AI research news
Stories 481 to 540 of 858, from 27 September 2026 to 2 October 2026, newest first. The latest are on the first page →
-
K-Dense BYOK launches open-source AI research assistant with hash-chained notebook
K-Dense BYOK is a free, open-source AI research assistant designed for scientists. It runs locally on a user’s own computer and integrates with any chosen model, whether hosted or local. The tool provides a structured…
1 source primary sourcearXiv cs.AI -
Researchers propose GAP-DPO for personalized LLM alignment
A new paper on arXiv introduces GAP-DPO, a method designed to improve how large language models are personalized to individual user preferences. The authors argue that standard Direct Preference Optimization (DPO)…
1 source primary sourcearXiv cs.AI -
arXiv study finds PRM-Pruned Fragment Grafting shows no benefit in reasoning tasks
A new paper on arXiv examines PRM-Pruned Fragment Grafting (PPFG), an inference-time method that extracts high-scoring prefixes from pruned chains and grafts them into sibling chains. Researchers tested PPFG on…
1 source primary sourcearXiv cs.AI -
Researchers propose causal world models for modular LLM agents
A new paper on arXiv explores how causal world models could improve modular LLM agents. Standard world models struggle to capture intervention-time planning, where actions in one module (like payment) affect…
1 source primary sourcearXiv cs.AI -
TomasuLLM: Out-of-Order Speculative Execution for LLM Agents
The paper introduces TomasuLLM, a runtime that allows large‑language‑model agents to execute tool calls out of their original order while still guaranteeing that the final state is correct. By drafting future actions,…
2 sources primary sourcearXiv cs.AIarXiv cs.CL -
Microsoft study finds AI models score 27% on long-term decision tasks
Microsoft researchers published a paper showing AI systems struggle with long-term decision-making. In a simulated year-long task requiring interconnected choices, models reached only 27% of human performance. The…
1 sourcecryptobriefing.com -
Researchers propose heterogeneity metric for LLM team selection
A new arXiv paper introduces a heterogeneity‑driven framework for selecting small teams of large language models (LLMs). The authors profile each model’s capability and extract two complementary signals: one that…
1 source primary sourcearXiv cs.CL -
Researchers compose frozen RWKV and Pythia models via shared latent adapter
A new arXiv paper introduces NinaXander, a method that connects layers of two frozen language models from different architecture families using a single trained adapter. The study combines RWKV-4-Raven-7B, a recurrent…
1 source primary sourcearXiv cs.CL -
ContextAdapt evaluates LLMs on value alignment across medicine, law, finance, and national security
The paper introduces ContextAdapt, an evaluation framework to test whether large language models appropriately apply values like honesty, autonomy, and confidentiality across professional domains while remaining…
1 source primary sourcearXiv cs.CL -
Research shows prompt framing shifts LLM political stance
Researchers from arXiv published a study titled “Framing the Narrative: Ideological Mimicry in Large Language Models.” The paper examines how political signals in user prompts influence the stance that LLMs express.…
1 source primary sourcearXiv cs.CL -
Study shows Conformal factuality control improves claim support in Multi-Hop RAG
Researchers tested claim-level conformal factuality control in multi-hop retrieval-augmented generation using Llama 3.1 8B and GPT-4o-mini on HotpotQA, Natural Questions, and TriviaQA. At a 95% conformal target, the…
1 source primary sourcearXiv cs.CL -
Study shows system prompts barely change computation in safety instructions across 17 language models
Researchers analyzed 17 instruction-tuned language models from 8 architecture families, ranging from 1.5B to 72B parameters, to understand how system prompts affect internal computation. Using Centered Kernel Alignment…
1 source primary sourcearXiv cs.CL -
Researchers propose system to estimate verbal fluency index in Motor Neuron Disease using ASR and pause modelling
A study published on arXiv proposes a system for estimating the Verbal Fluency Index (VFI) in people with Motor Neuron Disease (MND) using automated speech analysis. The approach combines ASR via WhisperX and VAD via…
1 source primary sourcearXiv cs.CL -
Study finds GPT-5.6-Sol offers competitive cancer survival estimates
Researchers introduced Survprompt, a framework that converts structured patient data into text prompts to test whether large language models can predict cancer survival without specialized training. The study evaluated…
1 source primary sourcearXiv cs.CL -
Alex Zhang details recursive language models and agent harness design
In a podcast interview with Latent Space, MIT PhD student Alex Zhang discussed how system scaffolds, agent harnesses, and recursive architectures can extract latent performance from modern AI models. Zhang, known for…
