Confidence-guided selection linked to higher robot-model scores
Preprint: A benchmark study reports higher scores across four robot-model measures, while the Semantics result remains uncertain.
Search
2,230 results
1061–1080
French ASR test reports high LLM agreement on transcript choices
Preprint: GPT-4.1 and Qwen3.5-35B closely matched human pairwise judgments, while embedding scores varied by model and configuration.
Study reports higher database code coverage with AI SQL tests
Preprint: DBcover had the highest reported line coverage in PostgreSQL and MySQL, and was also evaluated in two KingbaseES subsystems.
Feature-distilled CLIP scores higher on complex vision tests
Preprint: Hidden features from multimodal teachers were transferred to CLIP, with higher reported averages in compositionality, classification and retrieval.
AI Memory System Keeps Conflicting Evidence in View
Preprint: PolyMemDB stores the history behind an AI agent's memories and illustrates its evidence scoring in three worked scenarios.
Proof Establishes Exact Product Rule for Bernstein-Sato Polynomials
Preprint: A formal proof shows that the Bernstein-Sato polynomial of a product ideal equals the product of the two individual polynomials.
Mathematical model yields global solutions for two-phase flow
Preprint: A mathematical analysis reports global quasi-strong solutions and an energy equality for a Voigt-regularized model of two-phase viscous flow.
Chinese metaphor AI shows a trade-off between accuracy and consistency
Preprint: A comparison of four conditions found the strongest native score with fine-tuning, while an expert-informed Skill had the narrowest observed spread across three Chinese datasets.
Concept-Based AI Outperformed No Support, but Did Not Consistently Beat Label-Only AI
Preprint: Two experiments found higher accuracy with concept displays than no support in email and bird classification, while interactive controls showed no consistent full-sample advantage over a label alone.
Three post-training compression tests showed a sharp accuracy drop
Preprint: A graph-based framework was applied after training to three networks, with accuracy nearly steady at small allowed error values before sharp declines appeared.
Unified AI model leads most hate-speech benchmark tests
Preprint: HIPPO scored higher than separately trained systems in 14 of 17 comparisons, while cross-lingual scores were mixed.
Preprint identifies alloy families with large Hall signals
A computational screen of 2,251 tetragonal alloys highlights NiPt- and CoIr-based candidates, but experiments are still needed.
Rubric-based Qwen3-VL variant tops two benchmark averages
Preprint: A comparison of three Qwen3-VL-derived training variants reports higher average scores for rubric-based GRPO, but not across every benchmark.
Robot-learning method reports higher success on held-out tasks
Preprint: RA-VLA retrieves action-aware robot demonstrations during evaluation and reports higher average success on simulated LIBERO and real-world UR5e tests.
Models suggest dark-matter halos could weaken neutron-star lensing
A preprint reports lower modeled deflection inside extended halos, with the predicted effect far below current direct-imaging resolution.
Model finds directional volatility links among UK firms
This preprint reports a model that learned a sparse, directional volatility network for 20 UK-listed firms and had lower errors than preset networks.
Satellite Model Finds Coherent Links Could Reach More Power
Preprint modeling compares coordinated satellite links and finds quadratic received-power scaling under coherent coordination, but the result remains theoretical.
Local Fairness Can Demand Much Larger Clustering Slack
Preprint: A study of fair hierarchical clustering finds that local checks can require much larger slack across a full hierarchy.
NUCLEUS detector records 2.16 eV resolution in lab
This preprint reports electron-volt-scale resolution and stable operation in six cryogenic detectors, while calibration and reactor-site backgrounds remain open questions.
Smaller echo-control model reports similar results with 70% fewer parameters
Preprint analysis reports about the same benchmark performance from a smaller model at 14% of a medium model's computational complexity.