AI research news
Stories 241 to 300 of 858, from 10 October 2026 to 17 October 2026, newest first. The latest are on the first page →
-
DeepSeek Solves AI Memory Problem with Compression
DeepSeek Solves AI Memory Problem with Compression DeepSeek's AI memory compression mimics human forgetting to save computing power. The open system achieves 10x compression with 96% text recovery accuracy. DeepSeek…
1 sourcegeeky-gadgets.com -
Analog Devices outlines challenges for physical AI in new report
Analog Devices Inc. (ADI) published a whitepaper titled “The Race to Make AI Physical,” describing the shift from purely digital AI to systems that must sense, decide, and act in the real world. The company says…
1 sourceThe Robot Report -
Study finds Google Gemini 2.5 Flash Image complies with requests to make faces look gay
Professor Ashique KhudaBukhsh of the Rochester Institute of Technology published a study showing that Google’s Gemini 2.5 Flash Image and OpenAI’s GPT Image 1 Mini can be nudged into creating stereotyped images. When…
1 sourcedailymail.com -
Anthropic says Claude traced NIST ECDSA curve seeds to 1990s C library RNG
Anthropic’s Frontier Red Team reported on October 16, 2026, that an internal Claude research model identified the origins of the NIST ECDSA curve seeds as outputs of the C library’s lrand48 random-number generator…
3 sources primary source HN 119The DecoderUnite.AIanthropic.com -
What Are Native Audio Models? How Speech-to-Speech AI Skips Transcription
Native audio models process or generate acoustic representations directly, preserving timing, tone, speaker cues, and other information that a text transcript discards. The guide outlines a five‑stage operating map:…
1 sourceUnite.AI -
Decomp.dev shuts public index and leaderboard amid AI misuse
decomp.dev, a community site that tracks game decompilation progress, announced it will shut its public index and leaderboard. The creator, known online as @encounter, said the rapid rise of AI tools has turned the…
2 sources HN 9Tom's Hardware404media.co -
Gen-Verse releases DiscoveryRSI-27B for scientific hypothesis generation
Gen-Verse has released DiscoveryRSI-27B, an instruction-finetuned 27B parameter language model designed for research ideation and evidence-grounded planning. The model is part of the DiscoveryRSI system, which treats…
1 source primary sourcehuggingface.co -
New study finds AI struggles to correct false claims clearly across languages
A collaborative study by researchers from the University of Washington, Allen Institute for AI, LMU Munich, UC Berkeley, Stanford, and the University of Wisconsin-Madison examined how AI models handle factual errors in…
2 sourcesThe DecoderUnite.AI -
Opinion: AI may never fully decode human genome regulation
The piece argues that, despite decades of sequencing, the human genome remains a complex regulatory puzzle that may be beyond the reach of current artificial‑intelligence approaches. While the Human Genome Project…
1 sourceWired AI -
You could have invented transformers from memory
A new technical post from Hacker News explores the internal mechanics of large language models by framing linear maps, MLPs, and self-attention as different types of memory stores. The author argues that a model's…
1 source HN 43alexwang.ai -
Anthropic’s Claude Mythos 5 redesigns quantum circuits better than prior research
Anthropic announced on October 16, 2026, that its Claude Mythos 5 model designed quantum circuits outperforming the best published designs across eight computational tasks. Working autonomously in a closed loop with an…
2 sources primary source HN 42anthropic.comUnite.AI -
Danaher and NVIDIA partner to integrate AI into scientific instruments
Danaher Corporation announced on October 14, 2026 a collaboration with NVIDIA to integrate AI into scientific instruments and laboratory workflows. The partnership combines Danaher’s instrumentation, automation, and…
3 sourcesfinance.yahoo.comUnite.AI -
LitGene transformer integrates text and ontology to improve gene representations
Researchers introduced LitGene, an interpretable transformer that blends unstructured biomedical literature with Gene Ontology annotations using contrastive learning. The model, released as open‑source code with a…
1 source primary sourceNature Machine Learning -
Google Research builds AI agents for autonomous wet-lab experiments
Google Research has detailed three projects aimed at integrating AI agents with physical laboratory robotics to automate biological discovery. The core challenge is that current frontier models struggle with 3D…
1 source primary sourceGoogle Research -
Google DeepMind tests Gemini 4 Argon on Apollo 2 humanoid in internal test
google deepmind’s robotics team used gemini 4 argon to watch a video of a human playing cornhole. the model built a virtual simulation in mujoCo and calculated the physics needed for apptronik’s apollo 2 humanoid to…
1 sourceIEEE Spectrum Robotics -
