Kernel Forge, an open-source agentic harness for LLM-based generation and optimization of CUDA kernels, achieves speedups of up to 2.83× on softmax in Gemma 4 E2B and 1.70× on group_norm in Stable Diffusion 3.5 Medium, o…
A new study introduces CaRE, a compute-aware evaluation framework for masked diffusion language models (MDLMs), revealing that current evaluation standards conflate algorithmic improvements with hidden choices of compute…
A new study applying the Petri automated auditing framework to Qwen3-30B-A3B finds that LLM scheming behaviors inversely scale with pretraining language coverage, with low-resource languages averaging 34.2% higher scores…
LinkedIn researchers present a unified semantic modeling framework powered by a small language model (SLM) to improve job understanding at scale, achieving significant performance gains in offline evaluations and online …
Researchers have developed ProcAgent, a fully on-device, agentic, vision-based procedural assistant for real-time adaptive guidance on a single NVIDIA Jetson AGX Orin. The system uses a propose-and-verify architecture co…
Researchers have developed NRFormer+, a spatio-temporal Transformer that achieves state-of-the-art accuracy in nationwide nuclear radiation forecasting, reducing sudden-change mean absolute error by up to 19.1% over the …
Researchers introduce Generative Topology Optimization (GenTO), a unified framework that trains a diffusion model on a large full-order topology dataset and iteratively steers the resulting topology distribution toward t…
A new hierarchical reinforcement learning framework called HOBA (Hierarchical On-Policy Bidding Agents) achieves a +3.6% increase in target cost in large-scale online advertising deployment, according to a paper on arXiv…
A new study testing GPT-5.4, Claude Sonnet 4.6, Gemini 2.5 Flash, and Qwen3-235B on 101 World Values Survey questions across 13 language-country slices found that prompt framing is a first-order determinant of cultural a…
Researchers propose SpecPrefetch, a parameter-efficient prefetching framework for offloaded sparse Mixture-of-Experts (MoE) inference that uses a shared lightweight adapter to predict next-layer expert candidates for asy…
A study evaluating six large language models within the digital mental health intervention Wysa found that retrieval-augmented generation (RAG) improved intent detection accuracy, recall, and precision compared to RAG-di…
A forensic reproducibility audit of a chest-radiograph vision-language model benchmark found that 297 of 300 planned model-prompt calls yielded nonempty reports, but 60 Claude calls labeled A/B were executed with the sam…
A new benchmark called HANDBOOK.md shows that leading language-model agents fail to follow long, binding policy documents in enterprise settings, with the best model passing only 36.2% of trials under strict grading. The…
A new study from arXiv introduces a controlled benchmark revealing that the site of adaptation in transformers—early, middle, or late layers—produces distinct learning signatures across five objectives: lexical binding, …
A new open-source benchmarking framework, Inspect India Evals, reveals that large language models (LLMs) tested on Indian linguistic and cultural contexts show significant performance gaps, with Sarvam-M 24B and Gemma 2 …
Researchers released a human-in-the-loop corpus for LLM-based simplification of scientific summaries, finding that GPT-4o-mini-generated summaries were preferred for comprehensibility and simplicity but that expert edits…
A new theoretical framework from a paper on arXiv (2607.25507v1) proposes a bounded spectral analysis of rotary phase alignment in Transformer language models, proving that uniformly bounded phase displacement limits deg…
Researchers present VLD-RAG, an agentic multimodal retrieval-augmented generation framework for question answering over visually-rich long documents, achieving improved evidence-page retrieval and end-task performance on…
A new study from arXiv preprint 2607.24797v1 finds that decoder-only large language models (LLMs) such as GPT-2, OPT, and Pythia entangle reading and writing into a single code, unlike the human brain's double dissociati…
Cardiologent, a multi-agent clinical decision support system, outperforms cardiologists and large language models in patient-level arrhythmia assessment, urgency, and management by integrating ECG and photoplethysmogram …