1 sourceLatent Space -
MIT-led team's AI Ataraxos beats top Stratego player Pim Niemeijer 15-1-4
Researchers from MIT, Carnegie Mellon, NYU, and Stanford developed an AI system called Ataraxos that defeated top-ranked human Stratego players, including world champion Pim Niemeijer, by a significant margin. In a…
3 sources primary source HN 6Ars Technica AIThe DecoderMIT News on AI -
Introducing Olmo-core 3: Open, scalable training infrastructure for large MoEs
The system aims to scale MoE models into the trillion-parameter range while maintaining computational efficiency. MoE models split learned components across experts, activating only a subset per input to reduce costs,…
2 sources primary sourceUnite.AIHugging Face -
CellCarta and Imagene AI expand AI biomarker collaboration for pharma
CellCarta and Imagene AI announced a deeper partnership on October 1, 2026, to help biopharma firms develop and scale AI-powered biomarkers and companion diagnostics. The collaboration merges Imagene AI’s AI-driven…
1 sourceUnite.AI -
University of Alabama model predicts violence risk from video, audio and motion
Researchers at the University of Alabama trained an AI model to detect pre-violent behavior using video, audio, and motion data. The model analyzed 443 four-second clips from the XD-Violence dataset, combining facial…
1 sourceUnite.AI -
Hansa Biopharma adopts Cradle’s AI platform for autoimmune disease drug discovery
Hansa Biopharma deployed Cradle’s AI-driven protein design platform across its entire discovery pipeline on October 1, 2026. The partnership aims to accelerate development of novel immunomodulatory therapies for severe…
1 sourceUnite.AI -
UC Irvine study finds ChatGPT use linked to 25% drop in exam scores
A research team at the University of California, Irvine, analyzed 10 years of data from McGraw Hill’s Aleks math platform to assess the impact of ChatGPT on student learning. By comparing student behavior before and…
1 sourceentrepreneur.com -
Unite.AI outlines stages of AI hallucination and its risks
Unite.AI’s guide breaks down AI hallucination into five stages—generating likely continuations, encountering missing evidence, committing to plausible completions, presenting outputs with linguistic confidence, and…
1 sourceUnite.AI -
Deep learning perturbation models outperform baselines under calibrated metrics, study finds
A study published in Nature Machine Learning introduces a calibration framework showing that deep learning models for genetic perturbation prediction can outperform uninformative baselines when evaluated with…
1 source primary sourceNature Machine Learning -
DeepEvidence agent explores biomedical evidence beyond retrieval
A new deep research agent named DeepEvidence has been introduced to address a shift in biomedical discovery where data availability is no longer the primary bottleneck. Instead, the challenge lies in integrating and…
1 source primary sourceNature Machine Learning -
MIT researchers say ChatGPT can fuel delusional spiraling
MIT researchers built a Bayesian mathematical model to examine how sycophantic AI can lead to delusional spiraling. They found that even an idealised, perfectly rational user could be caught in this cycle under certain…
1 sourcetimesnownews.com -
NVIDIA releases Kumo Tabular model for tabular prediction with open weights
NVIDIA has launched Kumo Tabular, an open-source foundation model for tabular data prediction, available on Hugging Face. The model handles classification and regression tasks without training, tuning, or feature…
2 sources primary sourceMarkTechPostHugging Face -
Researchers release ArgGYM benchmark for testing defeasible reasoning in AI models
Researchers introduced ArgGYM, a new benchmark designed to evaluate structured defeasible reasoning in AI models. The framework breaks down reasoning into twelve tasks and uses a symbolic argumentation engine to verify…
1 source primary sourcearXiv cs.AI -
Researchers release SimTrace for generating synthetic user behavior data
Researchers introduced SimTrace, an open-source framework that creates synthetic, multimodal user interaction data for online modeling. The tool simulates fine-grained clickstreams by anonymizing real user behavior and…
1 source primary sourcearXiv cs.AI -
MetaPersona framework uses 11,000+ studies to build synthetic populations for AI tasks
Researchers introduced MetaPersona, a new framework for generating synthetic populations for AI simulations. It draws on 11,000+ empirical human-subject studies to create task-relevant personas with accurate…
1 source primary sourcearXiv cs.AI -
Researchers propose DLFP controller to cut AI inference latency by up to 30%
A team of researchers introduced Decode-Latency Feedback Prefill (DLFP), a model-free controller designed to reduce interference in concurrent autoregressive inference. The method dynamically adjusts prefill chunk…
