A system uses large language models to evolve and improve geospatial algorithms, taking over much manual design and testing by experts. It lets algorithms refine themselves for map region tasks, aiming to free researchers for bigger questions.
Toward Self-Evolving Geospatial Algorithms: An LLM-Driven Framework with a Case Study on Spatial Regime Discovery · Xiayin Lou, Peng Luo, Liqiu Meng · arXiv:2611.11669
A mathematical model combined phone and public transport movement data to study how lockdowns affected COVID-19 infections and economic activity in Helsinki. It offers cities insights for balancing public health measures with economic and social well-being in future crises.
Modelling socio-economic effects of the pandemic in a metropolitan area · Jan E. Snellman, Rafael A. Barrio, Cecilia I. Ventura et al. · arXiv:2611.12270
The paper investigates whether changing complex data, such as faces, can skip fully decoding hidden details and still keep identity while generating accurate what-if images. More efficient what-if image generation could help medical imaging and creative design.
Inversion-Free High-Dimensional Counterfactual Inference · Saketh Vishnubhatla, Pingchuan Ma, Shu Wan et al. · arXiv:2611.12815
Strategically choosing training data proportions, even imbalanced ones, can match rigid group balancing for reliable AI across user groups. It offers a more flexible and efficient way to build fair, dependable AI in diverse settings.
Guiding Data Allocation for Robust Subpopulation Generalization · Zhaoying Pan, Yipei Wang, Shenyu Lu et al. · arXiv:2611.12884
A score finds unusual data points by measuring how much each point adds to the dataset's overall diversity. It may identify critical outliers more robustly and efficiently without complex manual adjustments.
Vendi Anomaly Scores for Efficient and Accurate Anomaly Detection · Amey Pasarkar, Adji Bousso Dieng · arXiv:2611.12871
The statistical tools used to measure whether trade impact fades are biased and always show decay even when impact is permanent. It changes how researchers model and interpret the lasting effects of large trades on prices.
Is price-impact decay statistically identifiable by propagator estimation? Diffusive prices force the inferred kernel to be transient and fine-tuned · Shunta Fujiwara, Yuki Sato, Kiyoshi Kanazawa · arXiv:2611.10574
Training large AI models at 4-bit precision can slow their learning of rare facts, needing much more data even when overall scores look similar. Low precision training aims to make AI more efficient, but this work suggests a trade-off in detailed knowledge.
What the Loss Curve Misses: FP4 Pretraining Slows the Acquisition of Rare Facts · Chang Lu, Mengjiao Zhang · arXiv:2611.12832
A look at whether powerful AI models, like those for language and vision, are learning to represent information about the world in increasingly similar internal ways.
Where Does Alignment Emerge and What Does It Mean? A Closer Look at the Platonic Representation Hypothesis · Julia Herbst, Simon Roschmann, Quentin Bouniot et al. · arXiv:2611.12869
An algorithm compresses learning updates so reinforcement learning agents can learn almost as fast as with unlimited communication while sending far less data. It could help deploy AI in robotics and smart devices where communication bandwidth is scarce.
Adaptive Quantization Preserves Fast Rates for TD Learning over Rate-Limited Channels · Sreejeet Maity, Aritra Mitra · arXiv:2611.12878
MaxIM teaches an AI to summarize long interactions by keeping what is most useful for future tasks, helping it remember context with limited memory. It could let large language models keep context over long interactions without costly human guidance on memory.
MaxIM: Maximally Informative Incremental Summarization via Reinforcement Learning · Jihwan Jeong, Guy Tennenholtz, Yinlam Chow et al. · arXiv:2611.12825