phase performs a full-text analysis to assign secondary classifications, revealing latent, cross-topic connections across the corpus. Traditional natural language processing (NLP) methods, such as Bag-of-Words (BoW) and Term Frequency-Inverse Document Frequency (TF-IDF), confirm the resulting topical structure and also suggest that standalone word-frequency analyses may be insufficient for mapping fields with high diversity. Finally, a disjoint graph representation between the primary and secondary classifications reveals implicit connections between themes that may be less apparent when analyzing abstracts or keywords alone. The findings show that the approach independently recovers much of the journal's editorially embedded structure without prior knowledge of its existing dual-classification schema (e.g., biological studies also classified as engineering). This framework offers a powerful tool for detecting potential thematic trends and providing a high-level overview of scientific progress.

Language models are pretrained on sequences that blend statistical regularities (making text fluent) with factual associations between specific tokens (knowledge of facts). While recent work suggests that the variability of their interaction, such as paraphrases of factual associations, critically determines generalization ability, we lack a systematic analysis of these impacts. This paper introduces a flexible synthetic testbed that combines a statistical stream of generic tokens with an abstract factual stream of source-target token pairs, enabling fine-grained control over their interaction. The design enables the independent control of diversity nature by manipulating stream composition (contextual structure) and the diversity level by varying which statistical streams each fact appears in. Through controlled experiments, we find that while higher contextual diversity delays in-distribution (ID) factual accuracy, its impact on out-of-distribution (OOD) factual generalization depends critically on contextual structure. In some cases, OOD performance follows the same trend as ID, but in others, diversity becomes essential for non-trivial factual recall. Even when low diversity prohibits factual recall, optimal diversity levels depend on training duration. Beyond factual recall failures, we identify structures where statistical generalization fails independently, and others where both capabilities degrade. This shows how the interplay between contextual design and diversity level impacts different generalization aspects. Further, through a series of controlled interventions on the model components, we trace the OOD failures to distinct optimization bottlenecks, highlighting the importance of the embedding and unembedding layers. Our synthetic framework allows us to isolate effects that would be confounded in large-scale studies, offering a controlled testbed for future investigations.

Creating high-quality clinical Chains-of-Thought (CoTs) is crucial for explainable medical Artificial Intelligence (AI) while constrained by data scarcity. Although Large Language Models (LLMs) can synthesize medical data, their clinical reliability remains unverified. This study evaluates the reliability of LLM-generated CoTs and investigates prompting strategies to enhance their quality. In a blinded comparative study, senior clinicians in Assisted Reproductive Technology (ART) evaluated CoTs generated via three distinct strategies: Zero-shot, Random Few-shot (using shallow examples), and Selective Few-shot (using diverse, high-quality examples). These expert ratings were compared against evaluations from a state-of-the-art AI model (GPT-4o). The Selective Few-shot strategy significantly outperformed other strategies across all human evaluation metrics (p < .001). Critically, the Random Few-shot strategy offered no significant improvement over the Zero-shot baseline, demonstrating that low-quality examples are as ineffective as no examples. The success of the Selective strategy is attributed to two principles: "Gold-Standard Depth" (reasoning quality) and "Representative Diversity" (generalization). Notably, the AI evaluator failed to discern these critical performance differences. The clinical reliability of synthetic CoTs is dictated by strategic prompt curation, not the mere presence of examples. We propose a "Dual Principles" framework as a foundational methodology to generate trustworthy data at scale. This work offers a validated solution to the data bottleneck and confirms the indispensable role of human expertise in evaluating high-stakes clinical AI.