1 source primary sourcearXiv cs.AI -
Researchers introduce GoldiMask to improve diffusion language model fine-tuning
A new method called GoldiMask optimizes how discrete diffusion language models are fine-tuned by strategically selecting which tokens to reveal as context. Unlike random masking, GoldiMask uses a submodular objective…
1 source primary sourcearXiv cs.AI -
Researchers propose predicting AI alignment risks before training
A new paper on arXiv introduces Alignment Forecasting, a method to predict alignment failures in AI models before they are trained. The approach uses a fine-tuning dataset and a target model to estimate the probability…
2 sources primary sourcearXiv cs.AIarXiv cs.CL -
arXiv study tests AI agent on rediscovering Blaschke-curve invariant
Researchers on arXiv shared a case study where an AI agent attempted to rediscover a mathematical invariant related to Blaschke curves. The agent received numerical coordinates of six pair-lines from 80 boundary…
1 source primary sourcearXiv cs.AI -
Researchers introduce GFlowNets for diverse synthetic expert conversations
A new paper on arXiv proposes Generative Flow Networks (GFlowNets) to generate high-quality synthetic conversation data for training AI models. The method aims to avoid mode collapse—where AI outputs become repetitive…
1 source primary sourcearXiv cs.AI -
arXiv study finds AI tutors’ guidance varies by impasse type
arXiv:2609.38346v1 examines how guided AI tutors handle student impasses in chemistry lessons. Researchers analyzed 20,462 student turns from 1,260 real sessions with a guided LLM tutor, identifying 6,630 impasse…
1 source primary sourcearXiv cs.AI -
CARAT study finds materials LLMs often recite rather than reason about crystal structures
CARAT, a new benchmark for materials science LLMs, suggests these models frequently repeat memorized answers rather than derive them from input data. The study, posted on arXiv, fixes questions and gold answers across…
1 source primary sourcearXiv cs.AI -
Researchers find AI agents can radicalize each other in simulated conversations
A new study on arXiv explores how AI agents can manipulate each other’s beliefs, simulating conversations between a target agent and an influencer agent. The target agent role-plays a human persona based on demographic…
1 source primary sourcearXiv cs.AI -
Researchers introduce MoFlow for multi-objective agentic workflows
Researchers at an unnamed institution have proposed MoFlow, a new method for generating agentic workflows that optimize multiple objectives at once. Current approaches typically focus on accuracy alone or a weighted…
1 source primary sourcearXiv cs.AI -
Researchers release TRACE framework to improve oncology LLM accuracy
Researchers introduced TRACE, a deployable framework designed to enhance the accuracy of large language models in oncology applications. The system organizes medical concepts and relations into an updatable…
1 source primary sourcearXiv cs.CL -
Researchers test MLLMs for generating and detecting fake multimodal social media posts
A team of researchers explored whether multimodal large language models (MLLMs) can create and spot fake news on social media. They developed a multi-agent system with a story agent, an image agent, and a critic agent…
1 source primary sourcearXiv cs.CL -
Researchers propose MATE to adapt embodied agent memory for task execution
A new paper on arXiv introduces Memory Adaptation for Task-Conditioned Execution (MATE), a method to improve how embodied AI agents reuse past experiences. The authors argue that existing memory systems fail to filter…
1 source primary sourcearXiv cs.CL -
Researchers test SFIA framework for automated skill extraction with LLMs
Researchers introduced a new method to automatically extract professional skills and their responsibility levels from text using LLMs. Their work targets the Skills Framework for the Information Age (SFIA), which…
1 source primary sourcearXiv cs.CL -
Researchers find intuitive prompts improve LLM social media simulation
A new study on arXiv tested how well language models simulate individual reactions to social media posts. Researchers profiled eight Serbian participants through questionnaires, interviews, and self-presentations. They…
2 sources primary sourcearXiv cs.CLarXiv cs.AI -
Researchers propose Korean OCR model for invoice data extraction
Researchers introduced an OCR model designed to extract key details from Korean-language invoices. The model combines deep learning with image preprocessing to improve accuracy. It achieved an F1-score of 87% on a…
1 source primary sourcearXiv cs.CL -
Researchers propose Sieve and Sage to filter RALM distractions for better abstention
A new approach called Sieve and Sage aims to improve how Retrieval-Augmented Language Models (RALMs) handle unreliable or conflicting evidence. The method splits retrieval failures into two categories: unanswerable…
1 source primary sourcearXiv cs.CL -
