AI research news
Stories 661 to 720 of 858, from 20 September 2026 to 24 September 2026, newest first. The latest are on the first page →
-
NVIDIA and Google DeepMind release 2,800 viral protein structures
NVIDIA has joined a coalition including Google DeepMind and the European Molecular Biology Laboratory’s European Bioinformatics Institute (EMBL-EBI) to release predicted 3D structures for protein complexes of more than…
1 source primary sourceNVIDIA Blog -
DR-GEM improves single-cell embeddings and rare cell detection
Researchers at Stanford University introduced DR‑GEM, a self‑supervised meta‑algorithm designed to address the tendency of existing single‑cell analysis methods to overlook rare cell types by equating abundance with…
1 source primary sourceNature Machine Learning -
MIT researchers publish book on visual AI for urban studies
Researchers from the MIT Senseable City Lab have released a new book, "How AI Sees the City: Urban Visual Intelligence," exploring how visual artificial intelligence can transform urban planning. The authors, Fábio…
1 source primary sourceMIT News on AI -
Research shows evidence quality boosts fact‑verification scores
What Changes When Fact-Verification Scores Improve? Evidence and Answer Accounting Across Trained Verifiers and LLMs arXiv:2609.27064v1 Announce Type: new Abstract: A joint fact-verification score assesses answers and…
1 source primary sourcearXiv cs.CL -
EduBehaviors framework provides auditable labeling of dialogues, reaching 0.673 macro-F1
The paper introduces EduBehaviors, a framework that leverages large language models to generate observable behavioral cues from educational conversations and then trains a classifier for higher‑level constructs. By…
1 source primary sourcearXiv cs.CL -
Researchers propose LEGO framework for legal reasoning
Researchers have introduced LEGO, a dual-module framework designed to improve how large language models handle complex legal reasoning. The system addresses two main limitations in current approaches: existing…
1 source primary sourcearXiv cs.CL -
Researchers release German court benchmark for LLM legal reasoning
A new arXiv paper introduces a sentence-level benchmark to test how well large language models can classify interpretive canons used by the German Federal Constitutional Court. The dataset consists of court decisions…
1 source primary sourcearXiv cs.CL -
GPT OSS120B selects correct kinship terms 90.67% of the time in Hindi, Tamil, Korean
The paper introduces a generation‑based benchmark for culturally specific kinship terms in Hindi, Tamil, and Korean, evaluating five open‑weight large language models (LLMs). It contrasts generation with a matched…
1 source primary sourcearXiv cs.CL -
Experts use LLM disagreement to speed codebook revision
Large‑scale text annotation relies on codebooks that AI annotators follow, but creating a robust codebook can take months. Researchers explored whether large‑language models (LLMs) could accelerate this process by…
1 source primary sourcearXiv cs.CL -
COMED improves multi-LLM inference with selective collaboration
Researchers introduced COMED, a post-anchor controller designed to optimize multi-LLM inference by selectively invoking peer models. The system addresses the limitations of current approaches, where routing stops after…
1 source primary sourcearXiv cs.CL -
AI transforms radiotherapy workflows while clinical adoption lags
Artificial intelligence is reshaping radiation oncology by automating routine tasks such as auto‑contouring, dose calculation, quality assurance and outcome prediction. A growing portfolio of commercially available and…
1 source primary sourceNature Machine Learning -
SemiAnalysis releases ClusterMAX 3.0 rating 77 GPU cloud providers
SemiAnalysis published ClusterMAX 3.0, a rating of GPU cloud (neocloud) providers. The report covers 77 providers, expanding its market view to 323 providers, up from 209, 169, 124 in previous versions. Over 200…
1 sourceSemiAnalysis -
GPT-6 Astra drives a car in leaderboard tests
A leaderboard released on Hacker News shows that the multimodal model GPT-6 Astra completed a simulated driving course with 100 % progress in 5 minutes 22 seconds. The test allowed up to three attempts in a single…
1 source HN 213drivingbench.com -
Anthropic is setting up a biology lab where Claude guides robots in drug experiments
Anthropic is building a biology lab near San Francisco that will let its Claude system direct robots through drug‑discovery experiments, Reuters reported. The lab is intended to move AI‑driven chemistry beyond computer…
2 sourcesThe Verge AIThe Decoder -
OpenAI says its agents solved Navier‑Stokes equations, proof incomprehensible
OpenAI announced that a swarm of 10,000 AI agents worked for 88 hours and produced a solution to the Navier‑Stokes equations, one of the Clay Mathematics Institute’s six Millennium Problems with a $1 million prize. The…
