Tagged articles
foundation model
34 articles · Page 1 of 1
Mistral AI releases Timestral, a family of time series foundation models (5M–209M parameters) featuring Decoupled Horizon Forecasting (DHF), covariate-aware attention for exogenous variables, synthetic-first two-stage training, and direct 21-quantile probabilistic output, outperforming TimesFM-3 and Chronos-2 on GIFT-Eval and TIME benchmarks.
BenchmarkDHFMistral AI
0 likes · 15 min read

Tadpole introduces an autoencoder-based foundation model for 3D partial differential equations, pretrained via reconstruction on massive online-generated simulation data (202 TB equivalent) without storing it, enabling zero-shot reconstruction, dynamics prediction via lightweight adaptation, and generation across diverse 3D physical systems and resolutions.
3D PDEZero-shot Generalizationautoencoder
0 likes · 14 min read

Xiaomi releases Xiaomi-TabLDM, a foundation model for tabular data that uses large-scale synthetic pretraining, efficient model scaling with dual-stream feature groups and sparse MoE, and test-time scaling to achieve top-tier performance on four public benchmarks and real-world industrial tasks without per-dataset retraining.
BenchmarkMixture of ExpertsTest-Time Scaling
0 likes · 9 min read

The article examines the inherent limitations of the Transformer architecture—its three core shortcomings—and presents the Mobius framework, which decouples knowledge and reasoning to address these issues, offering up to four‑fold inference speed gains while maintaining accuracy and improving data efficiency.
AI scalingMobiusTransformer
0 likes · 14 min read

The TouchWorld model combines predictive tactile forecasting with fast reactive correction, enabling robots to anticipate contact patterns before motion and instantly adjust during execution, achieving up to 65% success on six real‑world tasks and outperforming baselines by over 15 percentage points.
Embodied AIfoundation modelpredictive modeling
0 likes · 14 min read

SurgMotion, the world’s first billion‑parameter surgical video foundation model trained on the 15‑million‑frame, 3,658‑hour SurgMotion‑15M dataset, introduces motion‑guided latent masking, spatiotemporal affinity self‑distillation and feature‑diversity regularization, delivering up to 16.5% gains in workflow recognition and 2.9% error reduction on static tasks while topping 17 benchmark evaluations and becoming the most‑downloaded surgical model on Hugging Face.
SurgMotionfoundation modellarge-scale dataset
0 likes · 8 min read

The paper introduces LEVIRDet‑159, the largest unified remote‑sensing detection dataset with 159 categories and 2.5 M annotations, and its foundation model LEVIRDetNet, which achieves state‑of‑the‑art performance on nine external benchmarks after a single training run, demonstrating strong cross‑scene generalization.
LEVIRDetObject Detectioncross-benchmark evaluation
0 likes · 9 min read

FTP-1, a new generalist tactile foundation policy trained on the 3,000‑hour FTP‑1‑Dataset covering 21 heterogeneous sensors from 26 sources, introduces a morphology‑aware token space and an independent tactile transformer expert, achieving up to 31.6‑percentage‑point gains on unseen sensors and consistently outperforming prior VLA baselines across 14 real‑world manipulation tasks.
DatasetRoboticsfoundation model
0 likes · 12 min read

A joint team from Cambridge, Aalto and Bristol introduces TESSERA, a pixel‑level remote‑sensing foundation model that leverages a Barlow‑Twins self‑supervised scheme and a novel d‑pixel data organization to achieve state‑of‑the‑art accuracy on classification, segmentation and regression tasks, especially when annotations are scarce.
Self-Supervised LearningSentinel-1Sentinel-2
0 likes · 12 min read

Samsung is rapidly building its own foundation model ecosystem, introducing the memory‑based Meki architecture to leverage ROM for edge AI, the multi‑domain M2RL reinforcement‑learning paradigm, and the LiveClawBench 3‑dimensional benchmark to evaluate physical‑world AI, signaling a strategic shift from cloud‑centric to physical AI deployment.
LiveClawBenchMemory-Based ArchitectureMulti-Domain RL
0 likes · 8 min read

Genesis AI’s new GENE‑26.5 embodied foundation model demonstrates long‑horizon robot capabilities—from cooking a multi‑step meal and solving a Rubik’s cube to playing a high‑speed piano piece—using a full‑stack system that combines human‑like hands, a data‑glove, extensive simulation, and ultra‑low‑latency control.
Data GloveEmbodied AISimulation
0 likes · 11 min read

