A new arXiv study (2608.21811v1) finds that in sparse-reward reinforcement learning for vision-language math reasoning, prior injection only helps when it actually reaches the policy: on 20,830 visual…
A new system called MemGuard, developed by researchers and posted on arXiv (2608.21867v1), improves LLM-agent memory reliability by treating verifier output as persistent lifecycle metadata, achieving…
Researchers introduced GuardianBench, a benchmark with 3,024 instruction-scene examples designed to test latent contextual risk in embodied AI, finding that state-of-the-art vision-language models ach…
A new arXiv paper (2608.22063v1) proposes the Domain-Oriented Tooling Pattern for MCP servers, where LLM agents select from domain-aligned tools instead of generating SQL at query time. In a benchmark…
Researchers introduced Hack-Verifiable Terminal Bench (HVTB), an adaptation of the hack-verifiable environments (HVE) methodology to Terminal Bench, a benchmark of real-world terminal and coding tasks…
A controlled study of three compact open-weight language models under five billion parameters found that no single model dominates on mathematical reasoning, with Qwen3:4b (Alibaba) most accurate on t…
Researchers from MIT and DARPA introduced an LLM-driven approach that generates dynamic algorithmic dispatch heuristics for high-performance linear algebra, using LLaMA 3 and a curated performance dat…
Embedd Ltd., a physical artificial intelligence chip infrastructure startup, has raised $2.7 million in a pre-seed funding round led by Seedcamp, with participation from Cocoa, Connect Ventures, 2100 …
Microsoft embeds a server-issued GUID as an invisible watermark in AI-generated images from Paint and Photos, linking the image to the user's prompt, according to an analysis by software developer Xus…
Glassbox, a new browser-based tool by solo developer David Dale, reveals how easily browsers can be fingerprinted, estimating identifiability from 99% in a standard Chrome session to 56% in Tor Browse…
Dogotix, the humanoid robotics unit of XPeng Motors, raised more than $900 million in its initial funding round, bringing its post-money valuation to over $6.3 billion. The company plans to use the fu…
Eigen Labs, a Seattle-based tech company, estimates that owners of Apple Silicon hardware can earn $120 to $200 per month by selling idle compute power for AI inference through its Darkbloom network, …
UAW members rejected Deere's contract extension offer, which included a 4% annual wage increase and $3,000 yearly bonuses, as union president Shawn Fain seeks a more lucrative deal amid Deere's AI-dri…
More than 1 million LinkedIn users have clicked the platform's new 'Seems like AI slop' button since its launch at the end of July, and LinkedIn Chief Product Officer Hari Srinivasan said those who co…
Australian Prime Minister Anthony Albanese will use Wednesday's national cabinet meeting to reassure premiers that new national controls on datacentre developments will complement state rules, amid op…
A METR study finds that AI has dramatically accelerated the rate of reported cyber vulnerabilities in 2026 compared with 2025, contributed modestly to mathematics research with arXiv submissions doubl…
Xiaomi unveiled three TSMC-fabbed Xring chips, including the 3nm O3 flagship phone SoC, the 6nm O100 NPU for on-device MiMo LLM, and the 3nm D100 for autonomous driving, aiming to reduce reliance on Q…
Anthropic's flagship models Fable 5, Mythos 5, Opus 5, and Opus 4.8 experienced elevated error rates for roughly three hours starting around midnight ET on August 24, taking down Claude.ai, the API, C…
Claude Code topped the August agent harness rankings, leading on depth of hooks, subagents and dynamic workflows, and is named the default choice for long autonomous coding sessions, according to the …
Cloudflare Inc. released three tools that enable AI agents to act autonomously: Kitesurf, a stateless Rust browser running on Cloudflare Workers; Wallets, which let humans delegate capped spending to …