RXed AI News

MUSE-Autoskill introduces self-evolving LLM agents with reusable skills via creation, memory, management, and evaluation. https://arxiv.org/abs/2605.27366v1

Alibaba's Qwen team releases Qwen3.7-Max, an AI model that autonomously optimized chip code for 35 hours. https://the-decoder.com/alibabas-latest-ai-model-ran-autonomously-for-35-hours-to-optimize-code-for-its-own-custom-chip/

AI agents improve by reusing structured procedural skills, study highlights domain-level and model-generated artifacts. https://arxiv.org/abs/2605.23899v1

Microsoft Copilot invents country differences in identical datasets when model selection defaults are used. https://the-decoder.com/why-you-shouldnt-leave-model-selection-on-default-in-copilot-gemini-and-other-ai-tools/Answer: Answer:

China requires top AI researchers at private firms like Alibaba and DeepSeek to seek official approval before overseas travel. https://the-decoder.com/china-reportedly-now-requires-top-ai-researchers-to-get-permission-before-leaving-the-country/

Researchers reveal RLHF alignment methods can be exploited to optimize misaligned biases via alignment tampering https://arxiv.org/abs/2605.27355v1

Benchmark compares vision LLMs vs OCR on long, image-heavy PDFs via MMLongBench-Doc. https://www.reddit.com/r/MachineLearning/comments/1tm0cqg/visioncapable_llms_vs_ocr_for_longdocument/

Vector Policy Optimization shows training diversity improves AlphaEvolve search performance. https://arxiv.org/abs/2605.22817v1

ML practitioners grapple with hyperparameter selection in non-monotonic loss SSL, relying on BYOL/JEPA/data2vec with unclear efficacy. https://www.reddit.com/r/MachineLearning/comments/1tmprdm/how_do_ml_practitioners_select_hyperparameters/