AI research news
Stories 361 to 420 of 858, from 4 October 2026 to 13 October 2026, newest first. The latest are on the first page →
-
Parsimony, an open-source benchmark, measures code added by AI models
A new open-source tool called Parsimony has been introduced to address a gap in current AI coding benchmarks. While existing metrics focus on solve rates and costs, they often ignore the volume of code generated, which…
1 source HN 42parsimony.lasu.dev -
Meta, Google and NIH invest $1.8B in AI cell simulation project
Meta, Google DeepMind, and the U.S. government are backing Biohub’s $1.8B Virtual Biology Initiative, expanding its original $500M funding. The goal is to train AI models on trillions of cell datasets to simulate how…
1 sourceThe Rundown AI -
OpenAI releases 722 math papers from unreleased model
OpenAI published 722 math papers from an unreleased internal model, grouped into 372 result families. The boldest claim is a proof of the quasi-Riemann hypothesis, a weaker version of the $1 million Millennium Prize…
1 sourceThe Rundown AI -
Harvard physicist uses Claude to publish 36 manuscripts in 18 fields in three months
Harvard physics professor Matthew Schwartz used Claude to produce 36 research manuscripts across 18 fields—including ecology, genetics, and economics—in just three months. He selected projects from about 400…
1 sourcemaccus.fr -
Python 3.15’s profiler finds line-ending code dominates web-edit app
Simon Willison profiled a Python 3.15 web-edit app using built-in sampling tools. The app’s preserveendings feature consumed 64.8% of CPU time across 15,296 recorded stacks, with SequenceMatcher.findlongestmatch…
1 sourceSimon Willison -
DeepMind and Biohub debate AlphaFold’s limits in protein folding
Google DeepMind’s Pushmeet Kohli and Biohub’s Sal Candido discussed why AlphaFold’s breakthrough in protein structure prediction did not fully solve biology’s challenges. They argued that scaling compute and data alone…
1 sourceLatent Space -
Periodic Labs introduces synthesis superintelligence for materials discovery
Periodic Labs, founded in September 2025 by Liam Fedus and Ekin Doğuş Çubuk, is pioneering a new approach to AI-driven scientific discovery. Unlike traditional AI models trained on digital data, Periodic’s team…
1 sourceLatent Space -
OpenAI publishes 722 math papers solving 90 of top 500 problems
OpenAI released 722 mathematical manuscripts from an unreleased internal model, addressing roughly 90 of the top 500 unsolved problems in mathematics. The results include the Quasi-Riemann Hypothesis, described by some…
1 sourceLatent Space -
Researcher predicts AI engineering acceleration but not superintelligence
A top AI researcher argues that rapid improvements in infrastructure and engineering will make AI models more capable at distributed tasks like GPU management within a few years. They note that while models will become…
1 sourceInterconnects -
MIT researchers show neural networks can build own learning rules
Scientists at MIT’s McGovern Institute and K. Lisa Yang Integrative Computational Neuroscience Center published a paper in PNAS on Oct. 9, arguing that artificial neural networks should be judged by their ability to…
1 source primary sourceMIT News on AI -
Sakana AI’s MASS method improves model self-judgment but risks baking errors into training
Sakana AI and UC Berkeley researchers introduced Multi-Agent Self-Supervision (MASS), a method where a single language model improves itself by solving tasks, evaluating results, and refining workflows—without external…
1 sourceThe Decoder -
Microsoft Research finds clearing chat history boosts AI agent performance by 66%
Microsoft Research published a study showing that long-running AI agents often plateau in performance despite more compute. The issue stems from accumulated conversation context, which restricts agents to similar…
1 sourceThe Decoder -
Elorian and Scale AI release benchmark showing AI lags in visual reasoning
Elorian and Scale AI introduced Humanity’s Sixth Sense, a benchmark testing how AI models interpret implied details in visual scenes. The test covers 522 questions across 288 images and 234 short video clips (17.6…
1 sourceImport AI -
AI labs buy data from dead startups for training models
AI labs are purchasing data from defunct startups to train models, as public and synthetic datasets prove insufficient. Companies like SimpleClosure and Sunset broker these deals, selling codebases, Slack archives,…
