Week of November 16, 2026
2026-W47: 20 curated papers with 20 free plain-English explainers.
Start with #1 → FORGE: Force-Aware Multi-Granularity Control for Efficient Humanoid Loco-Manipulation
FORGE: Force-Aware Multi-Granularity Control for Efficient Humanoid Loco-Manipulation
A humanoid can walk, balance, and push without a full-body motion-capture suit or wrist force sensors. FORGE takes sparse wrist poses, a coarse motion command,…
Read free explainer →
Unbundling Multimodal Vision Encoding
A multimodal model does not need a vision encoder that was pretrained in a separate stage. What it needs is vision capacity of its own, and that capacity can s…
Read free explainer →
K2V: Generative Keyframing For Physically Plausible Video Generation
Video diffusion can paint convincing texture and still break basic physics, because the whole motion has to be invented inside one denoising process. K2V lets…
Read free explainer →
Scalable Pauli propagation through $\mathbb{F}_2$-linear partitioning
Pauli propagation can follow more than 10^11 Pauli strings when the strings are split so that each gate talks only to a bounded set of neighbors. The wall was…
Read free explainer →
AgentSSL: Can MLE Agents Leverage Unlabeled Data?
Coding agents can assemble semi-supervised training programs that beat strong human baselines in 8 of 12 settings, by up to 21.8 percent. Choosing which of tho…
Read free explainer →
Partition of Unity Data-driven Finite Element Method for Anisotropic Diffusion in Magnetized Plasmas
Heat and particles in a magnetized plasma shoot along the field far faster than they leak across it. PU-DDFEM builds a reusable local basis for that extreme an…
Read free explainer →
ArchSpace: Recording, Reviewing, and Referencing LLM Architecture Research
Architecture ideas for large language models are tried in private and, in public, often compared at a size that does not predict the real training run. ArchSpa…
Read free explainer →
Impact of edge-localized resonant magnetic perturbation on L-H transition in the KSTAR tokamak
On the KSTAR tokamak, an edge-localized resonant magnetic perturbation kept the plasma from entering high-confinement mode. The edge flow shear fell below the…
Read free explainer →
Learning When Diffusion Language Models Should Regenerate
Diffusion language models can rewrite a passage, but they often reprint the same mistake when the surrounding words stay fixed. DiVR is a verifier that decides…
Read free explainer →
XBot-Omni: An Omni-Modal Model for Embodied Understanding and Interaction
A robot that understands a scene still has to answer with a voice, a posture, and a face. XBot-Omni adds an Actor next to the Thinker and the Talker so body mo…
Read free explainer →
SALA: Semantically Aligned Latent Actions Enable Dual-Axis Scaling of Embodied Foundation Models
Latent actions learned from human video often group clips by how the scene looks, not by what the hands did. SALA uses language descriptions of unpaired human…
Read free explainer →
Ensemble phase transitions in measurement-unitary hybrid quantum circuits
A circuit that mixes gates and measurements can change which measurements actually steer an entangled partner qubit, even while the average state stays complet…
Read free explainer →
PDO Drug Screening Data for Evaluating and Training AI Drug Response Prediction Models
Most drug-response models learn from immortal cell lines that are not the patient's tumor. Four screens on patient-derived organoids add multi-omics profiles,…
Read free explainer →
Peak Performance in Matched Filtering: Fused Correlation and Statistically Controlled Screening on CPUs and GPUs
Searching a huge bank of templates for a weak signal spends most of its time writing spectra it immediately throws away. Fusing the multiply, the inverse trans…
Read free explainer →
Silscore: modelling siRNA potency with positional thermodynamics and data-selected nucleotide-identity terms
Two small interfering RNAs aimed at the same gene can silence it very differently. Silscore predicts that potency from duplex stability, a position-by-position…
Read free explainer →
Euclid data release 1 (DR1) - SIR spectroscopic calibration
Euclid's slitless infrared spectra cannot calibrate themselves. Data Release 1 depends on separate calibration pipelines that treat the whole focal plane as on…
Read free explainer →
Inter-Nanoparticle Repulsion Effect on Self-Organization in Polymer Nanocomposites: A Langevin Dynamics Simulation Study
In a tight polymer box, turning up the repulsion between nanoparticles walks one droplet through four shapes: a compact drop, a slab, a crystal, and a disperse…
Read free explainer →
Library of peptides: Analysis of a complex reactive mixture with the use of statistical physics and inference
A self-assembled peptide mixture can hold dozens of species whose spectra sit on top of each other. A Potts-style energy model plus the law of mass action read…
Read free explainer →
HybridCT: ML Accelerated Tomography for Climate Relevant 3D Cloud Observations
Low clouds are a large hole in climate forecasts partly because their insides are hard to measure in three dimensions. HybridCT turns nine viewing angles into…
Read free explainer →
Upscale energy transfer from mesoscale convective systems to African easterly waves
West African storms do not only ride the larger easterly waves. In a kilometer-scale model, the storms feed kinetic energy up to those waves, especially inside…
Read free explainer →