AI research news
Stories 721 to 780 of 858, from 16 September 2026 to 22 September 2026, newest first. The latest are on the first page →
-
Cisco Talos unveils CAIRN tool that detects AI‑driven closedquorum malware
Cisco Talos researchers introduced CAIRN, a framework that flags AI‑integration traits in malware by analyzing metadata fingerprints and assigning unique IDs. Lead researcher Ryan Fetterman says the system looks for…
1 sourceWired AI -
DeepInstructor agentic framework improves idea evaluation using experience graph from 58,607 peer reviews
Researchers introduced DeepInstructor, an agentic AI framework designed to evaluate scientific ideas by reasoning over structured scholarly experience. The system builds an Experience Graph from 58,607 peer reviews and…
1 source primary sourcearXiv cs.CL -
Researchers identify Context poisoning as Extreme-Value attention interference in Long-Context language models
A new arXiv paper introduces the concept of context poisoning, describing how long-context language models struggle to use decisive evidence when irrelevant or confusing information is added to the prompt. The authors…
1 source primary sourcearXiv cs.CL -
AdaMem improves soft compression for retrieval-augmented generation
AdaMem is a new relevance‑guided soft‑compression framework for retrieval‑augmented generation (RAG). It maps passage‑relevance estimates to a query‑dependent allocation of a fixed memory‑token budget, allowing the…
1 source primary sourcearXiv cs.CL -
Researchers propose CLIC to compare coding behaviors of large language models
Researchers introduce CLIC (Code Learning for Identification and Comparison), a method to analyze and compare how different large language models generate code. Instead of relying only on performance metrics like…
1 source primary sourcearXiv cs.CL -
Climate campaigns on X raise happiness by 9.02 points, reduce action language
A new study published on arXiv examined 364,118 public X posts from climate‑related events over 19 years. Researchers used a weighted lexical model to estimate happiness, hope, collective capability, distress and…
1 source primary sourcearXiv cs.CL -
Researchers introduce SJR architecture that improves multimodal moderation by 23.6% F1
A new arXiv paper presents Summarize‑Judge‑Refine (SJR), a two‑model system that separates multimodal content understanding from policy classification. The multimodal Content Model first creates structured text…
1 source primary sourcearXiv cs.CL -
Researchers propose Zero-Inference prospective term that boosts personal Memory recall
A team of researchers introduced a prospective term for personal memory retrieval that requires no inference at query time. Commitments are stored in an explicit ledger and linked to memory items, giving them a…
1 source primary sourcearXiv cs.CL -
PsyAgentBench tests if LLMs truly mimic human psychological biases
Researchers introduced PsyAgentBench, a new benchmark designed to distinguish between LLMs mimicking human psychological patterns and actually possessing those biases. The study re-runs five classic psychology…
1 source primary sourcearXiv cs.CL -
Indy Autonomous Challenge features first road-course overtakes, Unimore wins
The Indy Autonomous Challenge held on Sept. 3 at Laguna Seca marked the first time autonomous racecars passed each other on a road course. Nine university teams from North America, Europe and Asia competed with…
1 sourceThe Robot Report -
Researchers unveil TangleDiff to design entangled protein hydrogels
A team of authors introduced TangleDiff, a deep‑learning framework that creates homodimeric entangled proteins with programmable features. In silico, TangleDiff generates diverse foldable sequences with a success rate…
1 source primary sourceNature Machine Learning -
University of Bristol proposes learning ensemble framework to test medical AI reliability
Researchers at the University of Bristol have suggested a systematic testing framework for medical AI, called “Learning Ensemble”. The approach mirrors drug approval processes by requiring a structured information…
1 sourceThe Decoder -
Anthropic Claude AI predicts chainlink to rise about 300% by 2027
Anthropic’s Claude AI forecasts that Chainlink (LINK) will reach $35 by January 1 2027, roughly a three‑fold increase from its current price around $12. The projection assumes full bull‑market conditions through the…
2 sources99bitcoins.comcryptonews.com -
Understanding loss functions and their five-stage operating map
Loss functions turn the gap between a model's predictions and the target values into a quantity that learning algorithms aim to minimize. The guide stresses that a loss function involves three practical…
1 sourceUnite.AI -
Chinese open-weight models lead US in downloads and benchmark scores, analysis finds
Chinese AI labs have overtaken U.S. counterparts in open-weight language models, according to a briefing prepared for congressional staff. Since April 2025, Chinese models such as Alibaba’s Qwen family, Z.ai’s GLM‑5.3…
