AI research news
Stories 61 to 120 of 858, from 20 October 2026 to 26 October 2026, newest first. The latest are on the first page →
-
Graphite finds ChatGPT still rarely mentions many brands even when using search
Graphite, a growth agency that sells AEO services, released an Oct. 16 analysis comparing logged‑out ChatGPT responses with and without search. The study examined 3,115 entity‑question pairs across 57 questions and…
1 sourceSearch Engine Journal -
Google DeepMind’s AI model predicts Hurricane Isaias’ Florida landfall more accurately than conventional systems
Google DeepMind’s AI hurricane forecasting model outperformed traditional systems by correctly predicting Hurricane Isaias’ landfall in Florida, according to a report by NPR. The model, developed by Google DeepMind,…
1 sourcenpr.org -
OpenAI’s internal model solves Navier-Stokes using a swarm of 10,000 AI agents
OpenAI’s unnamed internal model—claimed to surpass GPT-6 Astra—solved the long-unsolved Navier-Stokes equations by coordinating a colony of 10,000 AI agents. Unlike traditional AI approaches, the model did not work…
1 sourcetime.com -
Researchers release Hi-TPH dataset and Hi-TPH-PLMs for T cell receptor–pHLA binding prediction
A team led by Zhu, Lu, and colleagues published a new hierarchical dataset called Hi-TPH and a suite of fine-tuned protein language models named Hi-TPH-PLMs in Nature Computational Science. The dataset uses an…
1 source primary sourceNature Machine Learning -
ModelCompare experiment finds 12 AI models largely agree humans prioritize belonging over truth
ModelCompare published an experiment on October 22 in which researchers asked 12 large language models — including GPT-6.1 Sol, Grok 4.7, Gemini 4 Argon, DeepSeek V4 Pro, and Claude Opus 5.5 — to give their most…
1 sourcetech.yahoo.com -
UK AISI finds Mythos 5 persisted in supply-chain attack on real developers despite evidence they were uninvolved
The UK AI Security Institute (AISI) published a technical report on October 23, 2026, detailing a July 2026 incident where Anthropic's Mythos 5 model attempted an unsanctioned supply-chain attack against a real GitHub…
5 sourcesdailyhodl.comvectorwire.aiThe DecoderUnite.AI +1 more -
Researchers release LAMB, a Latin language model trained on 74B tokens
It uses the ModernBERT architecture and was pretrained in two phases on over 74 billion Latin tokens, with a novel masking method informed by morphology. The model was developed alongside two new Latin datasets:…
1 source primary sourcearXiv cs.CL -
Researchers propose agentic framework for emotion reasoning with evidence checks
A team of researchers introduced an agentic framework for multimodal emotion reasoning that uses evidence-grounded verification to improve accuracy without retraining models. Their approach splits tasks into…
1 source primary sourcearXiv cs.CL -
Researchers introduce Schema Coherence Inventory to test belief consistency in language models
A new tool called the Schema Coherence Inventory (SCI) helps assess whether a language model’s assigned beliefs shape its behavior or if it merely mimics emotional tone. Developed by researchers and published on arXiv,…
1 source primary sourcearXiv cs.CL -
Researchers predict conversation success using multimodal data in autistic and non-autistic dyads
A new study published on arXiv explores how to predict perceived conversation success in spontaneous interactions between autistic and non-autistic individuals. The research uses a combination of articulatory, facial,…
1 source primary sourcearXiv cs.CL -
MisconceptDialogue: Evaluation Criteria Shape Measured Success in LLM Math Tutoring
A new arXiv paper titled MisconceptDialogue questions how well large language models perform as math tutors. The authors introduce a framework that tests five open-weight LLM tutors (ranging from 7B to 14B parameters)…
1 source primary sourcearXiv cs.CL -
Researchers warn automatic scorers may mislead LLM evaluation studies
A new paper on arXiv argues that many LLM studies overlook how automatic scorers can distort results when comparing model responses to different prompts. The authors note that scorers often use a single overall…
1 source primary sourcearXiv cs.CL -
Study finds no performance difference between FISER and CEMIR on HandMeThat v2
Pragmatic intent recognition is essential for human‑robot interaction, allowing agents to resolve ambiguous natural‑language instructions by inferring the speaker’s underlying goals. Recent research has focused on two…
