Emergent misalignment (EM) occurs when narrow finetuning causes a model to behave dangerously outside the finetuning task. Standard training signals can miss this shift, making reliable detection costly if it depends on repeated behavioral evaluation. We ask whether emergent misalignment can instead be detected from internal representations during finetuning. Using seven alignment-relevant traits encoded as linear directions in activation space, we track representational drift across training checkpoints in four open-source 7-9B LLMs. EM-relevant drift concentrates on a low-dimensional axis that explains 65.5

To address toxic content on social media, we introduce SMARTER, a data-efficient 2-stage framework for explainable content moderation using Large Language Models (LLMs). In Stage 1, we leverage LLMs’ own outputs to generate synthetic explanations for correct and incorrect labels, enabling preference optimization with minimal supervision. In Stage 2, we refine explanation quality through cross-model training, allowing weaker models to align with stronger ones. Experiments on 3 benchmarks (HateXplain, Latent Hate, Implicit Hate) show SMARTER achieves up to 13% macro-F1 improvement over few-shot baselines using only 6-57% of training data. Our framework offers a scalable strategy for low-data settings by harnessing LLMs’ self-improvement for explainable moderation.

Despite the growing use of large language models (LLMs) for writing tasks, users may hesitate to rely on LLMs when personal style is important. Post-editing LLM-generated drafts or translations is a common collaborative writing strategy, but it remains unclear whether users can effectively reshape LLM-generated text to reflect their personal style. We conduct a pre-registered online study (n=81) in which participants post-edit LLM-generated drafts for writing tasks where personal style matters to them. Using embedding-based style similarity metrics, we find that post-editing increases stylistic similarity to participants' unassisted writing and reduces similarity to fully LLM-generated output. However, post-edited text still remains stylistically closer in style to LLM text than to participants' unassisted control text, and it exhibits reduced stylistic diversity compared to unassisted human text. We find a gap between perceived stylistic authenticity and model-measured stylistic similarity, with post-edited text often perceived as representative of participants' personal style despite remaining detectable LLM stylistic traces.

Research has documented LLMs' name-based bias in hiring and salary recommendations. In this paper, we instead consider a setting where LLMs generate candidate summaries for downstream assessment. In a large-scale controlled study, we analyze nearly one million resume summaries produced by 4 models under systematic race-gender name perturbations, using synthetic resumes and real-world job postings. By decomposing each summary into resume-grounded factual content and evaluative framing, we find that factual content remains largely stable, while evaluative language exhibits subtle name-conditioned variation concentrated in the extremes of the distribution, especially in open-source models. Our hiring simulation demonstrates how evaluative summary transforms directional harm into symmetric instability that might evade conventional fairness audit, highlighting a potential pathway for LLM-to-LLM automation bias.

Autonomous AI research agents aim to accelerate scientific discovery by automating the research pipeline, from hypothesis generation to peer review. However, existing benchmarks rarely test a fundamental bottleneck: whether Large Language Models can judge the methodological viability of a research idea before expending time and computational resources. We introduce SoundnessBench, a curated benchmark of 1,099 machine-learning research proposals reconstructed from ICLR submissions, labeled with reviewer soundness sub-scores, and audited against source papers. SoundnessBench should be interpreted as a benchmark for recoverable proposal-stage soundness rather than exact prediction of full-paper review outcomes. Across 12 frontier LLMs, we find a pervasive optimism bias: under standard prompting, models frequently rate low-soundness proposals as sound, while aggressive prompting largely shifts errors from false positives to false negatives. Additional controls for public-corpus contamination, paper-identifying phrases, surface features, and human audit quality suggest that this behavior is not explained by a single confounder. Our results indicate that current LLMs are not yet reliable as standalone first-gate evaluators for scientific rigor.

Deep Research (DR) is an emerging agent application that leverages large language models (LLMs) to address open-ended queries. It requires the integration of several capabilities, including multi-step reasoning, cross-document synthesis, and the generation of evidence-backed, long-form answers. Evaluating DR remains challenging because responses are lengthy and diverse, admit many valid solutions, and often depend on dynamic information sources. We introduce ResearchRubrics, a standardized benchmark for DR built with over 2,800+ hours of human labor that pairs realistic, domain-diverse prompts with 2,500+ expert-written, fine-grained rubrics to assess factual grounding, reasoning soundness, and clarity. We also propose a new complexity framework for categorizing DR tasks along three axes: conceptual breadth, logical nesting, and exploration. In addition, we develop human and model-based evaluation protocols that measure rubric adherence for DR agents. We evaluate several state-of-the-art DR systems and find that even leading agents like Gemini's DR and OpenAI's DR achieve under 68% average compliance with our rubrics, primarily due to missed implicit context and inadequate reasoning about retrieved information. Our results highlight the need for robust, scalable assessment of deep research capabilities, to which end we release ResearchRubrics(including all prompts, rubrics, and evaluation code) to facilitate progress toward well-justified research assistants.

