Live intelligence

Latest AI updates

5 updates
Research

LeCat: Measuring LLM Category Learning from Pretraining Data

LeCat is a two-part framework for analyzing how LLMs learn category knowledge from pretraining data, distinguishing memorization, generalization, and flexible use. Experiments on open-source models and a controlled 68M model show that verbatim repetition reduces flexible learning, while diverse knowledge expressions improve robustness. Contextual flexibility remains a persistent challenge.

LLM evaluationpretraining data analysiscategory knowledge
Research

DINOv2 Adaptation Generalizes Better Than Task-Specific Medical Segmentation Models

A new arXiv study compares adapting the DINOv2 foundation model with training U-Nets and ResU-Nets from scratch for medical image segmentation across colonoscopy, breast ultrasound and dermoscopy. The adapted foundation model generalized better on every independent external dataset, while task-specific models remained competitive in-distribution when training data were abundant and delineated boundaries more precisely. Only a small fraction of the foundation model needed adaptation, with the required capacity determined by the imaging domain rather than the tuning method; the authors also found that a widely used polyp test set duplicates images from another benchmark.

Medical image segmentationFoundation modelsDINOv2
Research

MARS: Multiscale Appearance-Residual Streams for Video Forgery Detection

Researchers introduce MARS, a two-path video detector that combines multiscale feature differences with appearance representations to improve deepfake detection across facial manipulations and fully generated videos. The method achieves the highest average AUROC of 91.43% across seven datasets, outperforming seven compared detectors. It also shows strong transfer to generated videos when trained only on facial manipulations, and provides interpretable patch and frame contributions.

deepfake detectionvideo forensicsmultiscale feature learning
Research Editor selected

SelfImprove-Bench: Evaluating Generalization in Self-Improving Agents

SelfImprove-Bench is a new evaluation framework for recursively self-improving (RSI) agents, organizing tasks into three core capabilities—Exploring, Planning, Reasoning—using controllable game environments. It introduces a generalization ladder (G0–G4) to measure how well capability gains transfer beyond training conditions. Experiments show that existing RSI methods achieve inconsistent generalization across these levels, highlighting the need for more robust self-improvement strategies.

Why it matters: Provides a rigorous, standardized benchmark for measuring generalization in self-improving AI systems, a critical gap in current RSI research.

recursive self-improvementAI evaluationgeneralization
Research

ID2Pattern: Pretraining Generalizable Identity Behavior Patterns for Transaction Risk Foundation Models

ID2Pattern is a pretraining framework that captures identity behavior patterns in transaction sequences using a Local Anonymous ID tokenizer and a multi-scale pattern codebook. It outperforms strong baselines on an industrial dataset with over 100M sequences and six public datasets, improving performance by 4.13% and 8.83% on average respectively. The method remains effective even when identity values are unseen during training, showing potential for generalization to new users and markets.

transaction risk modelingbehavioral pretrainingidentity encoding