Motion Models Can Match an Action While Losing Its Source
Preprint dated 29 September 2026: Tests suggest that action recognition can miss whether a model preserved a clip's source-specific movement.
Search
2,451 results
341–360
Story AI systems struggle most to keep images consistent
Preprint: In a 900-transition benchmark, orchestration led overall, but visual continuity remained the clearest weakness.
AI router reports stronger model choices with little evidence
Preprint results show RouteFM leading comparison routers on held-out and multimodal tests, with its largest reported margin when evidence was scarcest.
Mathematical Attractor Shows Patterns Can Fix Its Geometry
Preprint: A theorem-level analysis finds exact affine and conformal matches when neutral dynamical systems share the same combinatorial pattern.
MRI AI model tops test of enlarged perivascular spaces
Preprint: A model that predicted both MRI scores and PVS locations led the comparison, but the study's single data split limits how far the result can be taken.
Robust losses improve neural network feature selection under contamination
A preprint reports steadier prediction and feature selection in difficult settings, but no single loss wins every scenario.
AI assistant asks fewer questions and scores better on benchmarks
Preprint reports that REVOIR outperformed several baselines, but its evidence comes from simulated users and has not yet been tested with real users.
New mathematical method aims to preserve direction in data
Preprint: The approach separates geometric structure from directional flow, but its evidence so far comes from synthetic tests.
A mathematical model finds predator spread varies by direction
Preprint: A road-field model traces prey and predator expansion when predator diffusion is faster on the road.
Chip-scale photonic circuits demonstrate memory for optical computing
Preprint reports photonic circuits that retain state, detect patterns and support simulated drone and language-generation tests.
Interleaved ToF-SIMS Tracks Isotope Fractions Down to 0.07% in Lab Tests
Preprint: A laboratory method combined short and long pulse phases to measure much smaller isotope signals in one depth profile.
Local chemistry may explain why hafnia capacitors behave differently
Preprint: Local imaging suggests oxygen-related reactions may separate switching regions from leaky ones.
Coding agents struggle to turn big software plans into code
Preprint: A 100-task benchmark found that the strongest tested setup completed only 26.3% of large software changes.
Quantum model finds persistent nonclassical patterns in a double well
A preprint finds negative Wigner regions persist in the modeled ground state, while changing the representation smooths them away.
Ultrasound probes a hidden acoustic resonance in an open device
Preprint: A laboratory study reports far-field access to a symmetry-protected acoustic mode through nonlinear sound conversion, without adding a radiative channel.
New theorem sets a sharp bound for Dirac zero modes
Preprint: The theorem links a non-trivial Dirac zero mode on the sphere to a sharp lower bound on the conformally invariant magnetic-field norm.
Veer records fewer deceptive outcomes in web-agent tests
Preprint: Veer recorded far fewer deceptive outcomes across controlled benchmarks while keeping task completion competitive, but the evidence is not a real-world safety test.
Automated tools linked to lower energy forecast error in 18 tests
Preprint: An automated feature-building system was linked to lower forecast error and lower reported battery-storage regret, but the battery test covered one 55-day UK property case.
Offline Tests Show Gains for Less-Popular Items in Recommendations
Preprint: The authors report higher average top-20 scores across six datasets, but Google was a clear exception.
Video Models Reach 56 on a Benchmark for Preserving Input Objects
Preprint evaluates eight systems on 500 cases, using a fixed reference to test whether objects remain present, identifiable and structurally consistent.