DigestAI news desk

Cut through the AI noise.

Listen
Older stories · page 10 of 15

AI research news

Stories 541 to 600 of 858, from 23 September 2026 to 30 September 2026, newest first. The latest are on the first page →

  1. Research new

    Researchers propose NashEval for context-dependent AI agent evaluation

    A new paper on arXiv introduces NashEval, a framework for evaluating AI agents based on context-dependent preferences. Traditional methods like Bradley-Terry models assume transitive rankings, but human judgments often…

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

    Researchers test whether AI harnesses specialize or just repeat answers

    A new study on arXiv examines whether automated LLM harnesses truly specialize in tasks or simply repeat answers. The paper compares eight generated harnesses against nine identical baseline copies across 386 MATH-500…

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

    Study finds synthetic embeddings match text for LLM fine-tuning

    Researchers from arXiv have published a study questioning whether human-readable text is required for effective fine-tuning of large language models. The paper introduces a method called Desired-Update-Aligned…

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

    Researchers build neurosymbolic router that sends math to exact solvers on Raspberry Pi

    A paper on arXiv describes a neurosymbolic router that classifies incoming queries and routes structured tasks such as arithmetic and logic to deterministic solvers, while sending open-ended word problems to a small…

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

    Simon Willison shares GPT-6.1 Sol pelicans from OpenAI DevDay 2026

    Simon Willison posted a comment on Hacker News linking to his live blog of the OpenAI DevDay 2026 keynote. He shared SVG renderings of GPT-6.1-Sol pelicans, noting they are not notably different from the GPT-6 family…

    1 source
    Simon Willison
  6. Research new

    Swift-1.5-Qwen3.8-27b-oQ8e-mtp achieves 34.8 tokens per second on Apple M5 Max

    A Reddit user shared benchmark results for the Swift-1.5-Qwen3.8-27b-oQ8e-mtp model on an Apple M5 Max device on September 29, 2026. The model reached 34.8 tokens per second in a coding scenario, generating a playable…

    1 source
    llm-bench.io
  7. Research new

    Zhipu automates infrastructure with GLM-5.3’s outer RSI loop in under two weeks

    Zhipu AI, developer of the open-weight LLM GLM-5.3, detailed how it used its own model to accelerate internal infrastructure work. The team deployed an Infra Agent powered by GLM-5.3 to optimize GLM-5.3-Flash—a faster,…

    3 sources primary source
    madrobot.bloganthropic.comImport AI
  8. Research new

    Ohio State and Google launch AI research hub with DeepMind models and Gemini access

    Ohio State University and Google announced a five-year partnership on September 29, 2026, to create an on-campus AI research hub in Ohio State’s Innovation District. The collaboration includes access to Google…

    1 source
    finance.yahoo.com
  9. Research new

    General Intuition raises $220M at $6.2B valuation for AI agents trained on gameplay footage

    The total funding now exceeds $650 million, following a $320 million raise in June. Medal’s user-generated content, including 3 billion annual video uploads, provides a vast trove of real-world interaction data.…

    1 source
    gamesbeat.com
  10. Research new

    Anthropic’s AI Claude discovers CRISPR-like enzyme system ART

    Anthropic announced that its AI model Claude autonomously discovered a new enzyme system called array-associated reverse transcriptases (ART) in a DNA sequence database, noting structural similarities to CRISPR. The…

    26 sources HN 7
    MIT Technology Review AIcnet.comnote.comthedailystar.net +18 more
  11. Research new

    KDnuggets lists five free AI engineering courses for hands-on learning

    KDnuggets published a list of five free, open-source courses to help build AI engineering skills. The author, Abid Ali Awan, organizes them from beginner to advanced, covering LLMs, RAG, MLOps, fine-tuning, and…

    1 source
    KDnuggets
  12. Research new

    Jagged intelligence explains uneven AI performance in complex tasks

    Jagged intelligence describes how AI systems excel at demanding tasks while failing on simpler related ones. The concept focuses on five stages—mapping task performance, probing instruction variations, testing…

    1 source
    Unite.AI
  13. Research new

    Phylogeny-agnostic ML predicts strain-level phage–host interactions

    Researchers developed a phylogeny‑agnostic machine‑learning framework that predicts strain‑level phage–host interactions from genome sequences alone. The team evaluated over 13.2 million training runs across six…

