AI system tracks sentence-level sentiment targets in benchmark tests
Preprint tests SenseShift on stories and reviews, with stronger automated sentiment adherence but more mixed human judgments.
Search
2,451 results
1801–1820
Quantum checker reaches 100-qubit tests but hits limits
Preprint: QisMC handled selected large quantum programs, but its speed depended sharply on circuit structure and numerical precision.
Study reports lower errors for data-driven industrial-load models
Preprint: In three simulated industrial-load cases, two reduced formulations used fewer listed variables and had lower reported error scores than SAL and, where available, OVB.
Coordinate solver leads tests of larger absolute-value systems
An arXiv preprint reports that CPPC was fastest in tested systems of 1,000 or more variables, while spectral methods led in other regimes.
Preprint Finds Zero-Index Cases in Lie Superalgebra Modules
The proof-based study bounds Borel-subalgebra indices and extends zero-index results to nilradicals, their ideals and dual abelian-ideal modules.
AI image grader weighs visual quality against prompt fit
An arXiv preprint reports strong benchmark agreement from a model that combines collaborative and adversarial paths, while fresh human-rating tests remain needed.
Naphthalene’s Triplet State Shows a Distinct Infrared Signature
Preprint: The first reported laboratory measurement differs from the ground-state spectrum and suggests possible markers for naphthalene’s triplet state.
Quantum method targets uniform accuracy for amplitude estimates across depths
Preprint: A theorem and simulations report uniform angle guarantees across the full Grover-angle range, with different depth windows producing different error and shot requirements.
Adaptive satellite scheduler leads in simulated maritime tests
Preprint: An offline-learning version of ant colony optimization reported higher observation benefit across 14 simulated scenarios, while operational applicability remains untested.
Language-model method predicts target-task mechanisms before fine-tuning
A preprint under review reports that a small probing update can forecast which internal parts matter after targeted tuning, with the strongest pattern on controlled benchmarks.
Targeted Checks Cut AI Errors From Outdated Memory
An arXiv preprint found that directing one of two verification checks to a critical provenance path sharply improved decisions in controlled model tests.
New model reports 36.80 class-level contrast in STAR data
This arXiv Preprint applies a class-level model to Project STAR data, with the reported contrast depending on graph and normalization choices.
IAPO Scores Higher Than Matched GRPO on Several Qwen3-4B/8B Agent Tests
A preprint reports higher results for IAPO on several service-agent benchmarks, while multi-turn function-calling scores were comparable.
Correlated uncertainty model tops held-out likelihood scores
A preprint reports higher held-out likelihood when a low-rank model captures correlated uncertainty across outputs.
AI image system scores higher on unseen imaging combinations
A preprint reports stronger factor alignment for X-MULTI, while some sensor and viewpoint predictions remained unreliable.
ADEPS Tops Signal-Distortion Tests Across Mic Layouts
Preprint: In simulated speech scenes, ADEPS led on signal-distortion scores across the tested arrays, while other measures were mixed.
Flexible Self-Consistency reports lower token use in language-model benchmarks
Preprint: The method uses predicted output entropy as a continuous difficulty signal to adjust the reasoning-path budget for each question.
INCEPT posts strong results across varied EEG benchmark tasks
Preprint researchers report that INCEPT transfers across signal, brain-state and brain-health EEG tasks, while the benchmark does not test patient outcomes.
Preprint reports sublinear regret with feasible queries only
The theoretical method uses one feasible query each round, but has not been tested on data or real-world problems.
Quantum kernel tops synthetic tests, but trails on one fraud benchmark
Preprint reports leading scores on interaction-defined synthetic tasks and one fraud dataset, but a classical kernel led on another.