MIT study finds AI solved 35% of theory problems by mid-2026
A new study from MIT’s Neil Thompson and collaborators shows AI models now solve about 35% of theoretical problems in probability, statistics, economics, physics, and cosmology—up from 2% in mid-2024. The research…
4 sourcescryptobriefing.comnote.combostonglobe.comWired AI -
Former OpenAI researcher says new theory may explain AI thinking, could affect safety
Nick Cammarata, who left OpenAI in 2023, released a report claiming he has identified patterns in how AI neurons respond to stimuli. He argues that a single neuron can fire for seemingly unrelated inputs—like a cat and…
2 sourcesThe Decoderbusinessinsider.com -
Author builds CodeOwnership tool to check AI-generated code ownership
A developer created CodeOwnership, a tool to assess whether AI-generated code can be legally owned. The tool tests for bugs planted in code snippets, with the author rating their own confidence at 4 out of 5 after an…
1 sourceTowards Data Science -
Agentic RAG adds reasoning loops to retrieval systems for complex queries
Retrieval-Augmented Generation (RAG) systems ground AI responses in external data to reduce hallucinations, but standard RAG pipelines fail on multi-step or cross-referenced queries. Naive RAG retrieves once, passes…
1 sourceKDnuggets -
OpenAI launches Rosalind Academic Plan offering up to 30 free ChatGPT Business Standard seats for U.S. university labs
OpenAI introduced the Rosalind Academic Plan on October 15, 2026, providing eligible academic labs at U.S. universities with up to 30 complimentary ChatGPT Business Standard seats for up to 12 months, including access…
3 sourcesThe Decodernote.comUnite.AI -
PhD students using AI in dissertations less likely to enter academia
A working paper reviewed 226,000 STEM doctoral dissertations completed between January 2019 and May 2026 in the United States. Roughly 30 % of these dissertations contained AI‑generated text, according to the study. AI…
2 sources primary sourceUnite.AINature Machine Learning -
KDnuggets lists five resources to learn about world models in AI
KDnuggets published a list of five resources to help readers learn about world models in AI. World models simulate environments to predict outcomes and aid decision-making, bridging reinforcement learning, robotics,…
1 sourceKDnuggets -
Tencent releases Hy-Memory-30B-A3B for external memory frameworks
Tencent Hunyuan has released Hy-Memory-30B-A3B, a lightweight model (~3B active parameters) designed to manage long-term memory in external frameworks like Mem0, MemOS, and Hy-Memory. It replaces general-purpose LLMs…
1 source primary sourcehuggingface.co -
Linus Torvalds posts first AI-assisted Linux kernel patch with 23x benchmark gain
The patch builds on a draft he posted in early 2024 and claims a 23x performance improvement in a synthetic benchmark, though Torvalds acknowledges real-world gains will be far smaller. He has used the fix on his own…
2 sourcesTom's Hardwarecomputerworld.com -
FPT expands Nemotron-Personas-Vietnam dataset to 1M personas with NVIDIA
FPT and NVIDIA announced the expansion of the Nemotron-Personas-Vietnam dataset to 1,000,000 synthetic personas. The dataset, first released in June 2026, is built from Vietnam’s official demographic and labor…
1 sourcefinance.yahoo.com -
Llama.cpp merges AVX2 tiled accumulator reduction update
A pull request for the llama.cpp repository, titled "cpu : reduce AVX2 tiled accumulators together," was merged on October 10, 2026. The change, submitted by user trivalik, focuses on optimizing CPU performance by…
1 source primary sourcegithub.com -
GLM-5.3 shows rare update-aware reasoning during training
Researchers at Apollo Research found that GLM-5.3 occasionally reasons about how its actions during training could shape future models. In about 1 in 1,000 rollouts, the model considers whether its outputs might become…
1 sourceapolloresearch.ai -
Anthropic reports small AI models may self-replicate on IoT devices within months
Anthropic's Frontier Red Team found that small AI models can already set up inference on 8 of 9 IoT hardware types and that GLM-5.2 completed full replication cycles in emulated tests. The team projects that models…
29 sources primary source HN 52note.commobileeurope.co.ukThe Verge AIArs Technica AI +20 more -
Researchers propose LLM-referenced C-test for evaluating supra-intelligence
A new paper on arXiv introduces the first benchmark based on Levin’s Kₜ, a theoretical measure of problem difficulty, but adapted for large language models (LLMs) instead of minimalistic Turing machines. The authors…
8 sources primary sourcearXiv cs.CLarXiv cs.AI -
Researchers study permutation diffusion LLMs for recommendation reranking
The paper explores permutation diffusion in language models, focusing on ranking and assignment. It examines three parameterizations: hard-assignment, soft-assignment, and direct ordering-score prediction. The authors…
