New Research: LLM Judges Unstable, RAG Privacy, and More
New arXiv papers cover LLM judge instability (Wiggle Framework), privacy-preserving RAG (SEAG), multi-agent pricing, transformer expressiveness, constraint saturation, Dual-Flow Transformers, governed memory, and edge scheduling. Findings: judges flip verdicts 25-91% under pressure; SEAG hides sensitive entities with >74% accuracy; instruction following collapses beyond 5-6 constraints.
Why this appeared
- Multiple sources
These labels describe how the story was ranked for the feed — not editorial endorsements.
Sources and evidence
- arXiv
- arXiv
- arXiv
- arXiv
- arXiv
- arXiv
- arXiv
- arXiv
- arXiv
- arXiv
- arXiv
- arXiv
- arXiv
- arXiv
- arXiv
- arXiv
- arXiv
- arXiv
- arXiv
- arXiv
- arXiv
- arXiv
- arXiv
- arXiv
- arXiv
- arXiv
- arXiv
- arXiv
- arXiv
- arXiv
- arXiv
- arXiv
- microsoft.com
- arXiv
Summary last validated Aug 15, 2026
Reader actions
Report an issue
Use this for an incorrect summary, wrong source, duplicate story, or wrong category. Submissions are private and do not change the story automatically.
Related coverage
- industry
AI News in 60 Words
A roundup of AI developments: a new arXiv paper on mathematics in the age of AI, a blog explaining Claude's watermark, a Twitter thread on Claude writing a macOS driver, Linear's data on AI usage in software teams, an interactive model architecture tool, a Claude Code pipeline for iOS, a free book on inference engineering, and a hybrid RAG benchmark.
- open source
Mojo programming language goes open source
Mojo, the Python-inspired language for GPU programming, is now open source. The compiler and toolchain were released under an Apache 2 license, following the 1.0 release last week. Originally planned as a Python superset, Mojo is now its own language, prioritizing GPU programming ease.
- research
IBM Research: How Much Memory Does Your Agent Actually Need?
IBM Research introduces ALTK-Evolve-HMM, a method to evolve agent memory architectures using hidden Markov models. The approach automatically designs memory modules, improving performance on long-horizon tasks while reducing memory usage. The work is detailed in a Hugging Face blog post.
- research
Anthropic Explains How Claude's Text Watermark Works
Anthropic published details on how its text watermarking system for Claude operates, describing the technical mechanism used to embed detectable signals in AI-generated text. The blog post outlines the approach behind identifying content produced by Claude models, aiming to increase transparency about AI-generated text provenance.
- research
Hugging Face Reproduces 2,200 ICML Papers, Shares Lessons
Hugging Face published a blog post detailing an effort to reproduce 2,200 papers from ICML, sharing findings from the large-scale reproduction project. The post outlines what the team learned in attempting to verify and replicate results across this broad set of accepted machine learning papers, though specific quantitative findings were not detailed in the available excerpt.