    1 source primary source
    Nature Machine Learning
  14. Research new

    AI may narrow scientific exploration, authors argue

    A new opinion piece in Nature Machine Learning argues that while AI boosts individual researcher productivity, it is narrowing the scope of scientific inquiry. The authors note that academics using AI tools publish…

    1 source primary source
    Nature Machine Learning
  15. Research new

    LLM-generated ACSL contracts for numerical libraries

    A new approach uses large language models (LLMs) to automatically generate a small, human-certifiable Automated C Specification Language (ACSL) contract from function documentation. This combined with a deterministic…

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

    Visualizing RAG Conflicts: Temporal Semantic Divergence Score

    Researchers at arXiv have introduced the Trajectory Variance Score (TVS) to detect conflicts in retrieval-augmented generation (RAG) models. These models often struggle when retrieved context contradicts parametric…

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

    Study suggests dialects do not drive jailbreak success

    A recent arXiv paper investigates how dialectal variations affect large language model jailbreaks. The authors extended a classical Chinese red‑team framework to Shanghainese and Cantonese and ran a 36‑cell ablation…

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

    LLM-guided ontology construction from unstructured texts

    A new approach combines compact open-source Large Language Models (LLMs) with ontology learning to extract and structure knowledge from industrial text. This paper evaluates the method on a private French corpus of…

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

    Researchers test 72,000 RAG combos on Indian government documents

    The work evaluated eight hundred question instances across four distinct Indian central-government regulatory documents, with evidence strings validated against source text and manually reviewed for 10 percent of the…

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

    Researchers release ChestPheNoT for auditable radiology report analysis

    Researchers introduced ChestPheNoT, a compact 0.5–3B language model designed to extract structured medical findings, status labels (present/absent/uncertain), and verbatim evidence spans from radiology reports. The…

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

    BioDyad integrates biomedical discovery with ML program search

    BioDyad is a new agentic system that links biomedical evidence and machine learning engineering. It uses two hierarchies—one for scientific discovery and one for program engineering—to guide candidate construction,…

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

    Researchers propose MedCode to boost LLMs’ medical calculation accuracy by 20–30%

    A new framework called MedCode aims to improve large language models’ ability to perform precise medical calculations. Current LLMs struggle with tasks requiring exact numerical outputs, such as medication dosing or…

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

    Researchers release open protocol for distributed AI textbook formalization

    Researchers introduced Choir, an open protocol for distributed AI-driven formalization of textbooks and theorems. Unlike current centralized approaches, Choir splits tasks across independent contributors, each using…

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

    Researchers introduce countermem to improve AI agent memory with counterfactual checks

    Researchers at an unnamed lab have proposed COUNTERMEM, a reinforcement-learning framework designed to enhance AI agent memory by evaluating hypothetical outcomes. Unlike existing systems that rely solely on factual…

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

    Alibaba releases IndustryLLM for industrial procurement with 35B parameters

    Researchers at Alibaba have introduced IndustryLLM, an open-weight language model designed for industrial procurement tasks. Built from Qwen3.5-35B-A3B-Base, the model uses 35 billion parameters with 3 billion…

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

    arXiv paper finds low overlap in LLM judge validation risks wrong decisions

    A new arXiv paper argues that sparse overlap in LLM judge validation—when only a few items are labeled by multiple annotators—drives poor deployment decisions. The authors prove that at just 5% pairwise overlap,…

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

    Researchers release DriveHierarchy benchmark for VLM driving tests

    DriveHierarchy is a new benchmark for evaluating vision-language models (VLMs) in autonomous driving. The framework breaks down driving competence into four ranks: perceptual grounding, contextual memory, mental…

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

    Researchers propose a design framework for human-AI collaboration

    A new white paper on arXiv outlines a framework for designing human-AI collaboration in industrial settings. The authors argue that successful AI adoption depends on more than technical capabilities—it also requires…

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

    Researchers release CP-Agent for automated crystal plasticity simulations

    Researchers introduced CP-Agent, an AI agent designed to automate crystal plasticity (CP) simulations for metals. The tool uses natural language to execute full workflows, including configuring tools, managing data…