Large Language Models (LLMs) are increasingly involved in high-stakes domains, yet how they reason about socially-sensitive decisions still remain underexplored. We present a large-scale audit of LLMs’ treatment of socioeconomic status (SES) in college admissions decisions using a novel dual-process framework inspired by cognitive science. Leveraging a synthetic dataset of 30,000 applicant profiles grounded in real-world correlations, we prompt 4 open-source LLMs (Qwen 2, Mistral v0.3, Gemma 2, Llama 3.1) under 2 modes: a fast, decision-only setup (System 1) and a slower, explanation-based setup (System 2). Results from 5 million prompts reveals that LLMs consistently favor low-SES applicants—even when controlling for academic performance—and that System 2 amplifies this tendency by explicitly invoking SES as compensatory justification, highlighting both their potential and volatility as decision-makers. We then propose DPAF, a dual-process audit framework to probe LLMs’ reasoning behaviors in sensitive applications.

Large Language Models (LLMs) are increasingly used in healthcare, yet ensuring their safety and trustworthiness remains a barrier to deployment. Conversational medical assistants must avoid unsafe compliance without over-refusing benign queries. We present an iterative post-deployment alignment framework that applies Kahneman-Tversky Optimization (KTO) and Direct Preference Optimization (DPO) to refine models against domain-specific safety signals. Using the CARES-18K benchmark for adversarial robustness, we evaluate four LLMs (Llama-3B/8B, Meditron-8B, Mistral-7B) across multiple cycles. Our results show up to 42

Social science research has shown that candidates with names indicative of certain races or genders often face discrimination in employment practices. Similarly, Large Language Models (LLMs) have demonstrated racial and gender biases in various applications. In this study, we utilize GPT-3.5-Turbo and Llama 3-70B-Instruct to simulate hiring decisions and salary recommendations for candidates with 320 first names that strongly signal their race and gender, across over 750,000 prompts. Our empirical results indicate a preference among these models for hiring candidates with White female-sounding names over other demographic groups across 40 occupations. Additionally, even among candidates with identical qualifications, salary recommendations vary by as much as 5 between different subgroups. A comparison with real-world labor data reveals inconsistent alignment with U.S. labor market characteristics, underscoring the necessity of risk investigation of LLM-powered systems.

The propagation of offensive content through social media channels has garnered attention of the research community. Multiple works have proposed various semantically related yet subtle distinct categories of offensive speech. In this work, we explore meta-earning approaches to leverage the diversity of offensive speech corpora to enhance their reliable and efficient detection. We propose a joint embedding architecture that incorporates the input's label and definition for classification via Prototypical Network. Our model achieves at least 75% of the maximal F1-score while using less than 10% of the available training data across 4 datasets. Our experimental findings also provide a case study of training strategies valuable to combat resource scarcity.

The widespread use of social media necessitates reliable and efficient detection of offensive content to mitigate harmful effects. Although sophisticated models perform well on individual datasets, they often fail to generalize due to varying definitions and labeling of “offensive content.” In this paper, we introduce HateCOT, an English dataset with over 52,000 samples from diverse sources, featuring explanations generated by GPT-3.5Turbo and curated by humans. We demonstrate that pretraining on HateCOT significantly enhances the performance of open-source Large Language Models on three benchmark datasets for offensive content detection in both zero-shot and few-shot settings, despite differences in domain and task. Additionally, HateCOT facilitates effective K-shot fine-tuning of LLMs with limited data and improves the quality of their explanations, as confirmed by our human evaluation.

The Children's Data Network (CDN) is a data and research collaborative focused on the linkage and analysis of administrative records. In partnership with public agencies, philanthropic funders, affiliated researchers, and community stakeholders, we seek to generate knowledge and advance evidence-rich policies that improve the health, safety, and well-being of the children of California. Given our experience negotiating access to and working with existing administrative data (and importantly, data stewards), the CDN has demonstrated its ability to perform cost-effective and rigorous record linkage, answer time-sensitive policy- and program-related questions, and build the public sector's capacity to do the same. Owing to steadfast and generous infrastructure and project support, close collaboration with public partners, and strategic analyses and engagements, the CDN has promoted a person-level and longitudinal understanding of children and families in California and in so doing, informed policy and program development nationwide. We sincerely hope that our experience—and lessons learned—can advance and inform work in other fields and jurisdictions.