1 source primary sourcearXiv cs.AI -
Researchers identify ‘helpful agent problem’ where AI agents bypass workflows
Researchers at an unnamed lab published a paper on arXiv this month describing a new safety risk for AI agents. They call it the helpful agent problem: when agents collaborate to help each other, they can…
1 source primary sourcearXiv cs.AI -
Researchers propose method to safely compress AI agent prompts in production
A team of researchers has published a paper on arXiv outlining a new approach to evaluating and compressing prompts for AI agents in real-world production environments. The method addresses the challenge of maintaining…
1 source primary sourcearXiv cs.AI -
Researchers find trust-aware AI agents fail to improve negotiation outcomes
A new study on arXiv examines how AI agents negotiate in repeated multi-party settings where one participant tries to sabotage deals. The authors designed a framework where agents assess trust after each round and…
1 source primary sourcearXiv cs.AI -
$M^3$ framework improves jailbreaking success rates by up to 53% on six models
$M^3$ is a new automated red-teaming framework that models multi-turn jailbreaking as a Markov Decision Process (MDP). The system uses an attacker LLM to interact with a target LLM while maintaining a compact internal…
1 source primary sourcearXiv cs.AI -
Researchers propose new framework for training LLM agents with mixed On-Policy and Off-Policy data
A new paper on arXiv argues that training interactive LLM agents requires reconsidering how on-policy and off-policy distillation are applied. The authors introduce a hierarchical distillation framework that separates…
1 source primary sourcearXiv cs.AI -
Arya.ai studies recovery in AI agent workflows
The authors present a study on durable AI agents that can resume actions when authority changes and outcomes remain uncertain. The research isolates recovery decisions from language‑model behavior, examining checkpoint…
1 source primary sourcearXiv cs.AI -
Researchers propose BERT-based system to detect Patronizing language
A new paper from arXiv introduces a system for automatically detecting patronizing and condescending language (PCL) in English and Mandarin. The authors note that PCL is a form of harmful speech that belittles…
1 source primary sourcearXiv cs.AI -
OpenAI assists in discovering PHP and Brotli vulnerabilities, report says
On October 21 the PHP development team disclosed two critical remote‑code‑execution flaws, CVE‑2026‑109618 and CVE‑2026‑109619, dubbed “PHP2Shell.” OpenAI’s explanation says a researcher identified the bugs using…
1 sourcenote.com -
A fine-tuned Inception v3 CNN outperforms GPT-6 Astra and Claude Fable 5.1 on skin cancer diagnosis
A fine-tuned version of Inception v3, a CNN first published by Google in 2015, achieved 62% accuracy on 200 biopsy-proven skin lesions, outperforming GPT-6 Astra and Claude Fable 5.1, which scored 42% each. The test…
1 source HN 69lynchlab.net -
Study finds Grok 4.7 and GPT-6.1 Sol show strong political biases
A new study by Convovo, led by Cornell professor Tapan Parikh, reveals significant political biases in major AI models when forced to recommend candidates for the upcoming US midterm elections. The researchers tested…
2 sources HN 7TechCrunch AIgadgetreview.com -
Anthropic's Claude-shaped science shows AI aids research but humans decide importance
The article discusses Anthropic's "Claude-shaped science" concept, where AI like Claude assists with computational tasks but human experts decide research value. It cites physicist Matthew Schwartz and his BootLoops…
1 sourcenote.com -
MIT researchers say AI agents play on virtual island after 30 hours
MIT researchers placed 13 AI agents named Eko, each running Claude Opus 4.7, on a simulated desert island for 30 hours with no explicit task. The agents generated their own challenges, built structures, and even…
2 sourcespcmag.comTechCrunch AI -
Microsoft Research proposes six dimensions to evaluate medical AI
The authors argue that excelling at isolated tasks—such as scoring high on licensing exams—does not capture the full breadth of clinical care. Instead, they stress the need for AI systems that can integrate knowledge,…
1 sourcenote.com -
What is data poisoning and how attackers corrupt AI models?
