ResearchAnthropic identifies a hidden 'J-space' of internal tokens inside large language modelsAnthropic reports finding a previously unseen internal representation space—called the J-space—in its Claude models, populated by tokens that do not appear in outputs but seem to track reasoning steps and states.3 min read
ResearchClaude's expressions of value vary by language: warmth vs. strictnessClaude's linguistic behavior changes depending on the language of the conversation; the difference is most noticeable on the warmth versus strictness axis.1 min read
ResearchResearch into the causes and controllability of the language model Claude's value expressionsThe value manifestations of the language model Claude affect everyday conversations, but the developers do not understand why these values change.1 min read
ResearchAgent-as-a-Router and ACRouter: a dynamic, feedback-driven approach to model routingResearchers propose Agent-as-a-Router, an open-source paradigm that treats the model router as a learning agent using a Context-Action-Feedback loop to accumulate execution-grounded memory.6 min read
ResearchHungarian AI camera system monitors preterm infants' movements and positionsResearchers in Hungary developed a camera-based monitoring system that uses deep learning to track the spontaneous movements and body positions of preterm infants in incubators.2 min read
ResearchPositive assessment of the GPT-5.6 family of modelsThe article's author evaluated the GPT-5.6 family of models, finding that they deliver excellent performance.1 min read
ResearchGPT-5.6 Sol Ultra proved the 50-year-old Cycle Double Cover conjecture in under an hourAfter the public release of GPT-5.6 Sol Ultra, the model announced today that, with the combined work of 64 sub-agents, it proved the 50-year-old Cycle Double Cover Conjecture in less than an hour.1 min read
ResearchDeepSeek’s DSpark speeds up text generation with adaptive speculative decodingDeepSeek introduced DSpark, a speculative decoding module that uses a small draft model to propose token blocks which a frozen large model verifies in one pass, and released checkpoints for DeepSeek-V4 models under an MIT license.5 min read
ResearchBrain2Qwerty v2 improves MEG-based text decoding but error rates remain highResearchers from Meta and several French and Spanish institutions released Brain2Qwerty v2, an updated system that decodes text from magnetoencephalography (MEG) recordings.3 min read
ResearchAI uncovers prion-derived peptides with antibiotic activity, raising biosafety questionsResearchers at the University of Pennsylvania used an AI system to screen millions of short peptide fragments from prion and prion-like proteins and identified over a thousand antimicrobial candidates.2 min read
ResearchBartosz (@nasqret) mathematician solves previously unsolvable problems with help of GPT-5.6Bartosz, better known online as @nasqret, is a mathematician who uses the GPT-5.6 language model to solve mathematical problems previously considered unsolvable; this promises new methods and faster…1 min read
ResearchIterative synthetic financial headline generation using NVIDIA NeMo and NemotronResearchers developed an iterative pipeline to create a 502,536‑headline synthetic financial news corpus across 13 categories using NVIDIA NeMo Data Designer, NeMo Curator and Nemotron models.6 min read