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

    Researchers propose Witness Overlap to track Open-Weight model lineage

    Witeness Overlap: Directional Provenance Inside Open-Weight Model Families arXiv:2609.31784v1 Announce Type: new Abstract: Open-weight models are often released, fine-tuned, aligned, merged, and re-released, making…

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

    SMARtCARE uses privacy-preserving AI for clinical decision support

    Researchers introduced SMARtCARE, a privacy-preserving agentic AI system designed for clinical decision support in intensive care units. The architecture addresses a common limitation in long-context clinical AI, where…

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

    Maker tests GPT-6 Astra for mechanical design of a latte art camera

    A maker tested whether GPT-6 Astra could handle mechanical design by building a latte art camera called LATTE Df. The device uses a shutter-like button to drop cinnamon powder through a stencil onto coffee foam,…

    1 source
    note.com
  33. Research new

    Open-source decision models Jeff-Qwen3.5-0.8B and Jeff-Gemma4-E2B run in ~30ms on local hardware

    Jeff is a set of small, fast decision models fine-tuned from Qwen3.5 and Gemma 4, designed for zero-shot classification tasks. The 0.8B-parameter model runs in about 22ms on an RTX PRO 6000 and 28ms on an Apple M4 Max,…

    1 source primary source HN 59
    github.com
  34. Research new

    Google DeepMind’s AI model adds five days to hurricane warnings

    Google DeepMind’s WeatherNext Cyclones model won the 2026 Gizmodo Science Fair for predicting tropical cyclone tracks and intensity with unprecedented accuracy. During the 2025 Atlantic hurricane season, it gave…

    1 source
    aol.com
  35. Research new

    ConTP uses contrastive learning to predict transporter substrate specificity

    Researchers from King Abdullah University of Science and Technology (KAUST) and Ningxia Medical University introduced ConTP, a new framework for annotating membrane transporters. Traditional methods rely on…

    1 source primary source
    Nature Machine Learning
  36. Research new

    German researchers introduce ReImaGin for image-based chain-of-thought reasoning

    A team of researchers from Germany has developed ReImaGin, a vision-language model (VLM) that uses image generation as part of its internal reasoning process. Unlike previous systems that relied on bolted-on image…

    1 source
    Unite.AI
  37. Research new

    OpenAI contractors fired for using banned AI tools on ChatGPT review tasks

    OpenAI contractors reviewing ChatGPT responses were fired or removed from projects after violating a ban on using AI tools, according to an investigation by 404 Media. The workers, hired to evaluate AI outputs for…

    1 source
    ibtimes.co.uk
  38. Research new

    Artificial Analysis launches Cyber Index Alliance to benchmark AI agents on cyber defense

    Artificial Analysis introduced the Artificial Analysis Cyber Index Alliance, a partnership with Collinear, IBM, Nvidia, and Vercel to create a standardized benchmark for evaluating AI agents in enterprise cyber defense…

    1 source
    artificialanalysis.ai
  39. Research new

    Anthropic misses self-imposed safety deadline for provable-inference prototype

    Anthropic’s September 30 deadline for Phase 1 of its provable-inference prototype has passed without public confirmation of completion. The milestone, originally set for May 15, was delayed to focus on broader security…

    1 source
    finance.yahoo.com
  40. Research new

    Anthropic reports 26% of R&D reaches AI-led stage with 30,000 agents

    Anthropic announced on September 17 that 26% of its research and development work now falls under AL4—where AI handles most tasks independently, with humans providing only broad instructions. This marks a 26-fold…

    3 sources
    note.com
  41. Research new

    Worldmodeldata licenses 1 million hours of game data for AI

    A British startup called Worldmodeldata is packaging video game data to train AI world models, which are designed to understand physical physics and actions. Unlike large language models trained on text, world models…

    1 source
    Wired AI
  42. Research new

    Researchers release benchmark for AI in systematic review screening

    A new paper on arXiv introduces a benchmark dataset and framework for evaluating large language models in systematic review screening. The dataset contains labeled entries to test how well LLMs classify article…

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

    SlideLab framework generates scientific presentations from research papers

    SlideLab is a training‑free multi‑agent framework that builds scientific slide decks directly from research papers. The system first plans a presentation narrative, then iteratively refines a shared slide deck using…

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

    Cartograph reduces AI agent tool discovery from O(n) to O(k)

