News

Researchers propose a unified LLMOps architecture for real-time, enterprise-ready deployments of large language models, addressing knowledge staleness, catastrophic forgetting, hallucination, and weak feedback loops. The…

A study on arXiv (2608.00004v1) found that three cheap open-weight models — GPT-OSS 120B, DeepSeek-V4 Flash, and Gemma-4 31B — agree with human pass/fail decisions on natural-language mathematical proof grading at rates …

Researchers at an undisclosed institution introduced RubricReviewer, a rubric-driven peer review framework that generates paper-adaptive rubrics as an explicit intermediate step and combines a training-free agent (Scout)…

A study from arXiv (2608.00507v1) training a ~7M-parameter Transformer encoder on child-directed and read speech found that the learning objective, not architecture, determines perceptual narrowing in self-supervised spe…

Researchers propose the Capability-Driven Multimodal Scaling Law, the first cross-family framework that predicts vision-language model (VLM) benchmark accuracy from textual capability scores, trained on over 150 VLMs acr…

Researchers trained a small language model to route queries to specialized retrieval agents, achieving an NDCG@10 of 0.918 on mismatched queries, compared to 0.539 for Amazon Nova Lite and 0.490 for Claude Haiku 4.5, and…

A new arXiv preprint (2608.00585v1) shows that per-chunk verification fails on multi-hop retrieval-augmented generation, with entailment scoring reaching only 0.643, 0.523, and 0.560 AUC on HotpotQA, 2WikiMultihopQA, and…

A new study from arXiv (2608.00205v1) finds that human faithfulness annotations for text summarization exhibit an 'Averaging Bias,' where annotators accept summaries as faithful if most sentences are supported, rather th…

Researchers propose SeDeM, a selective decompression framework that decouples compact memory storage from decoder conditioning for long-context question answering. On four long-context QA benchmarks, SeDeM achieves highe…

A new arXiv preprint (2608.00355v1) finds that apparent progress toward harder tasks in large language models is mostly a ceiling effect, but a smaller hard-task effect persists after controlling for overall ability. Usi…

A new arXiv paper (2608.00582v1) introduces BPE-guided insertion, a method to adapt byte-level BPE tokenizers to underrepresented languages without changing the model vocabulary size. On Ukrainian adaptations of Nemotron…

Researchers at Observatorio Lázaro have built a self-populating database that has recorded more than two million English lexical borrowings (anglicisms) in the Spanish digital press, processing 1.88 million articles and …

A new arXiv preprint (2608.00078v1) presents a mobile deployment pipeline that uses QLoRA fine-tuning of an LLM to generate convolutional neural networks, with INT8 TensorFlow Lite export and on-device benchmarking on a …

A new arXiv preprint (2608.01014v1) introduces Cloud-ScPO, a topology-guided preference-mining framework that derives preference supervision from the hidden-state geometry of large language models (LLMs) for semi-supervi…

Researchers introduced SPARC-Rad, a manually curated multimodal benchmark dataset and evaluation pipeline for assessing spatial and anatomical reasoning in radiology vision-language models (VLMs). The dataset includes 30…

Researchers introduced TreeProbe, the first cultural-bias benchmark for Tibetan medicine, containing 4,719 expert-adjudicated items covering 467 diseases and 10 subtasks, and found that current large language models exhi…

Researchers introduced MedUPSQA, a dataset of 21,874 mid-stream clinical decision points from 5,535 real case reports, and MedUPS, an alignment framework that uses reinforcement learning to improve next-step accuracy in …

Researchers propose RSTG (Recovering Learning Signals via Adaptive Teacher Guidance), a method that selectively applies on-policy distillation to negative zero-variance prompts in reinforcement learning with verifiable r…

Researchers introduced OoO-Spec, an out-of-order semantic speculation method that speeds up LLM tool calling by predicting function choices and argument values in parallel. Across seven targets and three benchmarks, OoO-…

Opt.Gear, a new foundation model family from an unnamed research team, achieves up to 4.9x faster prefill and decoding speeds on NPUs compared to similar-scale models, with dense models of 1M, 270M, and 1B parameters sup…