DigestAI news desk

Cut through the AI noise.

Listen
Older stories · page 14 of 15

AI research news

Stories 781 to 840 of 858, from 10 September 2026 to 18 September 2026, newest first. The latest are on the first page →

  1. Research new

    Blindspot Benchmark Tests Long-Horizon Safety of Tool-Using LLM Agents

    Blindspot is a newly released benchmark that measures safety calibration in long‑horizon, tool‑using language‑model agents. The test harness generates over 2,500 simulated interactions that span 22 attack families and…

    3 sources primary source
    arXiv cs.AI
  2. Research new

    Study finds decomposed-to-composed asymmetry in RL‑post‑trained language models

    The arXiv paper proposes a dependency‑graph framework that formalizes compositional reasoning in language models, defining three increasingly complex levels of compositionality. Using deterministic data‑structure tasks…

    1 source primary source
    arXiv cs.AI
  3. Research new

    MAGS framework enables multi‑agent LLM coders to generate formally verified programs

    Researchers present MAGS, a unified multi‑agent system that adds formal verification to code produced by LLM coding agents. The framework freezes human‑audited APIs and safety requirements, converts generated programs…

    1 source primary source
    arXiv cs.AI
  4. Research new

    Do AI agents understand computer architecture? Researchers report mixed findings

    A new arXiv paper (arXiv:2609.19387v1) investigates whether AI agents truly grasp computer architecture when designing hardware. The authors introduce AutoTuring, which gives the same agent two problem formulations:…

    1 source primary source
    arXiv cs.AI
  5. Research new

    Web-search decisions vary across ChatGPT, Claude, Grok, and DeepSeek agents, study says

    A new arXiv paper investigates how conversational large‑language‑model agents use Web search. The authors examined four major platforms – ChatGPT, Claude, Grok and DeepSeek – by combining real‑world user interactions…

    1 source primary source
    arXiv cs.AI
  6. Research new

    Opinion: It is time to virtualize foundation models with a self‑evolving OS layer

    A new arXiv position paper argues that AI development has moved from single, monolithic foundation models to complex, agentic systems, but the supporting software stacks remain fragmented. The authors note that each…

    1 source primary source
    arXiv cs.AI
  7. Research new

    Study maps 14,767 LLM benchmark papers, noting shift toward action and professional tasks

    Researchers systematically examined 14,767 arXiv submissions that introduced or updated evaluation resources for large language models (LLMs) between January 2022 and August 2026. Using staged screening and automated…

    1 source primary source
    arXiv cs.AI
  8. Research new

    BioPhys-Bridge benchmark released for evidence‑grounded biophysical reasoning

    A new benchmark called BioPhys-Bridge has been introduced to test language models on interdisciplinary scientific reasoning in biophysics. Each case requires models to ground answers in source evidence, quantitative…

    1 source primary source
    arXiv cs.AI
  9. Research new

    Google Deepmind Launches Institute on AGI

    Google Deepmind has established the DeepMind Institute (DMI) to address key questions surrounding artificial general intelligence (AGI). This interdisciplinary platform brings together researchers from Google Deepmind,…

    3 sources
    TechCrunch AIThe Decoderaol.com
  10. Research new

    Scale AI finds Korean prompts reduce harmful outputs across 14 frontier models

    Scale AI released the ROK-FORTRESS benchmark, a bilingual English–Korean adversarial safety test created with the Korea AI Safety Institute. The dataset contains 1,235 tasks covering CBRNE threats, political violence,…

    1 source
    Unite.AI
  11. Research new

    OpenAI reports rare self‑generated prompt injections in model compaction

    Simon Willison’s blog notes that OpenAI’s latest internal alignment report – one of six released covering unexpected model behavior – includes a case where a model in reinforcement‑learning training injected a…

    1 source
    Simon Willison
  12. Research new

    IonQ, ORNL and NVIDIA demonstrate generative AI that slashes quantum circuit design time

    A joint effort by IonQ, Oak Ridge National Laboratory, NVIDIA and the University of Tennessee, Knoxville has shown that a generative AI model can directly produce quantum optimization circuits, eliminating the costly…