Objective To determine the population prevalence of diagnosed mental health disorders among Medicaid-insured children <18 years old in California based on levels of current and past child protection system (CPS) involvement. Study design In this retrospective, population-based study, we examined the full population of children enrolled in California's Medicaid program for at least 1 month between 2014 and 2015 and who had at least 1 claim during that period (n = 3 352 886). Records for Medicaid-insured children were probabilistically linked to statewide CPS records of maltreatment and foster care placements since 1998. A primary or secondary mental health diagnosis was classified using International Classification of Diseases codes. Results Overall, 14% (n = 470513) of all children insured through Medicaid in 2014-2015 had a documented mental health diagnosis. Among children with a diagnosis, the percentage with CPS involvement (ie, any report for maltreatment) was nearly twice that of the Medicaid population overall (50.4% vs 26.9%). This finding held across all diagnostic groups but with notable variations in magnitude. A graded relationship emerged between the level of CPS involvement and the likelihood of a mental health diagnosis. Diagnoses among children reported for maltreatment were common, regardless of placement in foster care. Conclusions Findings document high rates of both mental health diagnoses and past child protection involvement in a population of Medicaid-insured children. Most children reported for maltreatment will never be placed in foster care, underscoring the importance of ensuring that the children who remain at home receive the proper array and coordination of services.

The Children's Data Network (CDN) is a data and research collaborative focused on the linkage and analysis of administrative records. In partnership with public agencies, philanthropic funders, affiliated researchers, and community stakeholders, we seek to generate knowledge and advance evidence-rich policies that improve the health, safety, and well-being of the children of California. Given our experience negotiating access to and working with existing administrative data (and importantly, data stewards), the CDN has demonstrated its ability to perform cost-effective and rigorous record linkage, answer time-sensitive policy- and program-related questions, and build the public sector's capacity to do the same. Owing to steadfast and generous infrastructure and project support, close collaboration with public partners, and strategic analyses and engagements, the CDN has promoted a person-level and longitudinal understanding of children and families in California and in so doing, informed policy and program development nationwide. We sincerely hope that our experience-and lessons learned-can advance and inform work in other fields and jurisdictions.

Content warning: This work displays examples of explicit and/or strongly offensive language. Fueled by a surge of anti-Asian xenophobia and prejudice during the COVID-19 pandemic, many have taken to social media to express these negative sentiments. Identifying these posts is crucial for moderation and understanding the nature of hate in online spaces. In this paper, we create and annotate a corpus of tweets to explore anti-Asian hate speech with a finer level of granularity. Our analysis reveals that this emergent form of hate speech often eludes established approaches. To address this challenge, we develop a model and an accompanied efficient training regimen that incorporates agreement between annotators. Our approach produces up to 8.8% improvement in macro F1 scores over a strong established baseline, indicating its effectiveness even in settings where consensus among annotators is low. We demonstrate that we are able to identify hate speech that is systematically missed by established hate speech detectors.

The cryptocurrency market has gained significant traction in the last decade, becoming an alternative finance platform to traditional stock market trading. Despite its rapid evolution, legal regulations have not yet caught up to the cryptocurrency market's progress, attracting the attention of scammers looking to exploit legal loopholes for profits. Pump-and-dump schemes, a well-worn fraud device, has regained relevance in this new territory. In a typical pump-and-dump scheme, scammers organize and leverage media channels to artificially inflate the price of an alternative cryptocurrency, only to quickly sell them to profit off unsuspecting buyers. The disruptive nature of pump-and-dump schemes necessitates a system to reliably forecast pump targets and the magnitude of its success. In this paper, we propose an approach to predict the target cryptocurrency for each pump before its announcement using market and social media signals using Neural Network-based architectures while offering interpretable insights into their black-box nature. Additionally, we construct models that are capable of forecasting the highest price induced by the pump after the cryptocurrency's identity is revealed within 6.1% error margin. We examine the optimal temporal windows and describe the limitations of social data to predict the manipulations in cryptocurrency trade. Our experimental results serve as proof of a feasible forecasting expert system for identifying cryptocurrency pump-and-dump frauds using publicly available data.