This article examines the shortcomings of current AI assistants, outlines the ideal of long‑term memory engineering, reviews mainstream industry solutions such as hard‑context models and Retrieval‑Augmented Generation, proposes a four‑layer memory loop architecture, and looks ahead to online learning and collective intelligence for future agents.
AIAgentMemory
0 likes · 15 min read

Google's open‑source TimesFM model brings pre‑trained, GPT‑like capabilities to time‑series forecasting, offering few‑shot and zero‑shot predictions, extended context length, continuous quantile outputs, and easy integration via a simple PyTorch API for developers across domains.
GooglePyTorchTimesFM
0 likes · 7 min read

Google’s open‑source TimesFM is a decoder‑only Transformer foundation model that delivers plug‑and‑play time‑series forecasting with zero‑shot accuracy, larger context windows, quantile predictions, and a simple Hugging Face API, making it suitable for retail, energy, finance, monitoring, and IoT use cases.
Hugging FacePyTorchTimesFM
0 likes · 7 min read


Amap TechMar 30, 2026 · Artificial Intelligence
ABot-M0 is an open‑source Vision‑Language‑Action foundation model that unifies fragmented robot data, introduces Action Manifold Learning for smoother action prediction, and offers a plug‑and‑play dual‑stream perception architecture, achieving state‑of‑the‑art results on major manipulation benchmarks.
Embodied AIRoboticsaction manifold learning
0 likes · 4 min read

LongCat‑Next, Meituan’s new 3‑billion‑parameter foundation model, adopts a pure‑discrete DiNA architecture with next‑token prediction, converting vision, audio and text into unified tokens; it surpasses same‑size multimodal models on OmniDocBench‑EN, CharXivRQ and SWE‑Bench, avoids catastrophic forgetting, and introduces dNaViT, RVQ compression and a dual‑path detokenizer for high‑fidelity generation.
DiNALongCat-NextSWE-bench
0 likes · 10 min read

The scLong foundation model, trained on 48 million cells and 28 k genes, integrates full‑gene expression with Gene Ontology knowledge to outperform existing methods on genetic perturbation, chemical response, cancer drug prediction, gene‑regulatory network inference, and batch integration tasks.
bioinformaticsfoundation modelgene ontology
0 likes · 13 min read

RynnBrain, an open‑source unified spatiotemporal foundation model from Alibaba DAMO Academy, integrates perception, localization, physics‑based reasoning and planning across 2 B, 8 B and 30 B MoE scales, handles multimodal visual inputs, and outperforms existing models on over 20 embodied benchmarks.
AlibabaBenchmarkEmbodied AI
0 likes · 3 min read

Peter Steinberger announced his move to OpenAI and the conversion of OpenClaw into an independent foundation, sparking community debate over OpenAI's open‑source strategy, the future of AI agents, and the strategic implications of this partnership.
AI agentsOpenAIOpenClaw
0 likes · 4 min read

Walrus, a 1.3 billion‑parameter Transformer built by Polymathic AI, is pretrained on 19 diverse physics scenarios—including astrophysics, geoscience, rheology, plasma physics and acoustics—using techniques like patch jittering, adaptive compute tokenization and space‑time factorized attention, and consistently outperforms earlier foundation models on both short‑ and long‑term continuum dynamics predictions.
TransformerWalruscontinuum dynamics
0 likes · 13 min read

Skild AI raised about $1.4 billion in a C‑round led by SoftBank, with participation from Nvidia, Sequoia, Bezos Expeditions and others, to develop a universal foundation model—Skild Brain—that can be deployed across diverse robot platforms, leveraging large‑scale visual data and a hierarchical control architecture.
General AIRoboticsSkild AI
0 likes · 11 min read

Bid2X introduces a bidding‑environment foundation model that unifies heterogeneous ad‑bidding data, leverages variable and time attention mechanisms, handles zero‑inflated distributions, and demonstrates superior offline performance across eight large‑scale datasets and significant online gains in GMV and ROI when deployed on a major e‑commerce platform.
Advertisingbiddingfoundation model
0 likes · 20 min read

