AI coding prompts linked to higher costs in benchmark study
Preprint: In 2,700 controlled runs, prompt details and thinking effort were associated with large cost differences, and one cheap probe was associated with lower forecast error.
Search
2,111 results
761–780
Radio observations hint at hidden micro-flares on AU Mic
This preprint reports variable 12 to 26 GHz emission, two clear flares and radio behavior the authors say is consistent with trapped energetic electrons.
Unseen Camera Paths Expose Weaknesses in 3D Reconstruction
Preprint: A five-scene benchmark found lower image quality when models were tested on camera trajectories absent from training.
Real-time spatial-audio vocoder reports higher spatial scores
Preprint: CSAVocoder reported higher spatial scores in baseline and listening comparisons, while audio-quality results were mixed.
In one five-tier failure simulation, Standard Retry had lowest success
Preprint: A review of 200 GitHub repositories was paired with a three-scenario simulation of four retry strategies.
Non-Hermitian lattice models show distinct Euler-topology phases
Preprint: Three theoretical models link integer Euler labels with Wilson-loop winding and spectral signatures; the two indices are not reliable in modeled critical phases.
Lightweight AI keeps simulated drone network near optimum
Preprint study reports that a bandit-based controller exceeded 90% of an optimal data-rate benchmark, but it has not yet been tested in hardware or the field.
Explicit elliptic K3 surface reported with exact rank 17 over Q(t)
A version-1 arXiv preprint reports a fiber with rank at least 28 and infinitely many elliptic curves over Q with rank at least 19.
Exciton-polariton condensate develops coherence in two stages
Preprint: The condensate showed early growth, partial collapse and revival of coherence before entering a quasi-steady state extending across it.
Humanoid robot shows separate control of force and twisting
An arXiv preprint dated 26 Aug 2026 reports a controller tested in simulation and on a 23-degree-of-freedom Unitree G1 humanoid.
Phase maps expose hidden structure in modeled light scattering
A preprint proposes a two-index polarization phase that adds detail to simulated thin-film and chiral-sphere scattering maps.
Image AI Framework Reports Similar Accuracy With Less Data
An arXiv preprint reports a modular vision-transformer system that reached 85.83% ImageNet accuracy at 1.24 bits per pixel.
Adaptive retrieval preprint reports higher mAP with fewer vectors
Preprint: AdaptiveEmbed reported higher average mAP with around two embeddings per sample, compared with fixed-capacity baselines using 3 to 40 embeddings per sample.
Two-Room Simulation Shows Howling Suppression and Speech Trade-Off
Preprint: A two-room, three-terminal verification simulation showed qualitative howling suppression, while over-muting fragmented desired speech and reduced intelligibility.
RotDroid detected more Android rotation bugs in equal-budget tests
Preprint: With 30 minutes per app, RotDroid recorded more true bugs than DLD and a DOC-based method, but DOC reported slightly higher precision.
Metasurface reveals multiple topologies in entangled light
Preprint: OAM measurements exposed several distinct topological states from one high-dimensional quantum state.
New graph theory bounds sharpen the hunt for a fugitive
Preprint derives exact cop numbers for Fibonacci and Lucas cubes and improves a bound for finite median graphs.
AI Agents Make Early Drawing Gains, Then Often Lose Accuracy
Preprint: A 25-agent 2D benchmark found strong early progress but weak long-horizon correction, with a large gap to a data-policy reference.
AI tool builds 4D LiDAR labels without manual annotation
Preprint: A video model and sparse clicks produced labels on SemanticKITTI, while downstream scores remained below models trained on full ground truth.
Open Air-Combat Simulator Reports Faster Reinforcement-Learning Tests
Preprint: BVR Sim pairs heterogeneous aircraft models with a faster C++ backend, while its policy-transfer and multi-agent results remain limited.