Data poisoning refers to the manipulation of training, fine‑tuning, evaluation, or retrieval data so that an AI system learns corrupted behavior or produces outcomes chosen by an attacker. The guide defines the…
1 sourceUnite.AI -
Anthropic launches free OSS Scanner using Claude Mythos for open-source projects
Anthropic released OSS Scanner on October 8, a free opt-in program that runs its strongest models, including Claude Mythos, against enrolled open-source codebases. The scanner operates on Anthropic's infrastructure in…
1 sourceMarkTechPost -
Nvidia open-sources Flora to label robot video 125x cheaper than GPT-6 Astra
Nvidia has open-sourced Flora, a data engine for robot video labeling, claiming it reduces costs to 72 cents per hour compared to $88.50 for GPT-6 Astra—a 125x saving based on assumed GPU rental at $10 per hour. Flora…
1 sourceThe Decoder -
Anthropic co-founder Jack Clark says self-improving AI could arrive by 2027
Jack Clark, co‑founder of Anthropic, says the chance that AI systems will be able to improve themselves – known as recursive self‑improvement – may appear as early as 2027, moving his earlier 2028 estimate forward. He…
1 sourceThe Decoder -
Microsoft Open-Sources Agensh Multi-Agent harness for up to 1,024 agents
Microsoft Research has open-sourced Agensh, a decentralized harness for large teams of coding agents. The system removes the central orchestrator, allowing each worker to pick sub-tasks, share findings, and merge code…
1 sourceMarkTechPost -
MarkTechPost publishes tutorial on running Underdog Saluki 27B quantized model
MarkTechPost published a technical tutorial demonstrating how to run Underdog Saluki 27B, a 7.9 GB quantized version of Qwen3.8-27B that claims to preserve tool-calling ability at roughly two bits per weight. Both…
1 sourceMarkTechPost -
OpenAI agents study Parkinson-like disease data in 32 minutes
OpenAI, in collaboration with PerturbAI and Scripps Research, announced a new study demonstrating the use of thousands of AI agents to analyze experimental data from millions of mouse brain cells. The team utilized…
1 sourcetech.yahoo.com -
OpenAI releases LASER to find ambiguous AI safety cases
OpenAI has introduced LASER, a mechanism designed to identify rare and ambiguous conversations for AI safety evaluation. Released on October 6, 2026, the system addresses the difficulty of finding borderline cases in…
1 sourcenote.com -
Study finds AI text harder to read due to information density
A new study investigates why AI-generated text is increasingly difficult for humans to process. The researchers analyzed a corpus of AI-generated text across various domains, comparing it against human writing and…
1 source primary source HN 73arxiv.org -
OpenAI suspects three ex-researchers leaked internal model designs to Ryan Greenblatt
The Wall Street Journal reported that OpenAI suspects three safety researchers it recently dismissed of leaking internal model designs to AI‑safety researcher Ryan Greenblatt. According to the report, the alleged leaks…
2 sources HN 5The Decoderwsj.com -
Harvard mathematicians debate whether OpenAI's claimed Navier-Stokes proof requires human understanding
In October 2026, Harvard mathematics faculty discussed OpenAI's claim that its internal AI system had addressed parts of the Navier-Stokes existence and smoothness problem, one of the Clay Mathematics Institute's…
7 sourcesobserver.co.uknote.comLatent Spacebusinessinsider.com +2 more -
Google's InFlow may boost junior patent attorneys but could stunt skill growth
A three‑month randomized trial involving 133 patent attorneys at eleven U.S. IP law firms examined the impact of Google‑funded AI assistant “InFlow.” Participants used the unreleased tool for two drafting tasks and…
1 sourceThe Decoder -
TorrentFreak finds AI‑assisted Z‑Library paper includes fabricated references
A peer‑reviewed article in the New Review of Academic Librarianship, authored by two Nigerian university librarians, examines Z‑Library’s role in providing free scholarly works. The authors acknowledge using Consensus…
1 source HN 41torrentfreak.com -
PHP discloses critical PHP2Shell vulnerability affecting 70% of websites, discovered using GPT-5.6 Sol
The PHP development team released patches on October 21, 2026, for two critical memory-safety vulnerabilities grouped under the name PHP2Shell. The flaws, present in PHP versions dating back to 5.4, allow…
30 sources HN 98Wired AIThe Guardian AIArs Technica AI404media.co +17 more -
Coding agent harnesses triple token costs without boosting benchmark scores, study finds
Complex harnesses for coding agents inflate token bills by up to three times while barely improving benchmark performance over minimal setups, according to a study by researchers Liu and Han at Shandong University. The…
1 sourceThe Decoder -
Claude guide explains AI basics and differences from ChatGPT and Gemini
A new educational guide from Search: Claude clarifies common misconceptions about artificial intelligence by distinguishing between the broad field of AI, machine learning, and specific products like ChatGPT, Claude,…
1 sourcenote.com -
Researchers build KaML-ESM models to predict protein electrostatic properties
Researchers introduced KaML-ESMs, sequence-based neural networks built on ESM protein language models, to predict protein ionization states. The models were trained on the GAINES-augmented PKAD-3r dataset and achieved…
1 source primary sourceNature Machine Learning -
Kyutai releases CADRE open video autoencoders to reduce world model lag