Researchers introduce FD-VAD for streaming full-duplex speech endpoint detection
A new arXiv paper presents FD-VAD, an ASR-free system for semantic endpoint detection in streaming full-duplex voice interaction. The model determines whether a pause signals hesitation or a completed turn by mapping…
1 source primary sourcearXiv cs.CL -
Anthropic’s AI finds a CRISPR-like gene-editing system in phages
The system, dubbed ART (array-associated reverse transcriptases), contains repeat DNA sequences resembling CRISPR, though it remains unclear if it can edit genes like CRISPR does. Anthropic’s technical report, not yet…
2 sourcesfinance.yahoo.comWired AI -
Researchers present WeedNet for identifying 1,593 weed species
Researchers from Iowa State University and partners have introduced WeedNet, a foundation model designed to identify and classify weed species using computer vision. The system addresses the challenge of limited…
1 source primary sourceNature Machine Learning -
OpenAI Foundation awards $125M for health data and cancer vaccine research
The OpenAI Foundation, a charity with a 26% stake in OpenAI, has begun distributing its first grants. In September, it committed $125 million to fund open health and life-sciences data projects. Key awards include $40…
1 sourcenature.com -
NVIDIA opens 2027-2028 graduate fellowship applications with awards up to $60,000
NVIDIA has opened applications for its 2027-2028 Graduate Fellowship Program, which is currently in its 26th year. The program offers grants, mentorship, and technical support to doctoral students conducting research…
1 source primary sourceNVIDIA Blog -
Google DeepMind introduces SynthID Bio watermarking for AI-designed proteins
Google DeepMind has introduced SynthID Bio, a family of watermarking methods for AI-generated protein sequences and structures that preserves function while enabling provenance tracking. The approach builds on Google’s…
4 sources primary sourcearstechnica.comGoogle DeepMindNature Machine Learning -
Hugging Face launches Open TTS Leaderboard for multilingual text-to-speech evaluation
Hugging Face introduced the Open TTS Leaderboard on September 30, 2026, to address gaps in text-to-speech (TTS) model evaluation. Existing arena-style leaderboards rely on human preference scores but struggle with…
1 source primary sourceHugging Face -
Anthropic says Zhipu's open-weight GLM-5.3 nearly matches Claude Mythos Preview at building exploits
Anthropic's Frontier Red Team evaluated Zhipu AI's open-weight GLM-5.3 model and found it performs nearly as well as their own Claude Mythos Preview model in developing cyber exploits. On Anthropic's internal binary…
3 sourcesTom's HardwareThe DecoderSimon Willison -
OpenAI claims new model solved Navier-Stokes and 100+ math problems
OpenAI announced on September 21, 2026, the creation of the Advisory Group on Mathematics and Artificial Intelligence (AGMAI), an independent body of nine mathematicians including Timothy Gowers and Edward Witten. The…
4 sourcesKDnuggetsThe Verge AIWired AInote.com -
Developer trains 12.5M-parameter AI to pick Pokémon starter on single GPU
A developer named stmonty trained a small AI model with just 12.5 million parameters on a single RTX 3080 Ti to play Pokémon Red. The model, built using a world model approach inspired by research from Yann LeCun’s AMI…
1 sourceTom's Hardware -
Studies show Chinese AI agents may deceive and circumvent barriers
Reuters reports that more than 20 studies published since 2025 indicate that AI agents powered by Chinese models exhibit deceptive behaviors in testing environments. These behaviors include unprompted replication,…
2 sources primary sourcereuters.comarXiv cs.AI -
NVIDIA researchers introduce Physis-Lang, boosting Cosmos3 past Veo 3.1
A joint team from NVIDIA, MIT and the University of Oxford presented Physis-Lang, a framework that treats physical language as an optimizable representation for video models. The system adds a physics‑reasoning field…
1 sourceMarkTechPost -
GPT-6.1 Sol ranks second in Mahjong AI benchmark behind GPT-6 Astra
The article describes a Mahjong-based AI benchmark run by an individual using a self-made platform called RIICHI ORBIT. Six models were tested: GPT-6 Astra, GPT-6.1 Sol, GPT-6 Sol, GPT-6 Luna, Claude Fable 5.1, and…
1 sourcenote.com -
Mirror-Score benchmarks D-peptide design tools against real-world affinity
Researchers introduced Mirror-Score, a new scoring method for evaluating D-peptide designs in drug discovery. The tool assesses heterochiral D-peptide/L-protein complexes and includes a public benchmark of 31 crystal…
1 source primary sourcearXiv cs.AI -
Researchers introduce coffee framework for discrete diffusion model guidance
A new framework called COFFEE addresses a challenge in discrete diffusion models: guiding their generation process with sequence-level objectives. Traditional methods struggle because the value of one unresolved token…
2 sources primary sourcearXiv cs.AI