    Cartograph is a federated Model Context Protocol (MCP) proxy that changes how AI agents find and call tools. Instead of traversing every catalog entry, the system exposes only a few proxy tools, lowering the search…

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

    Survey reviews 211 fake review detection studies from 2018 to 2026

    A new survey published on arXiv examines the evolution of fake review detection methods, covering 211 studies released between 2018 and early 2026. The paper analyzes how the field has shifted from traditional machine…

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

    Researchers audit LLM-as-judge in text-to-SQL pipeline, find low agreement

    Researchers examined the agreement between an LLM-as-judge and human annotators in a production text-to-SQL pipeline. The deployed gpt-4o-mini judge achieved a Cohen's kappa of 0.04 on a disagreement‑enriched set and…

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

    Researchers introduce hierarchical memory system for LLM agents

    A new framework called HiCoMER addresses how collaborative AI agents manage and retrieve memories in team settings. The authors note that current systems treat all stored memories—both team-wide and individual—as a…

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

    Researchers propose attention-free model mixing with autoencoders for masked language tasks

    A new paper on arXiv explores an alternative to attention mechanisms in transformer-based masked language models. The authors introduce a method using autoencoder-based mixing modules to replace attention, reducing…

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

    Study identifies under 1% of BERT neurons driving AI-Text detection

    Researchers examined a frozen BERT‑base‑uncased encoder to determine which neurons support AI‑generated text detection. Using the RAID benchmark across six generators, they applied an L1‑to‑L2 sparse‑probing protocol…

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

    Researchers release Atelier for cryoEM map analysis via hypernetworks

    A team of researchers introduced Atelier, a self-supervised framework that uses hypernetworks to extract localized features from cryo-electron microscopy (cryoEM) volumes. The method leverages implicit neural…

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

    Researchers introduce Benchy, a standardized language for AI task benchmarks

    Researchers have proposed Benchy, a new semantic language and execution engine designed to standardize task-oriented AI benchmarks. The system defines benchmarks as a structured triplet of program, scoring function,…

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

    BioEVAL benchmark evaluates LLMs on bioengineering tasks

    BioEVAL is a global, multi‑institutional benchmark created to test large language and multimodal models on bioengineering tasks. The benchmark contains 608 evaluation items from 22 research groups: 380 multiple‑choice…

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

    Researchers propose S3KG framework to evaluate LLM context

    Researchers have introduced a new evaluation framework designed to test whether large language models truly understand context or merely perform pattern matching. The paper, published on arXiv, argues that traditional…

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

    Researchers train SchNet GNN to predict transmembrane protein topology from 3D structures

    A new paper on arXiv introduces a graph neural network (GNN) called SchNet to predict transmembrane protein topology using 3D structural data. Unlike prior methods relying on protein sequences or alpha-carbon features,…

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

    Synthetic Ground-Truth framework evaluates XAI methods

    The paper introduces a synthetic ground‑truth framework for evaluating explainable AI methods. It addresses the lack of reliable evaluation procedures and the absence of ground‑truth explanations by using controlled…

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

    Researchers propose MCP-based mediation layer for LLM agents and data spaces

    A paper on arXiv introduces an architectural mediation approach using the Model Context Protocol (MCP) to connect large language model agents with data space services. The authors implement the Eunomia Agent as a…

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

    Researchers find Multi-Agent code judge often declares both solutions equally good

    A study posted to arXiv examines how language models judge code correctness without access to ground truth. The researchers ran the MARCH framework on two code judging benchmarks and found it declared both solutions…

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

    ScopeBench benchmark tests agent scope adherence

    The authors introduce ScopeBench, a benchmark of 30 dead‑end agentic security tasks that measure whether autonomous agents stay within a defined scope. Each task is presented twice: once without a scope to gauge raw…

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

    Research paper outlines framework for autonomous systems

    The paper introduces a framework for designing autonomous systems, combining connectionist and symbolic AI. It proposes a generic agent architecture where behavior is built from cognitive functions organized around a…

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

    OpenAI discloses AI agents accessed US government websites in unexpected ways

    OpenAI has acknowledged that its autonomous AI agents interacted with multiple US government websites — including the SEC, Census Bureau, Education Department, Justice Department and Commerce Department — in ways the…

    9 sources HN 46
    bbc.comedweek.orgaol.comblockonomi.com +5 more
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