Technology & AI
BERT-LER: explainable AI reads 75 million health records
BERT-LER, trained on 75 million de-identified patient records, matches benchmark models on clinical prediction tasks while explaining its own reasoning.
AI Models
AI model releases and benchmarks — what each new model actually changes for the people who use it.
191 stories
Technology & AI
LLM corrections usually die with each session, arXiv preprint says
arXiv preprint argues fixing LLM errors is an operations problem, not a tooling problem, and governance for persisted corrections doesn't exist yet.
Technology & AI
Tencent AI-Infra-Guard: 5,200-star AI red team tool trends
Tencent Zhuque Lab's open-source tool scans AI agents, MCP servers and LLM jailbreaks. It has 5,200+ stars and a README that explicitly asks for them.
Technology & AI
Neurosymbolic world model transfers tasks without retraining
An August 2026 arXiv preprint splits RL world models so reward prediction uses only symbolic state, letting agents switch tasks without further training.
Technology & AI
FedLNS catches rogue clients in federated LLM training
A new arXiv preprint proposes screening federated LLM updates by tracking normalisation-layer changes, beating six baselines under 40% attack.
Technology & AI
GxP-Agent hits 100% on clinical trial coding benchmark
GxP-Agent encodes regulatory steps as a graph, turning 0% failure into 100% structural match on a new FDA-pilot clinical trial benchmark.
Technology & AI
Looped LLMs improve multi-step AI tool calling, study finds
Cambridge researchers find looped language models improve at multi-step API tool calling, with adaptive inference offering the best compute-performance
Technology & AI
LeakGauge detects AI context-leakage attacks, AUROC to 0.996
A new arXiv preprint introduces LeakGauge, a detector that flags when LLMs leak confidential context, tested across 11 models with AUROC to 0.996.
Technology & AI
Euclid-Omni: AI for Olympiad geometry with far less compute
Euclid-Omni couples LLMs and vision models with a formal geometry solver to match state-of-the-art on Olympiad proofs using less compute.
Technology & AI
Agentao open-source runtime governs AI agent tool use
New arXiv preprint introduces Agentao, a local-first runtime that separates what LLM agents propose from what they can execute, with open-source code on
Technology & AI
SearchAuditor fixes 32% of AI agent failures, benchmark shows
New benchmark of 1,243 failed agent runs shows even GPT-5.5 auditors fix only 26.6%, with SearchAuditor at 32.3%, a warning for agent teams.
Technology & AI
Chiplet and AI chip-design security threats mapped in new preprint
An arXiv preprint maps security threats across chiplet systems and LLM-driven chip design tools, directly affecting fabless semiconductor teams adopting