1 source HN 60Interconnects -
OpenAI's GPT-6 Astra and Anthropic's Claude Fable fail safety test in RoboHarm benchmark
Researchers at Robocurve introduced the RoboHarm benchmark to test whether AI models refuse dangerous commands when controlling robot arms. They evaluated three leading models—OpenAI's GPT-6 Astra, Anthropic's Claude…
2 sourcesTom's HardwareThe Decoder -
TatBLiMP benchmarks Tatar linguistic minimal pairs
Researchers have introduced TatBLiMP, the first benchmark for evaluating grammaticality in Tatar language models. The benchmark covers 1248 sentence pairs differing by a single morpheme, with each pair consisting of…
1 source primary sourcearXiv cs.CL -
Recursive Language Models generalize out of domain, study shows
The study examines how limiting a language model's view can affect learning. It compares chain‑of‑thought (CoT) with recursive language models that solve each subtask in isolation. In‑distribution, both approaches…
1 source primary sourcearXiv cs.CL -
Reviser proposes cursor-based text generation
Reviser is a decoder‑only Transformer that generates text as a sequence of cursor‑relative actions on a mutable canvas. At each step it predicts one of three tokens: INSERT(token), MOVE(Δ), or STOP, making the model…
1 source primary sourcearXiv cs.CL -
SAGE system raises grant review agreement to kappa 0.58, beating baseline
Researchers introduced SAGE, a schema‑guided LLM that converts grant rubrics into structured checks and links each judgement to evidence in the application package. In an initial evaluation on 35 nonprofit grant…
1 source primary sourcearXiv cs.CL -
Qwen2.5-Omni-3B adapters boost entity recall in accented conversational ASR
A new three‑stage pipeline targets accented English speech from India, Indonesia and Latin America, where standard ASR systems optimized for word error rate often miss named entities and filler words. First, heuristic…
1 source primary sourcearXiv cs.CL -
Researchers introduce DischargeBench to evaluate LLMs as hospital discharge educators
A team of researchers presents DischargeBench, a persona‑grounded simulation that pits a candidate LLM educator against a Virtual Patient in multi‑turn discharge education dialogues. An Education Monitor Agent oversees…
1 source primary sourcearXiv cs.CL -
Study finds token-level entropy blind in small language models, semantic entropy helps
A new arXiv paper investigates whether entropy‑based confidence signals can improve the accuracy of small language models (SLMs) with fewer than 3 billion parameters that run on consumer hardware. The authors evaluate…
1 source primary sourcearXiv cs.CL -
Embed‑TTT improves rule induction in ARC‑like tasks, authors say
The authors introduce Embed‑TTT, a two‑step test‑time training protocol for Vision ARC (VARC) that first fine‑tunes only a task‑specific embedding and then freezes it while fine‑tuning the backbone. They report that…
1 source primary sourcearXiv cs.AI -
Researchers introduce SpecOpt for agentic molecule specificity optimization
Off‑target protein binding causes many adverse effects in small‑molecule drugs, yet most design methods create new selective compounds instead of improving existing ones. The authors propose a new task, specificity…
1 source primary sourcearXiv cs.AI -
Researchers report agent-based HLS with RTL refinement speeds chip design 2.6×
A new arXiv paper investigates whether large‑language‑model agents can improve chip design by working at a higher abstraction level than traditional register‑transfer‑level (RTL) coding. The authors compare four…
1 source primary sourcearXiv cs.AI -
InterviewPlayground pilot shows LLM recovers 88% clinical items but underreports safety
Health systems that plan to deploy AI‑assisted psychiatric intake tools need a way to verify that those systems meet clinical quality standards. The authors propose a clinician‑grounded evaluation platform called…
1 source primary sourcearXiv cs.AI -
TinyCeNN-LM proposes quality‑gated conversion of attention in pretrained language models
The arXiv paper introduces TinyCeNN-LM, a post‑training conversion framework that swaps the attention mechanism in existing language models with CeNN‑inspired cellular‑recurrent layers. The approach adds bounded local…
1 source primary sourcearXiv cs.AI -
Fine-tuning reshapes LLM representations without aligning causal components, study says
Fine-tuning is a common method for adapting large language models (LLMs) to new tasks, yet its impact on internal mechanisms is not well understood. The authors examine changes in attention patterns and layer‑wise…
1 source primary sourcearXiv cs.AI -
Researchers propose topology-based method to detect LLM hallucinations
A new arXiv paper investigates how the shape of attention‑graph information flow can reveal hallucinated outputs from large language models. The authors compute Forman‑Ricci curvature on attention graphs to expose…