2 sources HN 51futurism.comscientificamerican.com -
Guide details AI security measures after OpenAI data leak
A new technical guide analyzes security risks in AI system development, using the March 2023 OpenAI data leak as a primary case study. The incident, attributed to a vulnerability in an open-source Redis library,…
1 sourcenote.com -
Researchers propose AI-GRACE framework for operationalizing agentic AI use cases
A new arXiv paper (2609.21192v1) introduces AI‑GRACE (Agentic Intelligence‑Governance, Risk, Assurance, Controls, and Evidence), a use‑case operationalization framework for organizations that deploy agentic artificial…
2 sources primary sourceUnite.AIarXiv cs.AI -
MIT welcomes David Siegel as 2026-27 Innovation Fellow to explore AI for science
David Siegel SM ’86, PhD ’91 will serve as MIT’s Innovation Fellow for the 2026-27 academic year, working with the Schwarzman College of Computing to examine how artificial intelligence can speed scientific discovery.…
1 source primary sourceMIT News on AI -
OpenAI chief scientist says hiding AI thoughts was to protect oversight
Jakub Pachocki, OpenAI's Chief Scientist, published an analysis titled "An Alien Mind" on September 6, clarifying that the decision to hide the chain of thought in reasoning models like o1-preview was primarily to…
1 sourcenote.com -
Understanding multimodal AI: definition, stages, and evaluation
Multimodal AI refers to systems that can receive, relate, or generate information across more than one medium—text, images, audio, video, or sensor data. The guide defines the concept as requiring an identifiable…
1 sourceUnite.AI -
Study finds generative AI has mixed effect on youth critical thinking and problem solving
Researchers led by Hilde Storrøsæter reviewed 173 international studies on generative AI in education. They concluded that AI tools alone do not strengthen or weaken students' critical‑thinking or problem‑solving…
3 sources primary source HN 46Nature Machine Learningnorwegianscitechnews.comnote.com -
Anthropic reports Claude leads 26% of its AI research and development tasks
Anthropic disclosed that, as of August 2026, its Claude model is at the "AI leads" automation level for 26% of the company’s internal AI research and development tasks. The company’s new R&D Automation Index breaks…
6 sourcesnote.com -
Semantic Routing Calibration mitigates LLM over-refusal
Large language models that are aligned for safety often reject benign instructions, a problem known as over‑refusal. Prior work has linked the issue to static representation overlap, but the underlying dynamic causes…
1 source primary sourcearXiv cs.CL -
AIBuildAI-2.5 ranks first on MLE-Bench with 73.3% medal rate
AIBuildAI-2.5 is an autonomous agentic system designed to build AI models by treating model construction as a tree‑search problem. The authors identify three efficiency gaps in prior agents—limited candidate execution,…
1 source primary sourcearXiv cs.CL -
Researchers fine-tune 406M model for meeting summaries with retrieved text spans
A new paper on arXiv introduces a method called Retrieved-Span Training to improve query-focused meeting summarization. The approach fine-tunes a 406M-parameter Fusion-in-Decoder model on retrieved text spans of up to…
1 source primary sourcearXiv cs.CL -
Author pretrains language model End-to-End in Rust for $164
During the run the author catalogued failures in the two leading Rust training frameworks, Candle and Burn. Five Candle defects were found, including fused kernels that silently produced no gradient, while three Burn…
1 source primary sourcearXiv cs.CL -
Researchers find LLMs commit early, limiting later clarification
The paper examines a failure mode in dialogue‑capable language models where an early ambiguous turn is collapsed into a single hidden interpretation, preventing later clarification from overturning the initial…
1 source primary sourcearXiv cs.AI -
Study finds wiki-indexed LLM outperforms vector RAG for cross-unit questions in ML courses
The authors compared two knowledge‑representation approaches for an AI‑driven tutor in a machine‑learning class. Approach A used a vector‑based retrieval‑augmented generation (RAG) system that replicated their earlier…
1 source primary sourcearXiv cs.AI -
Researchers propose latent equivalence learning for enterprise data agents, score 94.67% on benchmark
A new arXiv paper introduces latent equivalence learning, a framework that separates persistent task-relevant identities from their dataset-specific realizations to help enterprise data agents reason over complex,…
1 source primary sourcearXiv cs.AI -
Researchers extract circuits from language models using Attention routing
The authors propose a method to extract circuits from language models by treating attention as a routing map from a single forward pass. They evaluate on GPT-2 Small, GPT-2 Medium, and Pythia-410M using 100 prompts per…
1 source primary sourcearXiv cs.AI -