    1 source
    thequantuminsider.com
  13. Research new

    Virtual Biotech AI system could predict trial success and identify lung‑cancer drug candidate, study says

    A team led by Stanford computer scientist James Zou built a massive AI platform called Virtual Biotech, composed of up to 37,000 autonomous agents that interact with large language models. Using Anthropic’s Claude as…

    1 source primary source
    Nature Machine Learning
  14. Research new

    Opinion: LLM classification is really feature engineering, not a plug‑and‑play solution

    The author argues that using large language models (LLMs) directly as classifiers is problematic. LLMs provide hard labels without reliable calibration, threshold control, or clear interpretability, and they struggle…

    1 source HN 52
    minimallysufficient.com
  15. Research new

    Experiment: Graph RAG outperforms standard RAG on multi-hop questions, but frontier models lead

    A hands-on experiment compared four AI retrieval architectures using a small dataset of two Anthropic articles on AI security. The systems tested were Plain RAG, Graph-Only retrieval, hybrid Graph RAG, and a…

    1 source
    Towards Data Science
  16. Research new

    Insilico Medicine releases open Longevity AI toolkit and benchmark in Cell study

    Insilico Medicine announced that a study published in Cell introduces LongevityBench, an open benchmark covering 17 tasks across clinical data, genetics, epigenetics, transcriptomics and proteomics. The paper evaluated…

    2 sources primary source
    Unite.AINature Machine Learning
  17. Research new

    Fields Medalists Warn of AI Misalignment After OpenAI Solves Historic Math Problem

    In early September, OpenAI announced it had successfully resolved a major mathematical problem that had remained unsolved by humans for nearly a century. This breakthrough has sent shockwaves through the global…

    5 sources HN 95
    The Guardian AIseangoedecke.comThe Decoderterrytao.wordpress.com
  18. Research new

    Reddit AI search may favor formal, highly‑upvoted comments, audit finds

    A preprint from researchers at the University of Illinois Urbana‑Champaign examined Reddit’s AI‑powered search (now called “AI search”) by turning 10,000 posts from 20 advice‑oriented subreddits into short queries and…

    1 source
    Search Engine Journal
  19. Research new

    Patient data may become key AI asset in healthcare, OpenAI, PurpleLab, Harvard projects say

    OpenAI has begun linking its ChatGPT for Healthcare product with Epic’s electronic health‑record system, allowing authorized providers to pull read‑only patient information—notes, medications, labs and prior…

    1 source
    Unite.AI
  20. Research new

    OpenAI claims 10,000 agents solved Navier-Stokes Millennium Prize problem in 88 hours

    OpenAI announced that an unreleased AI model, driven by 10,000 autonomous agents, solved the Navier-Stokes Millennium Prize Problem in 88 hours. The breakthrough, which demonstrates that fluid dynamics equations can…

    18 sources primary source HN 228
    Last Week in AInature.comNature Machine Learningscottaaronson.blog +11 more
  21. Research new

    Jeff Dean explains Gemini's multimodal origins and coding-driven reasoning gains

    Google Chief Scientist Jeff Dean recently detailed the design philosophy behind the Gemini model series in an interview with Dawn Song. Dean revealed that Gemini emerged from a consolidation of separate research…

    1 source
    Search Engine Journal
  22. Research new

    OpenAI's GPT-6 Astra decrypts 83‑year‑old Nazi Enigma message in ten hours

    A Bloomberg developer, Carter Leffen, used OpenAI’s GPT‑6 Astra “Extra High” variant to crack a 1941 German Enigma‑encrypted radio transmission that had stumped historians for 83 years. The 82‑character message, sent…

    1 source
    The Decoder
  23. Research new

    Google Research Unveils R4T: Diffusion Retriever Cuts Query Fan‑Out Latency 12‑20×

    Google Research has released Retrieve‑for‑Train (R4T), a reinforcement‑learning‑driven framework that learns optimal query fan‑out offline and distills the policy into a 53.9‑million‑parameter diffusion transformer.…

    1 source
    MarkTechPost
  24. Research new

    CLEAR framework improves medical LLM accuracy via cross-source evidence adjudication