1 source primary sourcearXiv cs.AI -
LoRA boosts SAS vision transformer AUPRC to 0.679 in underwater target recognition
Researchers adapted a DINOv3 Vision Transformer (ViT) for synthetic aperture sonar (SAS) automatic target recognition, a task limited by scarce imagery and noisy acoustic backgrounds. Their three‑stage,…
1 source primary sourcearXiv cs.AI -
RBS-Attention provides Radius-Bounded sparse prefill for Long-Context LLMs
A new arXiv paper introduces RBS-Attention, a training‑free sparse‑prefill technique that uses a centroid base branch and a rescue branch to select attention blocks. The method independently thresholds each branch and…
1 source primary sourcearXiv cs.AI -
Stephen A. Weis factors RSA-896 with Anthropic Claude assistance
Researcher Stephen A. Weis has factored RSA-896, a 270-decimal-digit semiprime, using Anthropic's Claude AI model. The achievement on September 19, 2026, pushed the public factoring record from 862 bits to 896 bits.…
2 sourcesnote.comcryptobriefing.com -
UCI launches WRITE AI Center with $10 million federal grant
The University of California, Irvine’s School of Education has started the National Center for Writing Research to Improve Teaching Effectiveness with Generative AI, called the WRITE AI Center, after receiving a $10…
1 sourceocregister.com -
Emergence finds AI agents develop secret language incomprehensible to humans
Emergence ran an experiment that placed AI systems such as Claude, Gemini and OpenAI models into simulated societies modeled on the real world. When the agents began interacting without supervision, they quickly…
1 sourceaol.com -
Opinion: paradigm shift from rule‑based AI to data‑driven models
The essay traces AI from IBM's Deep Blue, which beat the world chess champion in 1997, to today’s conversational systems like ChatGPT. Early systems relied on human‑written rules and deterministic evaluation criteria,…
1 sourcenote.com -
OpenAI reportedly claims its AI solved Navier–Stokes Millennium Prize problem
OpenAI announced that an AI system produced a solution to the Navier–Stokes existence and smoothness problem, one of the seven Millennium Prize challenges. The solution demonstrates a singularity when a fluid vortex…
4 sources primary sourceThe Guardian AIWired AINature Machine LearningThe Decoder -
Prompt injection changes AI behavior by injecting untrusted instructions
Prompt injection is described as an attack or failure mode where untrusted content supplies competing instructions that alter an AI system’s intended output. The article stresses that the term should refer to a…
1 sourceUnite.AI -
Microsoft and University of Illinois launch StudentSim to train AI tutors
A joint effort by Microsoft and the University of Illinois introduced StudentSim, a system that creates digital replicas of individual learners from very limited data. The replicas mimic each student’s answers and…
1 sourceThe Decoder -
Opinion: author tests ChatGPT, Claude, Gemini predicting UFC 331 fights
The author launched a new series called “Watching Martial Arts with AI” and used ChatGPT, Claude and Gemini to predict the five main‑card fights of UFC 331. The author entered his own predictions a few minutes later at…
1 sourcenote.com -
Chinese robot brain startup may achieve ChatGPT-style breakthrough by next year
A Reuters piece by Laurie Chen, Ju-min Park and Eduardo Baptista reports that a Chinese startup developing a "robot brain" believes it could reach a ChatGPT‑style breakthrough as early as next year. The article does…
1 sourcestltoday.com -
Zhang et al. introduce B-BiLO framework for Bayesian PDE inference
A team of researchers presented a new Bilevel Local Operator Learning framework, called B-BiLO, to perform Bayesian inference for partial‑differential‑equation (PDE) inverse problems. The method samples posterior…
1 source primary sourceNature Machine Learning -
Study evaluates coding harness components for autonomous coding agents
Researchers conducted an empirical study of coding harnesses that translate large language model capabilities into long‑horizon software‑engineering tasks. They kept the execution loop constant while varying three…
1 source primary source HN 214arxiv.org -
Anthropic opens wet biology lab in bay area to run AI‑directed experiments
Anthropic has opened a wet biology laboratory in the San Francisco Bay Area, allowing its Claude AI models to direct robotic equipment while keeping human oversight. The company’s head of life sciences, Eric…
6 sourcesjang.com.pkijr.comTechCrunch AInote.com +2 more -
Google Deepmind introduces Dream-RSI to let AI agents reuse past searches
Researchers at Google and Deepmind have unveiled Dream‑RSI, a method that lets AI agents replay recorded search histories to evaluate new strategies without running costly live experiments. By treating past attempts as…