1 sourceMIT Technology Review AI -
Hugging Face’s ML-intern builds custom models from prompts in days
A Hugging Face user created seven custom AI models in under a week using ML-intern, an agent that automates training workflows. Each project started with a detailed prompt—including verified facts, baseline benchmarks,…
1 source primary sourceHugging Face -
Sakana AI's Multi-Layered Review catches 73.43% of core-claim errors
Sakana AI released a research paper, Beyond Imitation, describing an AI‑assisted peer‑review system called Multi‑Layered Review (MLR). The system uses three off‑the‑shelf Claude agents – Haiku 3.5 for appendix…
1 sourceMarkTechPost -
Is this idea from AI or humans? A detector claims it can tell
A new detector called idealens was presented by researchers to distinguish ideas generated by humans from those produced by ai. The system analyzes the outline of a document rather than its full text, stripping away…
1 source primary sourceNature Machine Learning -
Nature urges preservation of research data for AI integrity checks
Nature highlights that artificial‑intelligence tools can spot long‑standing errors in scientific papers, showing retrospective scrutiny is possible. The correspondence notes that such checks would be more robust if the…
1 source primary sourceNature Machine Learning -
Cilia tracks model classifies primary ciliary dyskinesia with 95-97% accuracy
The study shows that machine learning can detect primary ciliary dyskinesia (PCD) by analyzing cilia‑driven particle transport. Researchers captured fluorescent microsphere movement on cultured multiciliated cells from…
1 source primary sourceNature Machine Learning -
Researchers adopt offline LLMs to search literature, citing privacy and cost
Researchers are turning to locally‑run large language models to query their own literature collections after concerns that cloud‑based chatbots may retain input data. Ian Gaudet, a neuroscientist at Florida Atlantic…
1 source primary sourceNature Machine Learning -
Researchers release Tangermeme toolkit for interpreting genomic deep learning models
A team of researchers has released Tangermeme, an open-source Python toolkit designed to help scientists interpret what genomic deep learning models have learned about cis-regulatory logic. Published in Nature Machine…
1 source primary sourceNature Machine Learning -
Researchers fear AI tools may scoop unpublished work, limit usage
Scientists are growing concerned that commercial AI systems could solve research problems before they publish, effectively scooping them. In two recent cases, researchers found that AI companies announced results on…
1 source primary sourceNature Machine Learning -
Clinical microbiology needs AI to predict pathogen evolution
Clinical microbiology researchers argue that an artificial intelligence system capable of forecasting pathogen evolution and resistance is essential for proactive prevention. The article outlines how laboratories…
1 source primary sourceNature Machine Learning -
OpenAI posts 700 AI-generated math preprints, sparking backlash
OpenAI released over 700 mathematical preprints generated by AI on October 6, all published simultaneously on GitHub. The move has divided the research community: some mathematicians praised potential breakthroughs,…
1 source primary sourceNature Machine Learning -
Researchers retrofit LLMs to count letters using byteification
Minixhofer et al. describe a method called byteification that modifies existing token‑based large language models so they can operate on raw bytes instead of word‑level tokens. By exposing the underlying byte…
1 source primary sourceNature Machine Learning -
Opinion: MIT professor outlines path for AI in academia
Sasha Rakhlin, director of the MIT Statistics and Data Science Center, argues that universities must urgently rethink how they evaluate research and train graduate students as AI capabilities expand. In an essay…
1 source primary sourceMIT News on AI -
MIT professor uses AI to boost data center efficiency and cut energy waste
Christina Delimitrou, an associate professor at MIT, is applying machine‑learning techniques to make large‑scale data centers run closer to full capacity. She notes that many facilities operate at only about 15 percent…
1 source primary sourceMIT News on AI -
Google study shows AI improves patent drafting scores, senior lawyers gain judgment
