Research
Papers and lab releases from arXiv, OpenAI, DeepMind, Anthropic and more.
Weak Ties, Strong Signals: Efficient Training Data Detection in Diffusion LLMs via Independent Token Sampling
Diffusion large language models (dLLMs) offer a compelling alternative to autoregressive models, yet they may expose sensitive training data during denoising. Detecting such usage is challenging because dLLMs lack the efficient one-pass probability decomposition of causal architectures. Existing methods rely on random masking to obtain tractable token-wise detection signals under limited query budgets, but fail to control dependencies among masked tokens. We demonstrate that this token-wise approximation introduces a non-negative structural estimation error, which is theoretically characterized by the cumulative conditional mutual information (CMI) among masked tokens and can obscure subtle memorization signals. This insight suggests that reliable detection requires masked token sets with weak internal dependency. To avoid the prohibitive cost of directly estimating CMI over token combinations, we propose \textit{Independent Token Sampling} (ITS), a query-efficient framework that uses an attention-derived pairwise dependency proxy to approximate the CMI-aware selection criterion. ITS further incorporates a diversity-promoting strategy to improve token coverage across sampling rounds, yielding aggregated token-wise signals that are less affected by dependency-induced approximation error. Experiments on multiple datasets show that ITS consistently outperforms state-of-the-art baselines across different models and datasets, achieving an AUC improvement of 0.18 on the ArXiv dataset while maintaining strong performance under limited query budgets. The code is available at https://github.com/Chrisqcwx/DLLM-MIA .
A Comparative Framework for Evaluating Foundation Models on Tabular Data: A Case Study in Healthcare
Tabular data is the most common format in clinical practice, encompassing laboratory results, medication records, diagnostic codes, and patient demographics. As foundation models for tabular data have grown in number and variety, a practical question has become harder to answer: which model should a clinician or data scientist actually choose for a given task, and why? Existing surveys catalogue what these models can do, but they stop short of providing a structured way to compare them against the specific demands of a real application. We introduce \system{}, a comparative evaluation framework that scores and ranks tabular foundation models (TFMs) across six clinically meaningful dimensions: how well a model generalizes to new datasets, how effectively it protects patient privacy, how much data it needs to perform well, how it scales with growing datasets and feature spaces, how interpretable its predictions are to clinicians, and how fairly it performs across patient subgroups. Each dimension is broken down into measurable sub-components, and groups of sub-components can optionally be combined into supplementary compound scores, called super-metrics, that provide a diagnostic view of how a model performs across several dimensions simultaneously. To show how the framework works in practice, we apply it to two healthcare use cases, screening for iron deficiency and predicting heart failure, demonstrating how the same set of metrics leads to different model rankings depending on what matters most in each clinical context. We also provide a taxonomy of 45 TFMs organized by their underlying architecture, which serves as a reference for researchers and practitioners looking to navigate this rapidly expanding field.
ResearchFrom Latent Biomarkers to Clinical Rules: Embedding-Guided Rule Mining and Attribution-Based Translation for Interpretable Tabular Learning
Clinical decision support tools are most useful when accurate predictions are accompanied by understandable explanations. Rule-based models provide transparency, but rules derived directly from raw clinical measurements may miss patterns arising from interactions between multiple variables. We present a four-step pipeline that mines decision rules in the latent space of an FT-Transformer and translates them back into measurable clinical features. Embedding dimensions that consistently separate patient groups are treated as latent biomarkers, rules are mined using small decision trees, and selected rules are translated using gradient-input saliency and CLS attention attribution. We evaluate the framework on six public clinical and population health datasets at four embedding dimensions. Translated rules outperformed raw-feature rules in five of six datasets, with mean AUROC gains ranging from 0.04 to 0.23. On the heart disease dataset, embedding-space rules reached 0.98 AUROC, but translation reduced this to 0.72, showing that high-performing latent rules cannot always be represented by simple raw-feature conditions. These results show that latent-space rule discovery can uncover predictive patterns while translating them into clinically measurable features that can be evaluated by clinicians.
ResearchDecoupling Internal Representational Changes and Causal Importance in Fine-Tuned Large Language Models
Fine-tuning has emerged as a widely adopted approach for adapting LLMs to a variety of downstream tasks. However, how it reshapes their internal mechanisms remains poorly understood. To address this, we investigate how fine-tuning alters internal representations in LLMs, including attention patterns and layer-wise activations, and examine whether these changes are linked to task-relevant components identified by EAP (e.g., attention heads and logit-level activations) that drive task performance. We find that EAP-identified components are concentrated within specific layers, indicating a degree of functional localisation in how models internalise task-specific behavior. Notably, the distribution of these components across layers is largely uncorrelated with the layers undergoing the most substantial representational changes during fine-tuning. Furthermore, we observe that overlap in EAP-identified components across tasks does not translate into cross-task performance transfer if the tasks are different in nature (e.g. classification vs. generative tasks). More specifically, fine-tuning on one task can lead to a degradation of performance on another when the two tasks exhibit a high degree of overlap in their EAP-identified components.