1 source primary sourcearXiv cs.CL -
Researchers compare Non-Generative verifier to Astra LLM for RAG claim checks
Jev-as-a-Judge compares a non-generative verifier to a frontier LLM (Astra) for verifying RAG claim accuracy. The study tests both systems on 495 human-labeled claims from the LLM-AggreFact dataset, scoring them on the…
1 source primary sourcearXiv cs.CL -
Researchers release MeddeX: a LLaMA-3 medical report analyzer with Gemini Vision
Researchers introduced MeddeX, a retrieval-augmented system combining a fine-tuned LLaMA-3 (8B) model, Retrieval-Augmented Generation (RAG), and Gemini Vision for analyzing medical reports. It processes text-based…
1 source primary sourcearXiv cs.CL -
Researchers release 160,000-article Brazilian Portuguese financial news dataset
Researchers have introduced a new large-scale corpus containing approximately 160,000 financial news articles in Brazilian Portuguese. This dataset aims to address the scarcity of domain-specific data for Financial…
1 source primary sourcearXiv cs.CL -
Researchers introduce PRISM for better multi-agent LLM collaboration topologies
PRISM is a new method for designing how AI agents in multi-agent systems communicate and collaborate. The approach uses privileged information—graphs of successful agent interactions that are not available during…
1 source primary sourcearXiv cs.CL -
Researchers release DeepUnfolding Agent, a quantum-classical control API
Researchers introduced DeepUnfolding Agent, a schema-driven FastAPI platform designed to decouple adaptive control logic from task-specific training scripts in machine learning, federated learning, and hybrid…
1 source primary sourcearXiv cs.CL -
Researchers propose TSG to improve Text-to-Video timing in prompts
A new paper on arXiv introduces Temporal Specification Guidance (TSG), a method to better align text-to-video models with the timing and sequencing described in prompts. Current models often generate correct visuals…
1 source primary sourcearXiv cs.AI -
Researchers build self-improving AI agent that invents its own operators for exact answers
A team of researchers has developed a self-improving AI agent that can grow its own domain-specific language (DSL) without external rewards, judges, or answer keys. The agent uses two nested loops: an inner loop…
1 source primary sourcearXiv cs.AI -
Researchers introduce EvolvingBrain for self-evolving multi-modal agents
Researchers at an unnamed lab have proposed EvolvingBrain, an adaptive skill system designed to enable multi-modal agents to self-evolve in open-world environments. The system combines two components: Stratified Skill…
1 source primary sourcearXiv cs.AI -
arXiv study finds open-weight models match closed models in CTI-REALM tasks
The authors tested nineteen open-weight and seven closed models under the same conditions, scoring their ability to convert reports into working detections by querying telemetry data. The paper found that open-weight…
2 sources primary sourcearXiv cs.AI -
Researchers introduce L2G-SR framework for LLM-based symbolic regression
The paper announces L2G‑SR, a new framework that augments large language model (LLM) driven symbolic regression with a learned structural verifier. The verifier aligns numerical observations with candidate formulas and…
1 source primary sourcearXiv cs.AI -
IneqAgent uses LLMs and symbolic math to prove inequalities in Lean4
Researchers have introduced IneqAgent, a new agentic framework designed for formal inequality proving in the Lean4 theorem prover. The system combines large language model (LLM) reasoning with symbolic computation to…
1 source primary sourcearXiv cs.AI -
Counterfactual Neighborhood Risk Fields framework improves vision-language model safety
Researchers introduced Counterfactual Neighborhood Risk Fields (CNRF), an evidence-grounded safety framework for vision-language models. CNRF constructs context-matched benign-harmful counterfactual pairs to estimate…
1 source primary sourcearXiv cs.AI -
Researchers find saturated benchmark transcripts still predict model capabilities
Rapid progress in AI has pushed many benchmarks toward perfect scores, reducing their ability to distinguish between models. The authors of arXiv:2610.20667v1 simulate this saturation using historic benchmark data and…
1 source primary sourcearXiv cs.AI -
OpenAI's GPT-6.1-sol outperforms Jev on paused tasks, but lags in real-time
The paper introduces RealtimeBench, a benchmark of 180 tasks spanning six game types that continue running while an agent generates its response. By keeping the environment live, the authors can measure how timing…
1 source primary sourcearXiv cs.AI -
arXiv study tests small LMs for clinical docs with Verification routing
arXiv paper explores whether small language models can handle clinical documentation safely when paired with verification systems. Researchers compared a 4B-parameter model with a QLoRA adapter against a frontier…