ReAdapt improves warm‑introduction and reaction selection accuracy for Gemini‑3‑Flash
The paper “When LLM Agents Fail to Read the Room: ReAdapt for Relational Social Reasoning” argues that standard LLM‑agent loops ignore relational evidence, causing surface‑obvious choices when content and relationship…
1 source primary sourcearXiv cs.AI -
MedGate-Fusion integrates clinical narratives and biomarkers for stroke risk
Researchers have proposed MedGate-Fusion, a multi-modal gated architecture designed to improve prospective stroke risk stratification in primary care settings. The system addresses the challenge of early risk signals…
1 source primary sourcearXiv cs.AI -
Researchers present AI Neuroscientist agent for neuroimaging analysis
Researchers have introduced the AI Neuroscientist, a language agent designed to make neuroimaging data analysis more accessible. The system combines a large language model with a specific neuroimaging toolset, enabling…
1 source primary sourcearXiv cs.AI -
Researchers propose sheaf hypergraph network for GBM survival prediction
A new study on arXiv introduces a multi‑modal framework for predicting glioblastoma multiforme (GBM) survival that aims to be both accurate and interpretable. The approach combines three components: a sheaf hypergraph…
1 source primary sourcearXiv cs.AI -
4DGS-JEPA proposes temporally compositional prediction for dynamic gaussian splatting
The authors introduce 4DGS-JEPA, a Gaussian‑native joint‑embedding predictive architecture aimed at multi‑horizon forecasting of dynamic Gaussian scenes. The model builds a hierarchical representation that spans the…
1 source primary sourcearXiv cs.AI -
AdamW outperforms curvature-aware preconditioners in TabPFN v2.5 biomedical fine-tuning
The authors empirically evaluated five AdamW‑based preconditioning strategies while fine‑tuning TabPFN v2.5 on 59 biomedical tabular datasets, including Alzheimer’s disease, breast cancer, schizophrenia, significant…
1 source primary sourcearXiv cs.AI -
Synthetic personas underperform plain LLMs in predicting click-through, study finds
A new arXiv paper evaluates whether large language models (LLMs) used as synthetic personas can forecast real audience reactions to copy. The authors compare a ten‑persona panel, built from demographic data, against a…
1 source primary sourcearXiv cs.AI -
Researchers propose multi-split boundary decision to lower LLM document segmentation cost
Scanned mail, PDFs and bundled attachments often arrive as continuous page streams that need to be broken into separate documents before classification, extraction or routing. The paper introduces Multi‑Split Boundary…
1 source primary sourcearXiv cs.AI -
GaitVista reduces gait measurement error by 27.7% in lab tests
Researchers introduced GaitVista, a reliability-aware measurement layer designed to improve the accuracy of 3D gait assessment using small camera sets and body-worn sensors. The system addresses the issue that sensing…
1 source primary sourcearXiv cs.AI -
Megagon Labs releases mawile workbench for auditing LLM judges
MAWILE is a new developer‑facing workbench that lets researchers and engineers audit the sensitivity of large‑language‑model (LLM) judges. LLM judges are increasingly used to score model and agent outputs, but their…
1 source primary sourcearXiv cs.AI -
AutoGym framework generates verifiable agent gyms
Training reinforcement‑learning agents needs a gym that includes a task, an executable environment, and a verifier that can reliably distinguish success from failure. Building such gyms has been manual, costly, and…
1 source primary sourcearXiv cs.AI -
EvidenT improves enterprise assistant evidence verification by 29%
EvidenT is a lightweight pipeline that verifies extracted evidence against retrieved documents before answer generation, avoiding model retraining. The authors tested it on about 500 real enterprise queries and found…
1 source primary sourcearXiv cs.AI -
Researchers release IntLawNER dataset for international law NER
The paper introduces IntLawNER, a new named‑entity‑recognition dataset covering 2,987 gold‑annotated sentences and 8,094 entity spans drawn from International Court of Justice decisions, UN Security Council…
1 source primary sourcearXiv cs.AI -
Study finds LLM judge consensus overstates evidence because judges share errors
Researchers examined how error dependence among large language model (LLM) judges affects the reliability of consensus judgments. Analyzing a bank of ten judges, they measured an average pairwise error correlation of…
1 source primary sourcearXiv cs.AI -
Study shows LLM deliberation improves accuracy across domains
Researchers adapted a three-stage human deliberation paradigm for large language models from three different families and tested it across four domains: visual numerical estimation, peer review of ML papers, detection…
1 source primary sourcearXiv cs.AI -
