Category
343 papers
MGQL is presented, the first mechanized, small-step operational semantics for a substantial read-only fragment of GQL that is grounded in the ISO/IEC 39075 standard, and it is proved that the type system is sound, ensuring an end-to-end guarantee of well-formed queries yielding results that conform to their declared schemas.
The role of a frozen off-the-shelf instruct model as the teacher in on-policy distillation is investigated, and a key insight is revealed: the teacher reshapes the student's policy distribution so that subsequent RL converges to a superior solution that RL alone cannot reach.
A two-dimensional framework, specifically tailored for analyzing multilingual embedding benchmarks under dataset scarcity, is proposed and applied on the Slavic-language subset of the MTEB benchmark, revealing severe benchmark sparsity.
Ternary multiplicative adaptation is proposed, which represents discrete updates of ternary weights such as sign flips or zeroing through a low-rank Kronecker factorization into two small ternary matrices applied element-wise to ternary weights.
Reach audiences
Advertise in front of researchers, engineers, and readers.
A supervised ensembling framework that trains a classifier over heterogeneous UQ-based scorer outputs on a small, domain-specific dataset of labeled LLM responses, then applies it to out-of-sample hallucination classification without retrieval, tools, or reference documents is studied.
Parason is introduced, which reveals and learns both forms of parallelism in LLM reasoning, and identifies Trial Parallelism as the majority of parallelizable reasoning computation, and it becomes increasingly dominant on hard problems.
FPGAgent is the first task-specification-to-executable HLS generation framework experimentally validated on a well-established benchmark, and the value of end-to-end validation is demonstrated.
A multilevel conceptual pathway in wearable reflectance PPG is supported, in which mechanical conditions at the sensor-skin interface are associated with changes in PPG signal characteristics, derived features, and, in a smaller body of studies, downstream physiological estimation.
This paper argues that a small win does not show that routing did anything, the authors' or anyone else's, and argues that a small win does not show that routing did anything, theirs or anyone else's.
"List as many words as you can that start with M." The verbal fluency (VF) task is simple, yet even a typical university student only manages to produce about 15 words within 1 min, and there is substantial variability around this mean. The present study examined how linguistic and domain-general abilities contributed to VF performance in a large sample of healthy adult native speakers of Dutch (N = 571). We assessed the effects of linguistic knowledge, processing speed, short-term/working memory, and fluid intelligence on performance in the VF task. To examine whether linguistic and domain-general abilities contribute differently across VF task types, we included semantic trials (category-based generation: animals, food) and phonemic trials (letter-based generation: words beginning with S or M). We assessed the total number of correct words produced and the time to first response. Mixed-effects modeling showed that linguistic knowledge predicted the total number of correct responses in both semantic and phonemic VF. Short-term/working memory and processing speed were also significant predictors, but with smaller estimated effect sizes. Time to first response showed little effect of linguistic skills. We discuss how linguistic knowledge shapes the structure of the mental lexicon such that it affects both meaning-driven and form-driven access to lexical items. In addition, we provide updated norms for VF performance in Dutch and practical suggestions for using the task. (PsycInfo Database Record (c) 2026 APA, all rights reserved).
The results suggest that LLM data mixing should be treated not only as a prediction problem, but also as an experimental-design problem in which the proxy mixtures themselves can be chosen to improve statistical efficiency.
Agent harnesses record a failed tool call and its error message in the transcript and ask the model to continue, on the assumption that the error is corrective information, and it is found the gain is negative for every instruction-tuned model tested.
Data Management
What if pathology foundation models could do more with less? GigaPath-Flash and GigaTIME-Flash cut computational demands while maintaining strong performance, opening the door to larger studies and broader exploration. The post GigaPath-Flash and GigaTIME-Flash: Toward population-scale discovery with efficient pathology foundation models appeared first on Microsoft Research.
With millions of users across the world, Julia has been used to conduct cutting-edge research and to design new drugs, jet engines, heat pumps, and more.
Earth AI
Health & Bioscience