A new study from arXiv identifies a structural trilemma in large language models responding to emotionally vulnerable users, where models either protectively restrict, uninflectedly facilitate, or uni…
A new reference-free framework for auditing LLM-generated reasoning traces, using NLI-based hypergraph analysis and backward AND-OR search, outperforms direct LLM-as-judge baselines in both deductive …
Researchers introduce Surrogate Latent Policy Optimization (SLPO), a method that brings outcome-reward reinforcement learning to autoregressive latent reasoners, enabling test-time scaling in latent r…
A new arXiv paper introduces Rubric4Setwise, a training-free method that improves document set selection for large language models by using rubric-based evaluation criteria, achieving state-of-the-art…
Researchers have introduced VizRAG, the first retrieval-augmented generation (RAG) system to incorporate visual hypergraph structure awareness, outperforming strong baselines in experiments. The syste…
A regression-based approach to Arabic dialect geolocation that models dialectal variation as a continuous geographic space achieves a median localization error of 481.2 km under a leakage-free 5-fold …
A controlled comparison of language-specific versus cross-lingual knowledge graphs for implicit aspect identification in Arabic ABSA finds that a native Arabic KG outperforms a cross-lingual English K…
Researchers propose Efficient Chain-of-Modality Reasoning (ECoM Reasoning), the first framework to introduce compressed reasoning into spoken language models (SLMs), improving reasoning accuracy by 21…
A new 2.07-megabyte single-file German drama corpus, tiny_schiller, provides a drop-in counterpart to Karpathy's tiny_shakespeare for small language model prototyping, fine-tuning, and education. The …
Researchers propose Thinking Checklist Reward (TCR), a process-oriented reward for reinforcement-learning-based LLM preference alignment that evaluates reasoning traces against sample-specific checkli…
A new benchmark and training set called SelectBench, introduced by researchers in arXiv:2607.20090v1, shows that reinforcement learning via DAPO can modestly improve selective evidence adoption in ret…
A new benchmark, OpenSkillRisk, reveals that LLM-based agents fail to safely handle risky third-party skills, with even the safest configurations executing unsafe actions in about 17% of cases. The be…
Researchers introduce PyroDash, a cost-aware framework for token-level small-large language model collaborative inference that reduces large language model (LLM) usage while preserving reasoning perfo…
A new framework called HyGRL, detailed in a preprint on arXiv, outperforms state-of-the-art baselines in answer accuracy and reasoning fidelity for multi-entity compositional questions while maintaini…
Apple's M5 generation introduces a dedicated Neural Accelerator on every GPU core, and BaseRT, a native Metal inference runtime, exploits these units to deliver up to 6.4× higher prompt-processing thr…
Researchers propose Twin Agent, a privilege separation design pattern for LLM agents that uses two nearly symmetric agents—an Explore Agent inspecting untrusted information and a Safe Agent executing …
Researchers introduced PoTRE (Poly-Topological Reasoning Ensembles), a heterogeneous framework that decouples inference into four agents to improve complex reasoning in large language models. PoTRE ac…
A study by researchers on arXiv (2607.20301v1) finds that PortLLM, a training-free and data-free scheme for adapting large language models after continual pretraining, maintains long-term temporal por…
Amazon Web Services Inc. has extended its Security Hub to Microsoft Azure and added artificial intelligence-specific protections, positioning its security stack as the foundation for securing enterpri…
Shares of Alphabet and Tesla fell after the companies reported massive spending increases in their quarterly earnings reports, testing Wall Street's patience as AI investments overshadow near-term gro…