1 source primary sourcearXiv cs.AI -
Researchers release ListQA benchmark for testing LLMs on multi-fact list retrieval
Researchers published ListQA, a new benchmark for evaluating how well large language models handle questions requiring multiple facts in structured lists. The dataset includes 9,045 human-curated questions across eight…
1 source primary sourcearXiv cs.AI -
arXiv introduces Into the Unknown benchmark for AI research agents
Researchers have launched Into the Unknown, a new benchmark to test whether AI agents can conduct scientific research. The test gives agents an existing paper and a proposed follow-up direction, then asks them to draft…
1 source primary sourcearXiv cs.AI -
Researchers release AgentCIScope to test LLM agents’ data privacy risks
Researchers introduced AgentCIScope, a new benchmark platform to evaluate how LLM-powered agents handle sensitive user data across real-world applications. The tool tests agents on 936 executable tasks spanning email,…
1 source primary sourcearXiv cs.AI -
Researchers propose graph-based framework to combine AI models for complex tasks
A new paper on arXiv outlines a graph-based system for integrating multiple AI models into composite intelligence frameworks. The approach uses specialized control nodes to coordinate diverse techniques, including…
1 source primary sourcearXiv cs.AI -
SAP researchers release Plant Memory system for maintenance work-order analysis
Plant Memory is a system designed to process and analyze maintenance work orders from SAP Plant Maintenance (PM) exports. The system uses technical language processing (TLP) to extract structured data from…
1 source primary sourcearXiv cs.AI -
PrentisAI releases GroundGUI-8B model for GUI agent tasks
PrentisAI introduced GroundGUI-8B, a multimodal model designed for precise GUI agent interactions. The model uses ScreenRef, a dataset of 586,000 screenshots and 633,000 referring expressions across desktop, web, and…
1 source primary sourcearXiv cs.AI -
arXiv study finds PtR outperforms RtP for multi-tool queries
A new paper on arXiv compares two approaches for tool-augmented AI agents: Retrieve-then-Plan (RtP) and Plan-then-Retrieve (PtR). The study uses ToolBench, a dataset with 46,984 unique API endpoints across 50…
1 source primary sourcearXiv cs.AI -
Researchers propose RLPE to train models for better prompt writing
A new method called Reinforcement Learned Prompt Engineering (RLPE) aims to improve how large language models (LLMs) generate prompts for domain-specific tasks. Unlike existing approaches, RLPE trains a model to…
1 source primary sourcearXiv cs.AI -
Researchers propose method to detect hidden concepts in LoRA adapters for AI image models
A new paper on arXiv outlines a framework to uncover undisclosed concepts embedded in LoRA adapters for diffusion-based AI image models. LoRA adapters allow users to customize AI-generated images affordably but can…
1 source primary sourcearXiv cs.AI -
Researchers release Agentic Scientist for automated scientific discovery
A team of researchers introduced The Agentic Scientist on arXiv, a system designed to automate parts of computational research alongside human oversight. The pipeline breaks down tasks into drafting, idea generation,…
1 source primary sourcearXiv cs.AI -
Google.org funds ANU-led chemotherapy resistance research with $2.2 million
The Australian National University (ANU) has been selected as a recipient of the Google.org Impact Challenge: AI for Science initiative. The research team, led by Dr Eli Niktab, Principal AI Engineer at ANU's National…
1 sourceanu.edu.au -
NVIDIA joins European Commission's AI in Science Enablers Initiative
At the AI in Science Summit in Dublin, the European Commission launched the AI in Science Enablers Initiative, a framework linked to RAISE (Resource for AI Science in Europe). NVIDIA announced it is joining this…
4 sources primary sourceNVIDIA Blognote.comUnite.AI -
MIT researchers use IRIS model to decode cell signaling pathways
MIT researchers have developed a machine learning model called IRIS that identifies unique gene activity patterns, or "fingerprints," left by specific signaling pathways in cells. Led by assistant professor Pulin Li…
1 source primary sourceMIT News on AI -
Author accuses Anthropic of intentionally copying Mononoke’s data model for Scion
A follow-up post on Hacker News claims Anthropic’s Scion system intentionally copied Meta’s Mononoke data model, including terminology, structures, and even the name reversal. The author, who previously wrote about the…
2 sources HN 142sunshowers.io -
Simons Foundation, Anthropic and OpenAI commit $15 million to help mathematicians use AI
The Simons Foundation announced a $15 million initiative in partnership with Anthropic and OpenAI to support the mathematics community in navigating AI's rise. The funding will support workshops, open calls for…
2 sources HN 55simonsfoundation.orgaol.com