The paper introduces VidEmo, a multimodal video foundation model that uses a two‑stage emotion‑clue‑guided reasoning framework and a large emotion‑centric dataset (Emo‑CFG) to achieve state‑of‑the‑art performance on facial attribute, expression, and fine‑grained emotion tasks, surpassing Gemini 2.0.
AIComputer VisionDataset
0 likes · 15 min read

AION-1, developed by a consortium including UC Berkeley, Cambridge and Oxford, is the first large‑scale multimodal foundation model for astronomy that unifies images, spectra and catalog data via an early‑fusion backbone, achieving zero‑shot and linear‑probe performance that rivals or surpasses task‑specific models across diverse scientific tasks.
astronomycross‑modal generationfoundation model
0 likes · 18 min read

FinCast introduces a decoder‑only Transformer foundation model for financial time‑series forecasting that tackles non‑stationarity, multi‑domain diversity, and multi‑resolution challenges through input chunking with frequency embeddings, a sparse MoE decoder, and a PQ‑loss, achieving zero‑shot and supervised gains over state‑of‑the‑art baselines while running five times faster on consumer GPUs.
PQ lossTransformerfinancial time series
0 likes · 12 min read

Researchers present LucaOne, a Transformer‑based foundation model that unifies DNA/RNA and protein sequences using a 39‑token vocabulary, rotary positional encoding, and molecule‑type embeddings, and demonstrate through extensive multi‑task benchmarks that it outperforms domain‑specific models across seven biological tasks.
DNATransformerbioinformatics
0 likes · 5 min read

This article summarizes four recent time‑series forecasting papers, covering a universal delay‑embedding foundation model, a dual causal network that leverages exogenous variables, a distribution‑aware alignment plug‑in called TimeAlign, and a shapelet‑based framework for interpretable directional forecasting in noisy financial markets.
Representation LearningTime Seriescausal network
0 likes · 9 min read

DeepSeek‑V3.1 arrives on Amazon Bedrock as a fully managed foundation model offering two inference modes, improved benchmark performance over DeepSeek‑R1, support for over 100 languages, enhanced tool‑calling and agent capabilities, and detailed guidance for secure enterprise deployment.
Agentic AIAmazon BedrockBenchmark
0 likes · 7 min read

This article reviews Kronos, a unified and scalable pre‑training framework designed for financial K‑line data, detailing its tokenization approach, autoregressive architecture, large‑scale pre‑training on 12 billion records, and experimental results that show substantial gains in price prediction, volatility forecasting, synthetic data generation, and investment simulation.
Kronosautoregressive pretrainingfinancial time series
0 likes · 9 min read

Tencent's LFM4Ads introduces a full‑representation, multi‑granular knowledge transfer framework that moves user, item, and cross representations from a large foundation model to downstream tasks, achieving up to 2.45% platform GMV uplift across more than ten advertising scenarios.
Representation Learningads recommendationfoundation model
0 likes · 12 min read

Tencent's LFM4Ads foundation model introduces a full-representation, multi-granular knowledge transfer framework that moves user, item, and cross representations to downstream tasks, dramatically improving ad recommendation metrics across dozens of business scenarios.
Representation Learningad recommendationfoundation model
0 likes · 10 min read

This article reviews the Centaur foundation model—fine‑tuned from Llama 3‑70B on the Psych‑101 dataset—to assess its ability to predict human choices, brain activity, and decision rationales across diverse psychological experiments, while discussing generalization, over‑fitting, and future research limits.
CentaurCognitive ModelingPsychology
0 likes · 17 min read

On February 25 2025, Microsoft open‑sourced its first multimodal AI agent foundation model, Magma, which extends multimodal processing to images, video, and text, introduces Set‑of‑Mark and Trace‑of‑Mark techniques for spatial‑temporal reasoning, optimizes modular inference for edge devices, and integrates reinforcement learning for adaptive task execution.
Edge ComputingMagmaSet-of-Mark
0 likes · 6 min read


AntTechMar 1, 2024 · Artificial Intelligence
Ant Group introduced SkySense, a 2.06‑billion‑parameter multimodal remote‑sensing foundation model that outperformed 18 international rivals across 17 benchmark tasks, was accepted to CVPR 2024, and aims to support applications such as agriculture, urban planning, and disaster response.
Ant GroupCVPR 2024Satellite Imagery
0 likes · 6 min read
