Self-Driving Models Improve in Tests with Local Drone Data
An arXiv preprint reports lower offline errors after self-driving models were trained on aerial traffic data from Songdo, South Korea.
561–580
An arXiv preprint reports lower offline errors after self-driving models were trained on aerial traffic data from Songdo, South Korea.
Preprint outlines a two-device system that can be checked offline, but leaves major security and identity questions open.
Preprint: A theoretical study proves the empty-set-free result through height four while narrowing, but not resolving, the height-five case.
This preprint models a propeller and wing in cruise and hover, with cruise efficiency associated with advance ratio and hover metrics more closely associated with RPM.
Preprint: A benchmark of 15 image encoders found steep losses in cross-dataset accuracy and confidence, while CBR delivered a limited gain.
Preprint analysis finds that packing many binary variables into fewer qubits can shrink signal margins and increase decoding or measurement costs.
Preprint: In 4,400 simulated runs, competing messages were associated with different belief endpoints and a reversal in how network density tracked upward movement.
Preprint: A 20-design suite spans foundational 2.5D, basic 3D and complex hybrid architectures, using reusable virtual chiplet models.
Preprint: For every sufficiently large n, the minimum is 2n−4; a graph with at most 2n−5 edges can be relabelled to share at most three edges.
Preprint: Opus 4.8 scored highest among three Claude models, but none triggered a crash in 13 of 77 challenges.
This preprint describes a self-consistent calculation that gives model water, hydrogen and lithium systems constant-potential boundaries, while real interfaces remain untested.
Preprint: An audit of four automatic metrics found different strengths across timing and content disruptions, with no clear all-purpose score.
At weak coupling, one calculated channel grows below a critical loop size and decays above it under an exploratory infrared prescription, while the leading-in-colour channel remains unresolved.
An arXiv preprint models a 1,000 MW, 12-hour project in Rio de Janeiro and compares an optimized reservoir layout with an initial terrain-based design.
Preprint comparison reports higher BFN-RL success on Sokoban and DoorKey, but higher categorical-diffusion success on Empty-Random and BlockedUnlockPickup.
Preprint: A theoretical calculation finds weaker 21-cm absorption than ΛCDM, but it uses a ΛCDM-based spin-temperature input.
Preprint: Molecules in helium nanodroplets spun beyond expected limits, while alignment loss suggested different thermalization times.
The multi-view system restricts cross-view information and is reported as best on nine of 10 Real-IAD measures and all 10 MANTA-Tiny measures.
Simulations found broad monitor visibility and concentrated rewards, while leaving the bounty needed for participation and the market equilibrium unresolved.
Preprint analysis finds that closed-class word neighborhoods expand and contract earlier than open-class neighborhoods across transformer layers.