13 articles — 🔴 2 critical , 🟡 7 important , 🟢 4 interesting
Apple presented AFM 3 on June 8, 2026, a family of five models (Apple Foundation Models, third generation). AFM 3 Core records 45.6% user preference versus 23.3% for the previous model, while AFM 3 Cloud reaches 64.7% versus 8.7% — a relative gain of 36% in user satisfaction.
Reasoning Arena is a new language model training method based on reinforcement learning with verifiable rewards (RLVR). Instead of the standard group comparison of traces, it introduces head-to-head tournaments and the Bradley-Terry model for estimating relative advantages — eliminating the zero-gradient problem and accelerating training by 27–41%.
TheoremBench is a new benchmark that assesses the capability of large language models in formal theorem proving using Lean4. It covers approximately 100 classical theorems in standard and extended format with automatically extracted premises. Research shows that explicit premises significantly improve performance, while new metrics reveal that models favor easier sub-theorems and generate overly long proofs.
Ultra Flash is a cascaded streaming architecture for video generation that breaks the 480p barrier of autoregressive models for the first time. The system achieves ~30 FPS at 1K and ~18 FPS at 2K resolution on a single GPU, making it the first system for real-time generation of high-quality video.
VirBench is a new benchmark with 120 real-world queries for retrieving viral sequences across 40 pathogens, introduced by Anthropic as the standard for evaluating biological AI agents. Without specialized tools, SOTA models achieved 16.9–91.3% accuracy; with the gget virus tool GPT-5.5 reaches 99.7%.
SearchSwarm is a multi-agent research system that introduces delegation intelligence — the ability to decompose and delegate tasks to sub-agents while respecting context window constraints. The SearchSwarm-30B-A3B model achieves 68.1 on BrowseComp and 73.3 on the Chinese variant, outperforming all models of comparable size.
InquiTree is a new research framework that evaluates LLM agents on 30 real scientific papers using a DAG model of the research process. The study reveals cognitive tunneling — degradation of critical judgment in long sessions — and shows that performance drops on papers published after the model's training cutoff.
Anthropic has released Claude Code v2.1.169, an update to the CLI tool for agentic coding. The release introduces the --safe-mode flag for isolated troubleshooting, a new /cd command for changing the working directory without breaking the prompt cache, and fixes for enterprise MCP policies, UI freezes, and Windows issues.
'Context rot' is the name for AI configuration files such as CLAUDE.md and .cursorrules that reference outdated code, APIs, or paths. New research shows that stale references are present in 23% of analyzed repositories, and teams using them may silently degrade code quality without noticing.
Researchers from the Frontier Red Team reported that Claude Mythos Preview autonomously produces functional exploit code for patched vulnerabilities within hours. The model created 8 working exploits for Firefox SpiderMonkey in 12 hours, and 18 PoC codes for the Windows kernel in 6 hours, the first in just 31 minutes.
HPAA are typographic attacks that use spacing, emphasis, and spatial text layout to hide harmful content from automated detectors while remaining readable to humans. The study accepted at USENIX Security 2026 was tested on 13 systems and achieved >86% human recognition with detection below 1%.
VFUSE is the first feature-level virulence audit for open-weight protein design models (RoseTTAFold3, RFDiffusion3). Authors Michael Yu and Matthew Olson trained Sparse Autoencoders on diffusion-transformer activations and achieved AUROC 0.84 for detecting dangerous protein designs — with no performance loss on legitimate biosynthesis tasks.
RealityTest is a new benchmark from the UK AI Security Institute that tests whether AI systems disclose their identity when users ask them. Built on 3,152 real queries from 750 participants across five languages, the benchmark shows that disclosure rates vary between 8% and 92%, and a single suppression instruction can drop them to just 3%.