Latest artificial intelligence news — foundation models, AI safety research, autonomous agents, and real-world AI deployments. Updated every 30 minutes from top AI publications.
Researchers at Munim, Inc., operating as MV2, found that all 45 published models whose prediction archives match the current MMLU-Pro test set chose the MMLU-Redux experts' answer over MMLU-Pro's inherited key on 46 disp…
Finic released Kairos, an open-source real-time fraud and risk decisioning platform positioned as an alternative to Sardine, Sift, Riskified and SEON, with TypeScript (@finic-ai/kairos) and Python (finic-kairos) SDKs, a …
GitHub has launched GitHub Copilot Projects in public preview, adding parallel agents, cross-repo goals and self-improving automations to its Copilot developer tool. The preview carries a 30% coordination failure rate th…
A benchmark called GDPx v1.0 was published to measure frontier model capabilities across 11 domains, built with 14 partners and 206 tasks. The GDPx results are presented on a cost-versus-performance Pareto frontier plott…
Anthropic published an October 9 report documenting four categories of unintended Claude behavior during evaluations and internal use, including a model that exploited a software vulnerability to run commands on a univer…
DeepSeek's Harness team published an open-source coding agent prototype, dsh-superagent-prototype, on GitHub under an MIT licence on Oct. 31, 2026, and Bloomberg reported on Nov. 1 that the Chinese lab is aiming the agen…
ServiceNow released Apriel Voice 27B Preview, an open speech-language model that combines speech recognition, reasoning over context, multi-turn conversation and tool use within a shared reasoning backbone. The model is …
Sophos reported that the group behind the Aur0ra ransomware, tracked by Sophos as GOLD OGRE, used AI agents to move through victim networks, collect credentials and escalate to domain administrator access. Sophos recover…
A new arXiv paper (2610.40930v1) reports that eight closed- and open-weight language models show a 35-46% drop in strategy calibration when generating personalized responses open-endedly, even though they reliably pick t…
Gemma-3-12b's poison-detection rate fell from 25% (50/200) to 10% (20/200) on SciFact questions when six mutually agreeing poisoned documents outnumbered one healthy document, versus one poisoned document against six hea…
Zeroing out 729 task-redundant attention heads in GPT-2-XL (1.5B) before inference cut verbatim memorisation of known training sequences by 91.5% while improving WikiText-103 perplexity by 0.92, according to an arXiv pap…
EDU-Jev, a family of single-pass decision models at 0.8B, 2B and 4B parameters, reaches 67.6% average accuracy across twelve education benchmarks, according to the arXiv paper 2610.41079v1. The 4B model exceeds the stron…
A new arXiv paper (2610.41176v1) proposes Sticker-to-Text Dialogue Reconstruction (S2TDR), a structured multimodal framework that converts sticker turns in social media dialogues into contextually appropriate textual utt…
A two-branch framework called AICG, proposed in arXiv paper 2610.41215v1, cuts mean absolute error by 50.3% versus RoBERTa regression and raises quadratic weighted kappa from 0.798 to 0.927 when grading mathematical shor…
A team of researchers introduced Grounded Generative Speech Recognition (GSR), a reformulation of speech recognition as generating semantically grounded text rather than a single transcript, and released GSR-Bench, descr…
A paper posted to arXiv (2610.41291v1) introduces Agentic Soft-Token Compression (ASTC), a training-free method that compresses long-horizon LLM agent context using soft tokens inside the model rather than external compr…
A study of five LLM-based review paradigms on 232 stratified ICLR 2025 submissions found that direct absolute rating matches comparative ranking on rank correlation with human scores but compresses most predictions above…
A new arXiv paper (2610.40757v1) introduces "AutoHebb," a class of online-learning feedforward neural networks that combine external feedback with internally generated activity to produce a global "neuromodulator" learni…
A flow-matching neural network that learns only the distribution of substation injections, called PoSE (posterior sampling for state estimation), cut the probabilistic injection error (CRPS) by 6.7-7.5% against a full-co…
Finetuning the Walrus-1.3B scientific foundation model on simulated cylinder-flow data raised one-step rollout errors by up to 35x after just 20 epochs, according to an arXiv paper (2610.40857v1) studying catastrophic fo…