Quantum decoder matches exact results but slows on larger codes
A preprint reports near-exact decisions in simulations, while larger-code tests and hardware pilots expose runtime and model limits.
Search
2,230 results
1041–1060
Protein AI Models Contain Biochemical Signals Tied to Fitness
Preprint: Removing biochemical patterns from protein embeddings lowered classifier performance, while randomized projections had little effect.
Event-token model gains a small edge in recommendation tests
Preprint: An industrial-scale study reports modest offline gains in ranking and retrieval, but does not test user behavior or revenue.
Hybrid Method Reports Higher MapReduce Throughput in Simulation
Preprint: DDPG-CVX outperformed five allocation methods in a five-device MapReduce model, with the largest gap against MAXF.
SGD study finds adaptive bounds can justify earlier stopping
Preprint: A confidence-sequence method follows the observed SGD trajectory and is designed to certify target accuracy without a fixed horizon.
Preprint links seven usability factors to shopping app ratings
An analysis of five shopping applications used in Pakistan reports a strong in-sample model, but leaves questions about how widely it applies.
AdaVDR reports higher accuracy on two video research benchmarks
Preprint: An adaptive video-language system called AdaVDR reports higher benchmark accuracy than comparison models on VDR-EE and VideoDR, but the evidence is limited to benchmark tests.
Finite test identifies when toric blow-ups need no extra generators
This preprint reports algebraic criteria that reduce an infinite saturation problem to finite checks and gives exact results for several toric spaces.
Solar 'Fishbone' Structure Shows 18 Repeated Brightenings
Preprint: A 2015 solar active region showed 18 recurrent brightenings and moving plasma, but the data do not show that one reconnection site triggered the next.
Seven vision-language models struggle when first-person video and dialogue disagree
Preprint: A benchmark of seven vision-language models found low accuracy when first-person video and dialogue conflicted, along with wide variation in when models chose to intervene.
Sparse PINN tests find no single best pruning rule
An arXiv preprint reports lower Gray–Scott residuals for PI-SAP, while accuracy rankings shift with sparsity, width and equation.
New Formula Calculates Best-Worst Choice Polytope Dimension
Preprint gives an explicit formula for a mathematical structure built from all linear rankings and deterministic best-worst choices.
New optimization method reaches tight targets in fewer steps
Preprint: Decomposed Bilevel Search reached demanding SLOPE and group-lasso targets in fewer iterations than tested baselines, but evidence is limited to numerical benchmarks.
Model Links Anchor Density to Sudden Shifts in Hypergraph Cores
An arXiv preprint models how anchor links alter core transitions under first- and second-neighbor pruning rules.
Directed networks share a failure threshold, but redundant cores lag
Preprint: An arXiv analysis finds that giant strongly biconnected components emerge at the same threshold as giant strongly connected components, then grow more slowly.
Quantum gravimeter tracks a ship without satellite navigation
Preprint: On a 29-m research vessel, a hybrid quantum gravimeter kept its gravity-aided position within about 1 nautical mile during an approximately six-hour, 83-km evaluation.
LLM agent reports wider kernel coverage, mixed retargeting results
Preprint: KOPE reports broader AscendC benchmark coverage than CANNBot, while source-memory reuse is associated with higher K3 coverage but lower fixed-intersection speedup.
AGN model finds core heating balance in seven mock clusters
A preprint modeling lobed AGN feedback reports core heating offsetting cooling in seven of ten mock clusters, while cooling exceeds AGN heating above 100 kiloparsecs in seven even at a 100% duty cycle.
AI Emotion Steering Shows a Trade-Off in Model Scores
Preprint: Latent-vector steering was associated with higher model-rated affective scores in three instruction-tuned models, while stronger settings coincided with lower semantic-preservation measures.
Persistent rules were linked to steadier approval of AI interfaces
A Preprint study found that Maru users accepted more generated interfaces overall, but cross-topic rules drew more negative feedback.