1 sourceThe Decoder -
ICLR 2027 receives roughly 50,000 abstract submissions
The International Conference on Learning Representations (ICLR) for 2027 has collected about 50,000 abstracts, even though the deadline is still a week away. By comparison, the 2026 edition recorded around 19,500 valid…
1 sourceThe Decoder -
FlowCheck catches silent failures in vibe-coded apps where frontier models miss bugs
A study of vibe coding – where an LLM agent builds a web app from a natural‑language prompt – found that iterative modifications often introduce silent failures, such as UI actions that appear successful but do not…
1 sourceTowards Data Science -
Deepmind says chain-of-thought transparency is at risk
Researchers Rohin Shah and Anca Dragan, writing for the newly launched Deepmind Institute, argue that visible chains of thought (CoT) are a key safety advantage because they let observers see a model’s intermediate…
1 sourceThe Decoder -
Google adds Nobel laureates and top economists to AI & Economy team
Google announced an expansion of its AI & Economy Research Program after launching the AI & Economy ATLAS v1.0 and its interactive open‑access site. The initiative tracks how people use Google’s AI tools at work and in…
1 source primary sourceGoogle AI Blog -
Gartner outlines four AI tiers in warehouse automation
Gartner's analysis released this month identifies four operational AI tiers as logistics operators shift from software trials to live warehouse deployments. The firm attributes the move to persistent worker shortages,…
1 sourceAI News -
Decades-old anonymized medical data may cause AI misdiagnoses, study finds
A new collaboration between German and UK researchers warns that anonymized patient records from decades ago could lead AI systems to treat a 2026 patient as if they were still in the year the record first entered the…
1 sourceUnite.AI -
New framework optimizes LLM inference costs via adaptive model activation
Researchers have introduced "inference networks," a graph-based framework designed to optimize the cost-performance trade-off in Large Language Model (LLM) deployments. By treating different LLMs as nodes in a network,…
5 sources primary source HN 45KDnuggetsarXiv cs.CLarxiv.org -
UN System Data Commons launched, providing AI‑ready global statistics platform
On September 17, 2026 the United Nations and Google unveiled the UN System Data Commons, an open‑source, AI‑ready knowledge graph that aggregates statistics from UN agencies at data.un.org. The platform, built on…
4 sources primary sourcelivemint.comUnite.AITechCrunch AIGoogle AI Blog -
Qwen3.5-4B outperforms larger LLMs on new user-side conflict benchmark
Researchers present UC-Bench, a human‑annotated benchmark that evaluates whether a user’s follow‑up utterance conflicts with earlier intent in a dialogue with a large language model. The benchmark focuses on user‑side…
1 source primary sourcearXiv cs.CL -
Neo-Classic benchmark evaluates linguistic-aesthetic reasoning in Classical Chinese poetry
Researchers present Neo-Classic, a new evaluation benchmark that uses out‑of‑sample, contemporary metrical poems written by experts to probe linguistic‑aesthetic reasoning in large language models. Unlike prior tests…
1 source primary sourcearXiv cs.CL -
Study finds trust and friction issues in major generative AI app reviews
Researchers examined 17,012 English‑language reviews from Google Play and the Apple App Store for six leading generative AI applications—ChatGPT, Gemini, Microsoft Copilot, Claude, DeepSeek and Perplexity. Using…
1 source primary sourcearXiv cs.CL -
Study finds PCA can detect stylistic axes in LLM activations without training
Researchers from arXiv cs.CL present a training-free method to identify stylistic dimensions in large language models (LLMs) by analyzing hidden activations. The approach involves sampling completions at elevated…
1 source primary sourcearXiv cs.CL -
Study finds causal control in Subliminal prompting varies by model depth
Researchers from arXiv cs.CL present a study on subliminal learning, where language models transmit hidden traits through seemingly unrelated outputs. The paper challenges the common explanation of "token entanglement"…
1 source primary sourcearXiv cs.CL -
VRR lets LLMs verify, repair and generate candidates, improving code and reasoning results
A new arXiv paper proposes the LLM-as-an-Improver paradigm and introduces the Verify‑Repair‑Reselect (VRR) pipeline. VRR first generates a pool of candidate solutions, selects the top‑ranked winner with a verifier,…
1 source primary sourcearXiv cs.AI -
QVAC Genesis III: 191.43B-token synthetic STEM corpus improves small model performance
Researchers introduce QVAC Genesis III, a 191.43 billion‑token synthetic corpus focused on STEM education. The dataset spans 19 domains, multiple difficulty levels, and varied educational styles. It is generated with a…
1 source primary sourcearXiv cs.AI