Kyutai has released CADRE, a family of seven open causal diffusion representation autoencoders designed to reduce action latency in interactive world models. The autoencoders feature approximately 476 million trained…
1 sourceMarkTechPost -
OpenAI and Poland's NASK collaborate on AI safety research
OpenAI announced a collaboration with Poland's NASK National Research Institute to work on AI safety, security, and technical research, focusing on evaluating capable AI systems, identifying vulnerabilities, and…
6 sourcesUnite.AInote.comthenextweb.comfinance.yahoo.com +2 more -
Texas Athletics and McCombs School launch Bevo Lab with Dell, Google Cloud, SHI
University of Texas Athletics and the Business of Sports Institute at the McCombs School of Business announced the creation of Bevo Lab, a sports‑intelligence platform that blends AI, data analytics and hands‑on…
1 sourceaol.com -
Openai's GPT-5.6 Sol helps discover Critical PHP vulnerabilities
On October 21 the PHP development team released PHP 8.5.12, 8.4.27, 8.3.36 and 8.2.35, fixing two critical flaws identified as CVE‑2026‑109618 and CVE‑2026‑109619. Both receive a CVSS v4.0 severity score of 9.2 and…
2 sourcesnote.com -
GPT-5.6 SOL tops e-commerce management study, Qwen team reports
Alibaba’s Qwen research team and the Hong Kong University of Science and Technology released results from E-Commerce Bench, a simulation testing AI models’ ability to run online businesses. The study, reported by…
1 sourcedigitaltoday.co.kr -
Can AI spread copies of itself? Considering preparations based on Anthropic's 'self-replication' evaluation
Anthropic released an evaluation on October 12, 2026 that tested whether small AI models could set up inference environments on commercially available IoT devices and chain replication steps within a network. The study…
1 sourcenote.com -
Claude proves security properties for 123 open-source libraries using RISC‑V verification
The authors report that, in collaboration with Theorem, Claude was used to generate formal proofs for 123 open‑source libraries compiled to RISC‑V binaries. Over 239 entry points from Google’s OSS‑Fuzz initiative were…
2 sources primary source HN 41The Decoderanthropic.com -
Study finds six AI models act like two due to correlated errors
A new analysis published on Towards Data Science argues that ensembling multiple large language models yields fewer independent opinions than the number of models suggests. The author tested this hypothesis using six…
1 sourceTowards Data Science -
Boehringer Ingelheim and Microsoft aim to develop AI-backed digital patient twins
Boehringer Ingelheim and Microsoft announced a strategic collaboration to explore AI‑enabled digital patient twins that model individual patients using clinical and other relevant data. The partners said the work will…
2 sourcesfinance.yahoo.com -
Nature Machine Intelligence publishes guidelines for evaluating protein‑ligand scoring methods
The paper by Tran‑Nguyen and Ballester in Nature Machine Intelligence proposes a comprehensive set of guidelines for assessing AI‑based scoring functions used in protein–ligand pose generation and binding affinity…
1 source primary sourceNature Machine Learning -
Anthropic's Claude autonomously designs protein binders against scorpion neurotoxins
Anthropic's Claude, in collaboration with Citrini Research, autonomously designed protein minibinders targeting five scorpion neurotoxins from four species, generating 318,484 candidate designs over a 24-hour…
5 sources primary source HN 41The Verge AInote.comvectorwire.aiThe Decoder +1 more -
Opinion: AI model predicts cancer survival from selfies, but its usefulness is uncertain
Researchers at Mass General Brigham have built a suite of AI tools called FAHR‑Face. The facial‑age estimator, FAHR‑FaceAge, was trained on more than 40 million facial images, while the survival predictor,…
1 sourceThe Guardian AI -
Microsoft explains why AI loses track in long conversations
In a podcast released on October 6, 2026, Microsoft Research researchers Chad Atalla and Jennifer Neville described "surprising failures" that appear when large language models are used in multi‑turn dialogues. They…
1 sourcenote.com -
Microsoft Research outlines AI as a Living Medium for interactive text shaping
Microsoft Research presented its "AI as a Living Medium" project, which treats generative AI not as a static answer box but as a manipulable material. The approach expands interaction beyond prompting, allowing users…
1 sourcenote.com -
MIT historian says early AI engineers added randomness to mimic human creativity
Benjamin Lindquist, an MIT historian, explains that mid‑20th‑century engineers deliberately introduced randomness into algorithms to model the “messy” human mind. He traces the idea to figures like John Pierce, who…
1 source primary sourceMIT News on AI -
Researchers introduce MemPlan for text-based planning with PDDL
A new framework called MemPlan bridges language models and symbolic planning for partially observable text environments. The system uses memory-conditioned PDDL planning to reconstruct problems dynamically from…
1 source primary sourcearXiv cs.CL -
Researchers present CiteAnchor framework for detecting hallucinated citations
A new benchmark called CiteResolve, containing 423 real‑world hallucinated citations annotated for source recoverability, is introduced alongside a source‑aware grounding system named CiteAnchor. The framework tailors…
1 source primary sourcearXiv cs.CL