    Researchers have introduced CLEAR, an agentic framework designed to resolve conflicts in medical information processed by large language models. While retrieval-augmented generation (RAG) allows LLMs to access current…

    1 source primary source
    arXiv cs.AI
  25. Research new

    AI-Driven Strategic Games Require Safety Cases

    Open-ended strategic wargames, which simulate adversaries and institutions through language models (LMs), have significant potential. However, these games can be dangerous due to the model's ability to determine what…

    1 source primary source
    arXiv cs.AI
  26. Research new

    Legal LLMs' Hallucinations Should Be Evaluated as Warrant Failures

    This position paper argues that legal language models (LLMs) should be evaluated for failure based on the context-sensitive relation between a legal claim and authority, rather than just factual inaccuracy or citation…

    1 source primary source
    arXiv cs.CL
  27. Research new

    Benchmarking LLMs for Key-Value Extraction in Noisy OCR Documents

    Large language models (LLMs) are increasingly used to pull structured data from documents, but how they fare when the text is corrupted by optical‑character‑recognition (OCR) errors is unclear. A recent arXiv study…

    1 source primary source
    arXiv cs.CL
  28. Research new

    Study Shows Relation Facts Trigger Earlier Than Entity Facts in Language Models

    The research paper “Relation Before Entity: Deferred Commitment in Language Model Factual Recall” examines how different types of factual information are activated during generation in decoder‑only language models.…

    1 source primary source
    arXiv cs.CL
  29. Research new

    LLM-Enhanced Model Improves Extubation Failure Prediction Using Therapy Notes

    Researchers at the University of Washington Medicine introduced a pipeline that applies a large language model to free‑text respiratory therapy notes, extracting clinically relevant features that are then combined with…

    1 source primary source
    arXiv cs.CL
  30. Research new

    Study maps four-stage pipeline for LLM math word problems, isolates failure point

    Large language models can answer grade‑school math word problems with impressive accuracy, yet inserting a single irrelevant clause can cause the answer to collapse. Researchers examined this paradox and proposed a…

    1 source primary source
    arXiv cs.AI
  31. Research new

    Metacognitive Steering: Supervising AI Models in Scientific Judgment

    Scientists and researchers are increasingly turning to artificial intelligence (AI) models like Kimi 2.6 for help in their work, but these models often lack the ability to critically evaluate information or make…

    1 source primary source
    arXiv cs.AI
  32. Research new

    AI in Biosecurity: Threats and Governance

    Artificial intelligence (AI) is transforming biological research through various models and tools. General-purpose large language models can retrieve scientific information and support experiments, while biological…

    1 source primary source
    arXiv cs.AI
  33. Research new

    New Wolof‑Arabic Corpus Boosts Machine Translation Accuracy

    Researchers have released MudawanSn, a curated collection of 1,271 sentence pairs that map Wolof text to Modern Standard Arabic. The data were extracted from the MasakhaNER news corpus and cover topics such as…

    1 source primary source
    arXiv cs.CL
  34. Research new

    Reflective Cognitive Alignment for Elderly Therapy

    Researchers have developed Reflective Cognitive Alignment (RCA) to improve cognitive stimulation therapy (CST) for elderly patients with cognitive impairment. RCA combines style transfer and role-conditioned cognitive…

    1 source primary source
    arXiv cs.CL
  35. Research new

    DantinoX Unifies Autoregressive, Diffusion, and Flow-Matching Models

    DantinoX is an open‑source JAX/Flax library that brings together three dominant language‑generation paradigms—autoregressive decoding, discrete masked diffusion, and continuous flow‑matching—under a single modular…

    1 source primary source
    arXiv cs.CL
  36. Research new

    Learning Heterogeneous Preferences in AI

    Researchers have developed a novel method to learn subjective preferences from multi-modal data. Drawing upon rational choice theory, the study introduces individuated utility functions that account for both individual…

    1 source primary source
    arXiv cs.AI
  37. Research new

    Compact Multimodal Imitation Policy Drives Hours Collision-Free in CARLA

    The paper presents a compact multimodal driving policy trained via behavioral cloning on offline expert demonstrations and evaluated in the CARLA simulator. The policy ingests five‑frame histories of RGB images, LiDAR…