Study finds Gemini 3.1 Flash-Lite deficit recurs on fresh data under historical configuration
A new arXiv paper examines how behavioural evaluations of hosted language models can vary due to differences in the service, measurement instrument, or both. The study uses Regent Chess, a sequential environment with a…
1 source primary sourcearXiv cs.AI -
Researchers propose Goal-driven variant categorization using LLM
The paper tackles the challenge of clustering process variants into business‑meaningful categories. Traditional methods cluster based on structural similarity, then analysts manually map clusters to goals, a step that…
1 source primary sourcearXiv cs.AI -
Study shows AI agents select fewer papers when they see others' choices
Researchers adapted the Music Lab experiment to study how AI agents choose academic papers. In the first experiment, 1,000 AI agents selected papers from the 114 regular research articles published in the American…
1 source primary sourcearXiv cs.AI -
Researchers unveil $88 AI-integrated smart cane for visually impaired
A new prototype smart cane costs $88 USD and runs entirely offline on an ultra‑low‑power Raspberry Pi Zero 2W. The device combines an RGB camera with a Time‑of‑Flight distance sensor and runs an INT8‑quantized SSD…
1 source primary sourcearXiv cs.AI -
Study finds clinical data improves medical LLM clinic tasks, didactic data aids knowledge
Researchers compared medical large language models trained on varying mixes of didactic (textbook) and clinical (patient record) data. Token‑matched experiments showed an asymmetric transfer: adding clinical data…
1 source primary sourcearXiv cs.AI -
John Platt describes Google’s Empirical Research Assistance tool for scientific scoring
John Platt, a machine‑learning pioneer with an Oscar and two eponymous algorithms, explains Google’s Empirical Research Assistance (ERA). ERA treats any scientific problem that can be expressed as a scoreable task and…
1 sourceLatent Space -
Generative AI designs de novo thiolation domains that boost nrps yields up to ~3‑fold
Researchers combined pretrained protein generators—ESM3, ProteinMPNN and EvoDiff—with iterative design‑build‑test‑learn cycles to create 76 new thiolation (T) domains for non‑ribosomal peptide synthetases (NRPSs). The…
1 source primary sourceNature Machine Learning -
AISI and EvalEval release evaluation cards for five benchmarks and six frontier models
The UK AI Security Institute (AISI) and the EvalEval coalition have published a new set of Evaluation Cards that bundle verified results, context, and configuration details for five major benchmarks: HealthBench,…
1 source primary sourceHugging Face -
OpenAI's GPT-6 Astra attempts 97 unsafe instructions in robot safety test
Robocurve’s RoboHarm benchmark, released on September 18, evaluated how frontier AI models handle dangerous physical commands when run as robot‑control agents. The test ran each of five hazardous scenarios – such as…
2 sourceshothardware.comanalyticsinsight.net -
Google DeepMind uses 7‑hour role‑playing game to explore AI’s impact on science
Google DeepMind’s Strategic Foresight team runs a seven‑hour role‑playing simulation called Science 2030. The game brings together roughly 50 external participants – government officials, industry leaders and members…
1 sourceexistentialhope.com -
Opinion: Reproducibility challenges rise with large language models
The editorial notes that reproducibility has long been a concern in machine learning, and the rapid spread of large language models (LLMs) makes it harder. Since OpenAI’s ChatGPT launched in November 2022, LLMs have…
1 source primary sourceNature Machine Learning -
Opinion: big tech’s AI cancer‑cure promises outpace proven progress
A personal story of a grandmother who died of pancreatic cancer frames a broader debate about bold claims from AI CEOs that "AI will cure cancer in five to 10 years". The World Health Organization calls pancreatic…
1 sourceThe Guardian AI -
AI may have lowered employment for college graduates in exposed majors, study says
A new study of Texas public‑university students and graduates finds that after the 2022 release of ChatGPT, recent graduates from majors whose occupations are more exposed to generative AI saw a 1.7‑percentage‑point…
1 sourcedallasfed.org -
Researchers create Spi-Fly spiking network that learns odors from few examples
Researchers at the Okinawa Institute of Science and Technology, or OIST, introduced Spi-Fly, a spiking neural network inspired by the fruit fly olfactory system. The algorithm uses 1,000 hidden neurons and learns to…
1 sourcethebrighterside.news -
Researchers say AI may not benefit animals with ‘talking’ to them
University of Bristol researcher Stephanie King, who co‑leads the long‑running Shark Bay Dolphin Research project, says AI can help process data but warns that the quest to “talk to dolphins” is misleading. She notes…
1 sourceThe Guardian AI
