Official platform of Tsinghua Big Data Research Center, sharing the team's latest research, teaching updates, and big data news.
Visual and vision-language models excel under IID benchmarks but often fail on out-of-distribution data due to shortcut learning; this article examines the problem, explains its causes, and proposes data-level and model-level interventions—including StillMix, FLASH, and SPARCL—to improve OOD robustness.
SeqWM introduces a sequential causal decomposition of multi‑robot dynamics, allowing each robot to model its marginal contribution conditioned on preceding agents, which simplifies learning, improves sample efficiency, and yields natural collaborative behaviors both in simulation (Bi‑DexHands, Multi‑Quadruped) and real‑world tests on Unitree Go2‑W, outperforming prior methods.
A Hong Kong University of Science and Technology team has created a miniature AI‑powered wearable ring that uses a 0.0081 mm² olfactory sensor chip to non‑invasively capture skin‑emitted VOCs, identify diet and activity states, and even quantify alcohol intake, offering a new frontier for continuous health monitoring.
The article reviews the Kimi team's Attention Residuals approach, which substitutes traditional ResNet additive shortcuts with learned attention‑based weighting, explains the theoretical motivation linking depth to time, details full‑attention and block‑wise implementations, presents experimental results showing up to 1.25× compute efficiency and improved performance on reasoning and knowledge tasks.
This article introduces nine practical Python libraries—glom, boltons, beartype, result, whenever, pyinstrument, dirty‑equals, stamina, and pyfunctional—that address common development pain points such as nested data handling, missing standard‑library features, runtime type safety, error handling, timezone bugs, performance profiling, robust testing, retry logic, and functional pipelines, providing production‑ready solutions with concise examples.
SwiftTailor introduces a two‑stage, geometry‑centric framework that unifies pattern inference and mesh synthesis, dramatically cutting inference time to seconds while achieving state‑of‑the‑art accuracy and visual realism on the Multimodal GarmentCodeData benchmark for digital fashion.
The STEM architecture replaces the Transformer feed‑forward network with a static token‑indexed embedding table, enabling lookup‑based memory that decouples capacity from compute, improves training stability, expands addressable memory, and delivers consistent performance gains on long‑context and knowledge‑intensive tasks.
The article analyzes the emerging concept of Harness Engineering—combining AI models with structured environments, standards, and feedback loops—to enable agents that can work continuously, illustrated by OpenAI and Anthropic case studies, practical design guidelines, and a three‑week adoption plan.
The article presents LoGeR, a long‑context geometric reconstruction framework that combines test‑time‑training memory and sliding‑window attention to achieve minute‑scale, fully‑feedforward 3D reconstruction with superior accuracy on benchmarks such as KITTI and VBR.
The article analyzes the limits of screen‑bound AI agents, explains how Looki’s Proactive Intelligence Engine extends OpenClaw’s capabilities into the physical world through context‑aware perception, dynamic decision‑making, and privacy‑preserving pipelines, and discusses the technical and practical challenges of building truly proactive AI.