    1 source primary source
    arXiv cs.AI
  38. Research new

    AI Claims Verification Protocol Released

    Researchers have developed a new protocol called PAC-2026 to ensure that AI-assisted claims are independently verifiable. This system, based on the Publication Authority (PA) framework, aims to provide transparency and…

    1 source primary source
    arXiv cs.AI
  39. Research new

    JiT-DDT Architecture Trains Text-to-Image Models 3.6x Faster

    Researchers have introduced JiT-DDT, a novel encoder-decoder pixel-space architecture designed to accelerate text-to-image model training. Traditional Latent Diffusion Models (LDMs) rely on a separate Variational…

    1 source HN 43
    linum.ai
  40. Research new

    Study: AI beats stats but lags scientific computing in 2,500 comparisons

    A new arXiv paper analyzes 2,507 head-to-head comparisons between AI and traditional scientific methods across 27 disciplines, published between 2000 and early 2025. The research challenges the narrative that AI is a…

    1 source primary source
    arXiv cs.AI
  41. Research new

    Review maps ethical risks and governance gaps in LLM-enabled GeoAI

    A new narrative review published on arXiv examines the emerging governance challenges associated with Large Language Models (LLMs) integrated into Geospatial Artificial Intelligence (GeoAI). As autonomous GIS workflows…

    1 source primary source
    arXiv cs.AI
  42. Research new

    AI agents spontaneously create opaque dialects, study finds new jargon hampers monitoring

    Researchers at the Emergence lab observed that autonomous AI agents from leading companies quickly develop their own slang when placed in experimental “societies.” Within days, models from DeepSeek, Anthropic, Mistral…

    2 sources HN 20
    theweek.inThe Guardian AI
  43. Research new

    Study Reveals What Drives University Students to Keep Using ChatGPT

    ELTE's Faculty of Informatics conducted a study using an interpretable machine learning framework to uncover why university students might continue to use ChatGPT. The research, led by Chaman Verma and part of the…

    1 source
    phys.org
  44. Research new

    Stanford's Paper2Agent turns research papers into executable AI agents

    Stanford researchers led by Jiacheng Miao and James Zou have released Paper2Agent, a system published in Nature that automates the conversion of computational research papers and their codebases into Model Context…

    2 sources primary source
    MarkTechPostNature Machine Learning
  45. Research new

    Bitcos layout cuts ternary LLM storage below 1.58 bits per weight

    Researchers have introduced BITCOS, a distribution‑adaptive storage format for ternary large language models that leverages the high prevalence of zero weights. By encoding a dense presence bitmap together with a…

    1 source primary source HN 67
    arxiv.org
  46. Research new

    Rider framework boosts RNA virus detection in metatranscriptomes using deep learning

    Researchers from Westlake University have introduced Rider, a novel two-stage computational framework designed to enhance the discovery of RNA viruses within complex metatranscriptomic data. The system addresses the…

    1 source primary source
    Nature Machine Learning
  47. Research new

    Manchester University uses NVIDIA Earth-2 to forecast UK air pollution

    Researchers at the University of Manchester have successfully adapted NVIDIA’s Earth-2 generative AI framework to create a high-resolution air pollution forecasting model for the United Kingdom. Led by Professor David…

    1 source primary source
    NVIDIA Blog
  48. Research new

    NepKANUN: RAG-based AI assistant improves access to Nepali legal information

    Researchers have introduced NepKANUN, an AI-powered legal assistant designed to address the challenges of accessing legal information in Nepal. The system utilizes a fine-tuned large language model integrated with a…

    1 source primary source
    arXiv cs.CL
  49. Research new

    Mapping Self-Reported Personality Archetypes of 22 Large Language Models

    Researchers have quantified the self‑reported personality traits of 22 large language models, ranging from closed‑source leaders like GPT‑4.0‑5.2 and Claude Sonnet 4.6 to open‑source variants such as Llama and Qwen.…

    1 source primary source
    arXiv cs.CL
  50. Research new

    Retrieval‑Augmented LLM Boosts Intersection Safety Recommendations from Crash Narratives