A three‑month randomized trial involving 133 patent attorneys at eleven IP firms tested a then‑unreleased Google Labs AI patent‑writing assistant, now part of Gemini Notebook. Participants were split into treatment…
1 source primary sourceGoogle Research -
Lincoln laboratory publishes survey of AI accelerators
The Lincoln Laboratory Supercomputing Center (LLSC) has been running the Lincoln AI Computing Survey (LAICS) since 2018. The survey now spans six papers, covering more than 120 AI accelerators, from CPUs and GPUs to…
1 source primary sourceMIT News on AI -
AI models can provably hide arbitrary capabilities
Researchers have demonstrated that neural network models can conceal arbitrary computations by embedding encrypted sub‑networks within the host model's weights. The hidden circuits are activated only when a secret…
1 source primary source HN 262arxiv.org -
PyTorch tutorial shows how to cut drone data usage by 94%
A new tutorial on Towards Data Science explains how to use a small AI model to synchronize a remote machine, such as a drone, with a cloud server without constant data streaming. The method involves running two copies…
1 sourceTowards Data Science -
Article outlines five practices for building robust Python AI libraries
The article from KDnuggets details five best practices for creating production-ready Python AI libraries. It highlights common pitfalls like unhandled API errors, bloated dependencies, and unreliable model outputs. The…
1 sourceKDnuggets -
Vector databases explain how AI retrieves similar data at scale
Vector databases store, index, and search embeddings to retrieve items by similarity efficiently. The article breaks down their five-stage operating map: generating and storing vectors with metadata, building…
1 sourceUnite.AI -
PhAI Labs introduces JEPA-Anything framework for cross-domain world models
Researchers from PhAI Labs, CUHK, Fudan, Stanford, Oxford, and Princeton released JEPA-Anything, a domain-agnostic framework that applies a single learning recipe across seven fields: vision, biology, clinical…
2 sourcesThe DecoderMarkTechPost -
Researchers introduce VIMA for finding disease patterns in tissue data
A new deep learning method called variational inference-based microniche analysis (VIMA) has been introduced to identify disease-associated patterns in spatial molecular data. Unlike previous approaches, VIMA does not…
2 sources primary sourceNature Machine Learning -
AI may create scientific monocultures, researchers warn
Lisa Messeri and Molly Crockett argue that the widespread adoption of artificial intelligence in research could erode scientific independence. Their concern is that AI tools might introduce a common source of error…
1 source primary sourceNature Machine Learning -
Opinion: Author explains difference between Strong AI and Weak AI
Author explains that Strong AI refers to general‑purpose intelligence like a human, while Weak AI specializes in specific tasks. The piece clarifies that generative models such as ChatGPT, which can write text, create…
1 sourcenote.com -
Anthropic model refutes 3SUM and APSP hypotheses with subquadratic algorithms
Anthropic’s internal research model has produced the first deterministic algorithms that break the long‑standing quadratic and cubic barriers for two classic problems. The new method solves 3SUM on n integers in…
1 source primary sourcearxiv.org -
Llama.cpp adds MTP decoding for Qwen4Exp and GLM-5.3-Flash hybrid model
llama.cpp version 0.6.0 introduces MTP (Mixture of Token Predictors) speculative decoding for the Qwen4Exp model, claiming a ~1.5x speedup on DGX Spark hardware. The update also adds support for the GLM-5.3-Flash…
2 sources primary sourcegithub.com -
Anthropic study: task understanding beats job title for AI success
Anthropic released a study analyzing approximately 400,000 uses of Claude Code between October 2025 and April 2026. The research examined how users interacted with the AI to complete coding and document tasks,…
1 sourcenote.com -
Researchers test fixed token codes for language models at 100B-token scale
A new paper on arXiv explores whether language models need trainable input embeddings to function well. The authors trained three decoder-only models from scratch with identical tokenizers, backbones, and training…
1 source primary sourcearXiv cs.CL -
Baibaichuchu at the NTCIR-19 FinArg-3 Task: When Is Maximum Possible Profit Predictable from Investor Text?
