Wasserstein Robustness: Surrogate Losses Misalign with Robust Accuracy
New analysis shows WDRO risk rankings can diverge from robust accuracy, and proposes Wasserstein distributional attacks to measure it directly.
Research Digest
Preprints and papers, rewritten in plain English.
Sketching Your Way Through Debugging
A new interface lets programmers use pen strokes to set breakpoints and control execution, potentially making debugging more fluid.
RadGeoLoc: Radar and Particle Filtering for GNSS-Denied Localization
A new arXiv paper proposes using radar and particle filtering to localize autonomous vehicles against satellite imagery when GNSS is unreliable.
Comparing Suitability Metrics for Heterogeneous Multi-Robot Task Allocation
A new arXiv preprint compares twelve ways of scoring robot–task suitability, noting that the term has no single agreed definition.
Convergence Guarantees for LLM Agent Teams
A new theory gives LLM agent teams provable convergence, with experiments showing speedups over single agents.
Model Ability Shapes When AI Agents Seek Human Help
A new arXiv study suggests that as language-model agents become more capable, reduced reliance on human partners may reflect appropriate task division rather than poor collaboration.
FLINT: A Fixed-Budget Approach to Late-Interaction Retrieval
A new arXiv paper proposes a fixed-budget late-interaction model to reduce the storage overhead of token-level retrieval.
Knowledge as Code: Epistemology for Coding Agents
An arXiv paper argues that coding agents may be shifting knowledge from language into code, a change with epistemological consequences.
EmbodyHOI: Generating Hand-Object Motions That Fit the Robot
A new arXiv paper proposes embodiment-aware generation of hand-object interactions so reference motions are both semantically rich and kinematically feasible for humanoid robots.
Learning to Be Different: Specializing Components in Diffusion Policies
A new arXiv paper proposes making components specialize in compositional diffusion policies to improve multitask robotic learning.
LiveUIBench: Benchmark for Adaptive App Workflows
New benchmark aims to evaluate AI agents on app tasks where user requirements vary and change.
Energy Optimization Targets Artifacts in Neural Implicit LiDAR Maps
A new paper applies energy optimization to reduce jitter and artifacts in memory-efficient neural implicit maps of large-scale LiDAR scenes.
Benchmark Tests Depth Estimators on Jetson AGX Orin
A 2026 arXiv preprint benchmarks nine depth-estimation model families on the Jetson AGX Orin, measuring accuracy, speed, and energy.
Channel-Level Masks Aim to Make Learned Image Compression Slimmable
A new arXiv paper proposes channel-level mask guidance to let a single learned image compression model adapt to different compute and storage constraints.
Set-Valued Classification Improved for Long-Tailed Data
A new arXiv paper proposes using set learning to improve set-valued classifiers on long-tailed data, where fixed-size predictions often fail.
Large-scale study maps real-world LLM inference configuration practices
A new arXiv study examines how developers configure LLM inference in real-world software, treating configuration as a recurring engineering task.
Video-Point Model Forecasts Multi-View Scenes via 4D Trajectories
A new world model learns to predict future frames across cameras by forecasting point trajectories in a shared 3D scene.
WildMAE Brings Structure-Aware Masked Autoencoding to Wildlife Imagery
A new arXiv paper proposes a label-free model that reconstructs camera-trap wildlife images despite wide variation in appearance, lighting, and clutter.