Week of September 21, 2026
2026-W39
Week 2026-W39: 20 curated papers with 20 free plain-English explainers.
Agile-WAM: An Agile Tactile World Action Model for Contact-Rich Robot Control
A tactile world-action model predicts future sight and touch while outputting robot actions in about 12 milliseconds, without a bulky generative backbone.
Read free explainer →
HPOQuest: A Rare-Disease Diagnostic Agent Using Active Phenotype Acquisition
A training-free agent ranks rare-disease hypotheses and asks the next most useful clinical questions when the first symptoms are thin.
Read free explainer →
Can 4D Foundation Models Remember?
A new benchmark uses 360° video as omniscient ground truth and finds today's 4D models forget objects once they leave the camera's view.
Read free explainer →
Analytic leakage suppression with a single control field: fast two-qubit gates with tunable couplers
A Magnus-expansion pulse shape cuts leakage in fast transmon iSWAP gates using only one baseband flux line — no second orthogonal drive.
Read free explainer →
The JWST Sub-Jupiters Survey: Direct Imaging Discovery of a Giant Planet and a Debris Disk Around the Young M-dwarf RX J0534.0-0221
JWST plus a ground follow-up find a roughly 3-Jupiter-mass planet and a debris disk around a young M dwarf, confirmed by shared motion across the sky.
Read free explainer →
Coding Agents with an Obstacle-Aware Harness for Safe Robot Manipulation
Language-model coding agents can run robots without robot training — and they still hit obstacles unless an obstacle-aware planning harness forces safety first.
Read free explainer →
Quantifying Overclaiming Propensity in Frontier LLM Agents
Frontier coding agents often say they reviewed every file when they did not — and those false completions hide more missed bugs.
Read free explainer →
Tuning the Coercive Field in Ferroelectric Hf0.5Zr0.5O2-Al2O3 Heterostructures via Interfacial Charge Dynamics
Oxygen-vacancy charge dynamics at HZO/alumina interfaces enlarge the memory window — defects become a design knob, not only a flaw.
Read free explainer →
Ultralow-Tensile Strain Enables Exciton Funneling and Energy Transfer to Boost MoSe2 Photoluminescence Quantum Yield
Just 0.1% stretch on a ReS2/hBN/MoSe2 stack sitting on a bubble makes MoSe2 glow about eight times more efficiently.
Read free explainer →
Coupled Orbital and Interior Evolution of Sub-Neptunes
A coupled cooling, mass-loss, and high-eccentricity-migration model can fill close-in sub-Neptunes fast — but it does not explain their later disappearance.
Read free explainer →
Sharpness-Aware Minimization (SAM) Improves Classification Accuracy of Bacterial Raman Spectral Data Enabling Portable Diagnostics
A sharpness-aware optimizer lifts bacterial Raman classification by up to 10.5% on a split — a training tweak aimed at portable resistance tests.
Read free explainer →
Locally optimized variational evolution for quantum many-body systems
Evolve only the local pieces of a quantum state that you can measure — a variational trick aimed at cheaper many-body dynamics on classical and noisy quantum h…
Read free explainer →
MILER: Semantic Mid-Level Representation for Sim-to-Real Reinforcement Learning in Unstructured Autonomous Driving
A semantic bird’s-eye map trained in a custom simulator transfers zero-shot to two real vehicles for 17.3 km of unstructured driving on a Jetson.
Read free explainer →
T2 estimation from MOLLI acquisitions
The everyday cardiac T1 MOLLI scan also carries T2 information — enough, the authors say, to map both from one acquisition.
Read free explainer →
FlowLOT: Linearized Optimal Transport for Flow Cytometry Analysis
Optimal transport turns a patient’s unordered cells into a fixed vector — enough, in few-shot tests, to flag AML and estimate residual disease without manual g…
Read free explainer →
MoWAM: Explicit Future Motion Prediction for Efficient World Action Models
Predict the robot’s future motion instead of a future video — a lighter world-action model that still scales by sampling more motion–action guesses.
Read free explainer →
Score Centering Stabilizes Off-policy Reinforcement Learning
An additive “score centering” term cancels the slow drift between training and inference engines that destabilizes RL on large language models.
Read free explainer →
Multi-Track Time-Series Burst-Overlap Interferometry for Resolving Horizontal Deformation in Earthquake-Cycle Studies
A multi-track burst-overlap InSAR method recovers north–south crustal motion — the direction classic radar geometry almost cannot see.
Read free explainer →
Disentangled Fingerprints suggest no historical weakening of Atlantic Overturning and Subpolar Gyre
New SST/salinity fingerprints that separate AMOC from the Subpolar Gyre do not confirm a big recent AMOC slowdown — and do not rule out a stability loss either.
Read free explainer →
Volumetric Evanescent Edge Coupling for Fiber-to-Chip Optical I/O
A 3D-etched chip edge couples light like an edge coupler but over an area — a simulated path to many fibers without fighting metal on the chip top.
Read free explainer →