The BAIBAICHUCHU team participated in the NTCIR-19 FinArg-3 Social Media Subtask, ranking Chinese investor posts by Maximum Possible Profit (MPP). Their ensemble model—combining lexical features, a MacBERT ranker…
1 source primary sourcearXiv cs.CL -
Researchers introduce JEVal benchmark to test General decision models
A new paper on arXiv compares general decision models like Jev with traditional LLMs. The authors created JEVal, a benchmark of 11,257 instances across 36 datasets and 10 domains, to measure performance in structured…
1 source primary sourcearXiv cs.CL -
Researchers release OncoNoteBERT for oncology note processing
A team of researchers introduced OncoNoteBERT, a specialized BERT-style encoder for natural language processing of real-world outpatient oncology notes. The model was trained on a governed UK dataset of 290,026 notes…
1 source primary sourcearXiv cs.CL -
Study finds brain-alignment and Cross-Lingual scores can mislead when probes fail
A new arXiv paper challenges the use of similarity scores as evidence that language models share structure with the human brain or across languages. The authors examine two common settings. In cross-lingual transfer, a…
1 source primary sourcearXiv cs.CL -
Study shows large language models can classify Fine-Grained emotions in app reviews
The paper evaluates how large language models (LLMs) can perform multi‑label emotion classification on mobile app reviews, a task useful for requirements engineering and feature prioritisation. Researchers compared…
1 source primary sourcearXiv cs.CL -
Researchers propose FEM-ASM to separate storage, execution, and coordination in language models
The article introduces FEM-ASM, a finite-element-method-inspired organization for language models that separates contextual computation, persistent storage, and exact execution. It uses independently constructed…
1 source primary sourcearXiv cs.AI -
Researchers propose surrogate log-probabilities to audit LLM agents
A new arXiv paper introduces a method to audit black-box LLM agents by using a low-cost, open-weight surrogate model. The approach addresses the problem that frontier chat APIs hide token probabilities, making it…
3 sources primary sourcearXiv cs.AI -
Behavioral history outperforms descriptions for LLM synthetic personas
The study examined how well synthetic personas, created by large language models, can predict individual survey responses. Researchers used a two‑wave panel of 845 U.S. adults who answered 14 behavioral bias questions.…
1 source primary sourcearXiv cs.AI -
Researchers introduce ROAR to unify AI-driven research system runs
A new paper on arXiv describes ROAR, an infrastructure that aggregates and normalizes outputs from heterogeneous AI‑driven research systems (ADRS). The platform uses a relational schema and a parsing layer to reconcile…
1 source primary sourcearXiv cs.AI -
CPW-Drive uses Confucian thought to guide autonomous driving decisions
Researchers have introduced CPW-Drive, a new framework that integrates Chinese philosophical wisdom into autonomous driving decision-making. The system uses a closed-loop retrieval-augmented generation (RAG) approach,…
1 source primary sourcearXiv cs.AI -
Tool-using multimodal models fail to refuse harmful requests, study finds
A new research paper reveals a critical safety vulnerability in agentic multimodal large language models (MLLMs) when they use tools. The study found that top open- and closed-weight MLLMs are significantly less…
1 source primary sourcearXiv cs.AI -
Is the Place to Invent AI Moving Outside of Companies?
Star researchers are leaving corporate labs to launch independent AI research companies. Ilya Sutskever founded Safe Superintelligence (SSI), Mira Murati started Thinking Machines Lab, David Silver created Ineffable…
1 sourcenote.com -
Google launches prototype satellite for Project Suncatcher to test AI chips in space
Google announced on October 1, 2026, that it launched a prototype satellite developed with Planet into orbit as part of Project Suncatcher, a long-term research project announced in 2025. The satellite, launched via…
2 sourcessfist.comnote.com -
NVIDIA shares streaming robotics pipeline for Cosmos3-DROID dataset
NVIDIA’s tutorial demonstrates how to build a streaming robotics learning pipeline for the Cosmos3-DROID dataset—707 GB in size—without downloading it locally. The approach uses metadata graphs, HTTP byte-range access…
1 sourceMarkTechPost -
AI agents identify two room-temperature magnetic semiconductor candidates
A team of AI agents and a researcher designed a new Luttinger-compensated magnet, YBaMnFeO₅, and rediscovered KV[Cr(CN)₆], a 1999 compound, as potential materials for spintronic memory. Both are predicted to have…
1 source HN 64vals.ai -
OpenAI claims to solve Navier–Stokes problem, sparking math community debate
At the Heidelberg Laureate Forum in September, mathematicians discussed how AI companies like OpenAI, Anthropic, and Google are rapidly advancing in solving complex mathematical problems. OpenAI announced it had solved…
2 sourcesThe Verge AIIEEE Spectrum AI -
Microsoft Research unveils Quine, a biological world model for life sciences
Microsoft Research has announced Quine, a new AI system designed to function as a 'biological world model.' Unlike specialized models that handle only proteins or gene data, Quine integrates diverse biological…
1 sourcenote.com -
Researchers present ProFormer for classifying proteomic data
A research team from the University of Heidelberg has introduced ProFormer, a deep learning pipeline designed to classify biological samples directly from mass spectrometry data. The model uses a transformer…
1 source primary sourceNature Machine Learning -
Simon Willison tests Qwen3.8-27B arithmetic in words
Simon Willison replicated a two-year-old experiment by Colin Frasier, which tested GPT-4o's ability to sum large numbers and return the result in words. Willison ran the test on local hardware, specifically a DGX…
1 sourceSimon Willison
