AI Code Search Rankings Shift When Identifiers Are Renamed
Preprint: An offline evaluation found that carefully chosen identifier changes altered retrieval across models and five programming languages.
Search
2,451 results
1681–1700
Electrical Pulses Are Linked to Selective YBCO Resistance Shifts
A preprint reports branch-specific resistance changes, opposite inferred oxygen-vacancy directions and faster relaxation at higher temperatures.
Preprint outlines a route to weighted bounds on Morrey spaces
An expository mathematical note presents a model sparse-operator estimate, carries it to Calderón–Zygmund operators under a stated condition, and marks the proof's limits.
Grounded dialogue policies retrieved images similarly; DCP-B lagged
Preprint: Three policies had similar image-retrieval scores, while DCP-B showed weak retrieval and very large high-revision grounding shifts.
LLM data agents can get answers right despite flawed reasoning
Preprint: A 100-example test found that answer accuracy and trace validity often told different stories.
New Algorithm Reuses Sparse Graphs for Dynamic All-Pairs Queries
Preprint: A randomized method maintains approximate maxflow and effective-resistance answers as an undirected graph only gains or loses edges.
Mathematicians find a rotated semicircle law in random polynomials
Preprint: A theoretical analysis says permanental roots from Gaussian matrix ensembles approach a semicircle on the imaginary axis as matrices grow.
Quasar disk study finds size pattern at odds with thin-disk theory
Preprint analysis of seven lensed quasars reports a 13.3-light-day characteristic source size at 5007 Å and a wavelength trend that differs from the thin-disk prediction.
Critical operator limit reflects prime-coordinate concentration
Preprint: A proof-based analysis finds a uniform boundary limit, exact one-prime equality cases and certified finite approximations.
Adaptive wireless coding reports gains in changing networks
Preprint simulations and a 10-device testbed report stronger delivery under changing conditions, but the evidence remains limited to modeled and proof-of-concept settings.
3D-Printed Steel Shows Higher Sputter Yields at Higher Ion Energy
Preprint: Tests measured sputtering in 316L steel and briefly operated tungsten-rhenium grids in a KDC-40 ion source.
GaN study finds emitters at interface and a target depth
Preprint: An emitter was observed at the depth of a low-temperature GaN layer in a design targeting 5 micrometres below a sapphire surface, while natural emitters aligned with the GaN-sapphire interface.
AI Recovery Method Reports Fewer Hidden Attacks in Hardware Code
Preprint reports lower attack rates and higher functional scores after teacher-student recovery, but tests used synthetic poisonings and fixed benchmarks.
A Mathematical Model Puts a Value on Human Expertise
An arXiv v1 preprint dated 26 Aug 2026 asks whether a decision-maker’s belief can improve worst-case choices when key parameters are unknown.
Sparse expert model reports gains against three attack norms
An arXiv preprint reports higher combined accuracy for Robust CurveMoE than the strongest named baseline across three tested threat models on CIFAR-100 and ImageNet-100.
New Preprint Finds Polynomial Rings Can Lose a Finiteness Property
Explicit constructions show that the strong finite type property can fail after one polynomial variable is added, with examples across several characteristics.
Linear multiphase-flow method shows stability, but one cavity test fails
Preprint: A second-order linear scheme reports near-second-order time accuracy and stable benchmark runs, but an unregularized cavity case loses boundedness.
Numerical method recovers scattering data from dark solitons
Preprint: A 26 August 2026 arXiv version extends the classical Boffetta–Osborne transform and tests it on three modeled dNLSE wave fields.
New analysis improves row-coreset bounds for subspace problems
Preprint: The analysis reports ε⁻² accuracy dependence in both p<2 and p>2 settings, but the p>2 rank-accuracy comparison remains open.
LoRA theory ties adapter rank to attention error
Preprint: A mathematical framework derives conditional error bounds for matching a fixed attention target, but does not predict task performance.