ResearchRank Portability Does Not Imply Feasibility Portability: Target-Specific Evaluation of Joint Hardware Constraints
Cross-device hardware evaluation often assumes that if architecture rankings transfer across devices, a proxy device can support target-side model selection. We stress-test this assumption for joint latency-energy feasibility across two public architecture families. On NAS-Bench-201, cross-device rank correlations are moderate, while target-comparable feasible-set overlap remains incomplete. A faithful AdaProxy diagnostic substantially improves latency ranking, showing that the observed boundary failures are not simply due to weak adaptation. Exact finite-sample split-conformal analysis also exposes an evidence bottleneck: a finite one-sided 90% threshold requires at least nine calibration observations. We then replicate the phenomenon on 10,000 GPT architectures across 13 HW-GPT-Bench devices. Relative to an RTX3080 proxy, target latency SRCC ranges from 0.951 to 0.996, yet proxy-reuse violation risk ranges from 33.3% to 100% under matched joint constraints. These results show that rank portability, feasibility portability, and target-specific decision support are distinct evaluation objects. Cross-device evaluations should therefore report which target environments actually support the operating point being claimed.
ResearchCogGym: Towards Large-Scale Comparative Evaluation of Human and Machine Cognition
Understanding and modeling human intelligence are parallel goals shared by artificial intelligence (AI) and cognitive science. As AI systems grow increasingly capable, in what ways do model responses resemble human responses, and where do they systematically diverge? The sheer breadth and diversity of the tasks humans can perform and think about pose a challenge for scalable and rigorous comparison between humans and models. We introduce CogGym, a scalable, unified framework grounded in cognitive science for systematically comparing model and human behavior on matched experimental trials. CogGym uses a semi-automated, human-in-the-loop pipeline to standardize diverse experimental paradigms into a task-agnostic Experiment Markup Language (EML), enabling reproducible and faithful comparison at scale. For initial release, we curate and standardize 258 cognitive experiments from 100 papers that focuses on human commonsense reasoning, and evaluate 50 large language models against human responses. We find a clear scaling trend where larger and more recent AI models better reproduce human judgments. Yet AI models' improvement on such common reasoning tasks is considerably slower than the gains observed on formal-reasoning benchmarks like math and coding, and model--human fit remains well below human splithalf reliability ($R^2 = 0.93$ on text, $0.95$ on image, and $0.92$ on video) with the best models achieving $R^2 = 0.59$ on text, $0.58$ on image, and $0.43$ on video experiments. We intend for CogGym to provide a living evaluation framework that continually incorporates new cognitive science experiments to characterize where model behavior resembles human behavior, where it systematically diverges, and how those patterns change as models and experiments evolve.
ResearchSummarize, Judge, Refine: Decoupled Content Understanding and Policy Learning for Multimodal Content Moderation
Content moderation systems traditionally entangle multimodal understanding with policy-specific classification, requiring full pipeline retraining for every policy change and suffering from label scarcity since multimedia cannot be meaningfully augmented. We propose Summarize-Judge-Refine (SJR), a two-model architecture that decouples these concerns via a natural language interface: a multimodal Content Model produces structured text summaries, and a text-only Policy Model classifies them against policy definitions. An iterative co-training loop refines the Content Model via GRPO to produce policy-relevant summaries, while text-space augmentation generates adversarial summary variants---an augmentation pathway impossible on raw multimedia---enabling few-shot policy bootstrap. Every decision is grounded in a human-readable summary, providing interpretability as a structural byproduct. On misleading advertisement detection, SJR achieves +23.6\% relative non-misleading F1 over a zero-shot chain-of-thought baseline, outperforming end-to-end SFT, STaR/RFT, and RLFT. Notably, a variant trained on zero real violating examples---with all positive-class data synthetically generated---matches the full-data model within 0.2\% relative on violating F1, demonstrating that new policies can launch without any real violation data.
ResearchEnabling Vision and Cross-Modal Learning for Multimodal Stroke Recurrence Prediction: An Interpretable Two-Step Framework
Multimodal stroke recurrence prediction requires effective integration of heterogeneous clinical and imaging data, yet modality imbalance often causes models to over-rely on dominant modalities and underutilize complementary information. While self-supervised pretraining and selective parameter freezing are commonly employed to improve representation learning and fine-tuning stability, their effect on modality contributions and cross-modal behavior in multimodal medical models remains largely unexplored. In this work, we investigate whether image pretraining on 3D CTA scans reduces modality imbalance and improves cross-modal integration for stroke recurrence prediction, a clinically critical task we recently addressed. To this end, two multimodal neural networks are pretrained in a self-supervised manner and subsequently fine-tuned using two distinct freezing strategies. Their performance and modality utilization are compared against both the baseline model from our previous work and models trained entirely from scratch in this study. Our results demonstrate that self-supervised pretraining enables more effective utilization of the multimodal image-tabular dataset, outperforming both the prior baseline and all non-pretrained models. Notably, the best-performing Vision Transformer based neural network successfully overcomes unimodal collapse. Synergy analysis reveals significant interactions between vision and both gender and CHD, suggesting clinically relevant patterns for stroke recurrence. Overall, our findings demonstrate that self-supervised pretraining and strategic fine-tuning support more balanced modality utilization and enable meaningful cross-modal interactions. Code is publicly available at https://github.com/ChristianGappGit/SSL_Pretraining.