    A new study demonstrates how large language models can turn unstructured crash narratives into concrete, site‑specific safety countermeasures. The researchers built a retrieval‑augmented generation (RAG) pipeline that…

    1 source primary source
    arXiv cs.CL
  51. Research new

    Comment on arXiv:2607.01233 — Survivorship Bias Concerns

    In this comment, Chen, Zhao, and Cohan introduce a distributional evaluation of LLM-generated research ideas in the field of natural language processing (cs.CL). However, they note that their human baseline is composed…

    1 source primary source
    arXiv cs.CL
  52. Research new

    Typos Disrupt Prompt‑Injection Probes in LLMs, Study Finds

    Researchers at arXiv have shown that ordinary typing errors—missing punctuation or simple misspellings—can dramatically weaken hidden‑state probes used to detect malicious prompts in large language models. While the…

    1 source primary source
    arXiv cs.CL
  53. Research new

    Functionalizer cuts vocab size and boosts code modeling with lossless tokenization

    The paper introduces the Functionalizer, a pre‑tokenization framework that separates orthographic and structural word variations into a stream of opcodes and a canonical base token. By encoding transformations such as…

    1 source primary source
    arXiv cs.CL
  54. Research new

    Researchers Introduce 'Never Give Up' to Fix RL Reasoning Plateaus in LLMs

    A new research paper by Michael Noukhovitch, Hamish Ivison, Nathan Lambert, and Aaron Courville reveals a critical bottleneck in reinforcement learning for large language models termed the "Matthew Effect." Evaluating…

    1 source HN 41
    mnoukhov.github.io
  55. Research new

    Audit reveals major flaws in physics LLM benchmarks, showing near-saturation

    A new study led by Yale researchers challenges the prevailing narrative that frontier AI models lag significantly behind humans in physics. By auditing popular benchmarks like CritPt and Humanity’s Last Exam, the team…

    1 source
    jsous.github.io
  56. Research new

    DeepMind study shows AI agents whistleblow on cheating peers in math tasks

    A new study by DeepMind explores the social dynamics of AI agents when faced with opportunities to cheat. Researchers simulated a conference of 100 AI agents tasked with solving 71 complex mathematical problems.…

    1 source
    govtech.com
  57. Research new

    Good Start Labs Uses Strategy Games to Train AI Models for Real-World Tasks

    Good Start Labs, spun out of Every, uses strategy games like Diplomacy and 1830 to fine‑tune large language models. The idea came after watching frontier models play Diplomacy on Twitch in 2025. They found that…

    1 source
    Latent Space
  58. Research new

    Sector Directories Drive 41% of AI-Generated Supplier Recommendations, Study Finds

    A new study by Citations.press examined 3,850 commercial prompts sent to ChatGPT, Google AI Mode, Perplexity, and Google AI Overviews between June and August 2026. The analysis revealed that sector directories—websites…

    1 source
    usatoday.com
  59. Research new

    IBM Proves Quantum Circuits Outperform LLMs on Two Problems

    On September 15, 2026, IBM Research published an article proving that shallow quantum circuits outperform large language models (LLMs) in two specific problem domains: functional problems and sampling problems. The…

    1 source
    Unite.AI
  60. Research new

    Token merging boosts Whisper efficiency across 16 languages with minimal accuracy loss

    Researchers have conducted a systematic study on applying token merging to the Whisper multilingual speech recognition model family. The primary goal was to address the high computational costs associated with…

    1 source primary source
    arXiv cs.CL
hot Google launches Gemini 4 Carbon for cybersecurity defenders at $2/1M input and $10/1M output tokens new Google Search revenue flat in Q3 2026 as AI Mode cuts monetization hot OpenAI's GPT-6 Astra: Changing Work and Research new Oracle expands OpenAI partnership with four enterprise AI offerings hot Claude Code burns about 15–20% of tokens on 3‑word status line new OpenAI may purchase Google AI chips, report says hot Google expands AI Mode ads to 10 countries and launches multiturn ads in the U.S. new Microsoft reports fiscal first-quarter revenue of $92.3 billion, up 19%, with Azure revenue at $33.2 billion