Week of October 5, 2026
2026-W41
Week 2026-W41: 20 curated papers with 20 free plain-English explainers.
KaliBench: A Fine-Grained Benchmark for Cybersecurity Tool Use on Kali Linux with Runtime-Free Verifiable Rewards
A manuscript-grounded Kali Linux benchmark of 8,504 natural-language-to-CLI pairs shows no tested open-weight model tops 42% exact-command accuracy without too…
Read free explainer →
Can LLMs Reliably Annotate Bioassay Metadata to Improve Data Readiness?
PubChem’s ~2 million bioassays are almost empty of BioAssay Ontology format and detection labels; seven LLMs can often recover those fields from assay text — a…
Read free explainer →
InterEvolve: Test-Time Evolution of Reward Programs for Humanoid Loco-Manipulation
InterEvolve lets a humanoid reuse a frozen controller on new loco-manipulation tasks by evolving staged reward programs at test time — in simulation and, after…
Read free explainer →
Reconstruct, Practice, Go Real: Guided Self-Improvement for Embodied Agents
RPG improves a frozen robot execution stack by practicing in simulation and editing skills plus the system prompt — 28.6% to 95.0% on 22 held-out sim tasks, th…
Read free explainer →
VISTA: A Visual Harness for Reasoning in an Interactive World
VISTA gives a general multimodal model a lossless visual memory and active retrieval, lifting Claude Opus 5.0 to a perfect 100 Relative Human Action Efficiency…
Read free explainer →
HumanoidToolBench: Benchmarking Humanoid Tool Use from Selection to Mobile Execution
An 18-task humanoid tool-use benchmark plus 3.1k ToolBook demos shows a wide gap between picking the right tool and actually finishing the job — including on a…
Read free explainer →
Beyond Light Cones: State Preparation Complexity in Quantum Spin Glasses
For dense quantum p-spin Hamiltonians, near-ground-state energy needs Ω(n²/log n) one- and two-qubit gates — even with leftover ancillas — using a profile-comp…
Read free explainer →
SkeleWAM: Skeleton World-Action Modeling for Efficient Robotic Manipulation
SkeleWAM is a 57.1M-parameter world-action model that plans from a sparse 3D skeleton of joints, object centers, and contact points — 85.9% on LIBERO-Plus, 3.7…
Read free explainer →
A provable quantum advantage for approximate optimization via decoded quantum interferometry
In an oracle model of folded optimal polynomial intersection, decoded quantum interferometry beats every polynomial-time classical algorithm on approximation r…
Read free explainer →
ROWBench: Do Video Models Render What the Program Specifies?
PROWBench (presented as ROWBench) checks whether programmable world models actually depict the events a program logged — 170 constructed episodes, 600 proxy vi…
Read free explainer →
The Impact of a Clumpy Ambient Medium on the Dynamics and Synchrotron Emission of AGN Jets
3D relativistic MHD jets with Lorentz factor 10 pick up far more mass in a clumpy nuclear ISM — 5.3–8.8 solar masses versus 0.43 in a smooth medium — and look…
Read free explainer →
Multi-Scale Temporal Flows for Peptide Trajectory Generation
PepTIDE trains one peptide trajectory model across many time strides and an interval-aware clock, recovering rare conformational transitions better than fixed-…
Read free explainer →
Frustration-induced multiferroicity in hauerite MnS2
Below 48 K, antiferromagnetic order in pyrite MnS₂ doubles the cell, breaks inversion, and — via dipoles on S₂ dimers — yields a type-II ferroelectric polariza…
Read free explainer →
GeoFWI3D: Large-scale 3D Velocity Model Dataset for Deep Learning-assisted Seismic Imaging
GeoFWI3D releases 10,000 geologically staged 96³ velocity models with Vp, reflectivity, geologic time, and fault/salt masks — plus baseline tasks from FWI to 3…
Read free explainer →
Bayesian Image Reconstruction with Spatially Variant PSFs in X-ray Astronomy
Information-field-theory deconvolution plus a fast spatially variant PSF model beats Richardson–Lucy variants on a synthetic test and sharpens Chandra images o…
Read free explainer →
Fold'EM: Direct atomic structure inference from Cryo-EM particles
Fold'EM skips the usual density-map middleman and infers atomic models from few cryo-EM particle images by combining protein generative priors with the particl…
Read free explainer →
Weak non-Landau-type contributions to the diamagnetism of graphite
First-principles bookkeeping finds graphite’s extra, non-Landau diamagnetism is about −2.7×10⁻⁷ emu/g — comparable to in-plane graphite and to C60/C70 — while…
Read free explainer →
Analysis of grid instabilities in particle-in-cell codes based on a meshfree approach
A meshfree-then-grid analysis of PIC aliasing shows energy-conserving discretizations stay stable for a cold plasma, while momentum-conserving ones produce con…
Read free explainer →
SW1D: A Python package for seismic surface wave solutions in 1-D elastic media
SW1D is a Python port of a legacy propagator-matrix code for Rayleigh and Love dispersion and synthetic seismograms in layered elastic media, benchmarked again…
Read free explainer →
Birth-Potential Multigroup Fluid Model for Ballistic Electrons: Breakdown and Cathode Sheath Regimes
A 10–30-group fluid model that bins runaway electrons by birth potential reproduces argon Paschen’s left branch and cathode-sheath I–V to within a few percent…
Read free explainer →