Modified-gravity model fits cosmic expansion, but results vary
Preprint analysis finds the square-root exponential model favored by some supernova compilations, while a power-law model stays close to flat ΛCDM.
Search
2,451 results
681–700
New AI training method shows higher average scores on math tests
Preprint: OPDVR combines teacher guidance with verifier feedback; its average scores were higher in two settings, but not on every benchmark.
Kernel ensemble matches or beats baselines while cutting computation
Preprint: KENDO led most benchmark comparisons in optimization and active learning, but its kernel menu still must be specified by hand.
Targeted neuron tests flag sharp drops in LLM code scores
Preprint: A gradient-free scoring method selected small sets of language-model neurons for suppression tests, where code and math scores fell more than scores on other benchmarks.
Preprint tightens a key finite-scale test of Weil positivity
A version-2 arXiv study certifies positivity at support 1.6 and finds upper bounds that fit a proposed Landau–Widom decay law, while leaving the main asymptotic claim unproved.
Product-form centers and linear codes match the optimal covering rate
Preprint: A theoretical study finds that product-form centers and linearity preserve the asymptotic rate, but explicit codes reaching it remain an open problem.
Models link magma chemistry to starkly different sub-Neptune atmospheres
Preprint models connect interior composition with contrasting gases, radii and synthetic spectra, while stressing major thermochemical uncertainties.
AI Router Balances LLM Quality and Cost Across Four Tests
An arXiv version 2 preprint dated 30 Aug 2026 reports benchmark results from a system that changes model choices as a workflow unfolds.
Instruction-tuned AI leads on a Sanskrit glossary benchmark
Preprint: Fine-tuned systems beat prompting and alignment baselines, but phrase segmentation and word order remain major challenges.
Endpoint regularity is proved with uniform source control
Preprint: The theorem covers superquadratic Hamilton-Jacobi equations and shows bounded source norms alone are not enough.
Graphene device links resistance peak to magnetic history
Preprint: Low-temperature measurements on a single device found transport states that varied with gate-voltage path and in-plane field.
Mobility model improves predictions across four disease groups
Preprint: DeMMO delivered the strongest held-out predictions across five clinical outcomes, while its mobility markers and disability thresholds still need clinical validation.
Battery-free sensor system reaches every first rendezvous in tests
Preprint analysis reports complete first contacts and faster sustained exchanges for MagPie, but says controlled tests do not establish field-scale energy efficiency or multi-hop performance.
Chinese search correction system reports fewer misspellings
Preprint: GUIDE combines phonetic and visual clues to correct short Chinese queries, with tests on benchmarks and production search traffic.
New preprint reports proof of a three-part edge split for cubic graphs
An arXiv version 2 preprint dated 3 September 2026 says the 3-Decomposition Conjecture holds for finite connected cubic loopless multigraphs, while other extremal questions remain open.
Study Links Varied AI Refusals to Less Effective One-Vector Attacks
Preprint findings from OLMo-2-0425-1B-Instruct link varied refusal starts with higher stable rank and smaller changes after ablation.
Quantum code study ties 'magic' to word-set geometry
An arXiv preprint reports exact witness thresholds for selected code families, while identifying a major limit for larger ones.
TailorCoPilot Linked to Higher Novice Success in Pattern Tests
Preprint: TailorCoPilot users finished garment-pattern revision tasks faster, with lower workload and higher expert-rated quality than baseline users.
AI Memory Stores Film Facts, But Struggles to Find Them
Preprint: LoRA adapters stored some film knowledge, but query-based retrieval could not reliably find the right memory.
Quantum model estimates tiny pair creation in binary stars
Preprint: A theoretical study derives formulas for scalar, fermionic and vector fields and applies them to an idealized weak-gravity binary-star model; its relativistic enhancement remains speculative.