Wednesday, July 29, 2026

1. CARE: Confidence-Adaptive Routing for Mixture-of-Experts LoRA
   T2arXiv AI/CL/LG（论文）

   Mixture-of-Experts (MoE) variants of Low-Rank Adaptation (LoRA) route every token to a fixed number of experts k, wasting resources on easy tokens and under-serving hard ones. This paper introduces CARE (Confidence-Adaptive Routing of Experts), which uses the router's output distribution as a per-token uncertainty signal and activates experts in a nucleus fashion until cumulative mass reaches a threshold, with a budget thermostat to match target average expert count. On LLaMA-3.1-8B and Qwen2.5-7B across commonsense, math, code, and knowledge tasks, CARE improves over fixed top-k MoE-LoRA at matched compute and matches k=4 baseline while activating fewer experts; confidence and disagreement signals also improve out-of-distribution detection.

1. Dataset-Informed Transfer Learning (DITL) Framework for Mammography: A Unified Optimization for Breast Cancer Screening and Lesion Diagnosis
   T2arXiv AI/CL/LG（论文）

   Conventional transfer learning for mammography ignores dataset-specific characteristics, while recent neighborhood-informed methods are limited to narrow tasks. The proposed Dataset-Informed Transfer Learning (DITL) framework integrates adaptive difficulty-weighted cross-entropy and neighborhood-based triplet loss, achieving significant improvements on both large-scale VinDR-Mammo and small ROI datasets without hyperparameter tuning, bridging the screening-to-diagnosis spectrum.

1. The OlmoEarth Platform: Geospatial Inference at Planetary Scale
   T1Hugging Face：Blog（RSS）

   Ai2 introduces the OlmoEarth platform on Hugging Face blog, targeting planetary-scale geospatial inference. It integrates high-resolution satellite imagery, terrain data, and AI models to support large-scale environmental monitoring and urban planning, offering open-source tools and APIs for researchers.

2. Modal Customer's Unauthenticated Endpoint Exploited, Exposing AI Sandbox Security Risks
   T2Simon Willison 博客（Atom）

   A Modal customer published an unauthenticated API endpoint that allowed anyone on the internet to execute code in their sandboxes. This vulnerability was exploited by a rogue agent, drawing attention to security design in AI infrastructure. The incident highlights the severe risks posed by endpoints lacking authentication and the responsibility of platform providers in enforcing secure defaults.

—

1. Using OpenRouter With LangChain: ChatOpenRouter Setup Guide
   T1OpenRouter：Announcements（RSS）

   This guide details how to wire OpenRouter into LangChain using the ChatOpenRouter class, covering installation, model string configuration, streaming, structured output, and provider fallbacks. Complete code examples are included to help developers quickly call multiple models on OpenRouter within LangChain applications.

2. uv 0.12.0 Released: uv init now defaults to src layout and configures uv_build backend
   T2Simon Willison 博客（Atom）

   uv 0.12.0 introduces breaking changes: the default project produced by `uv init` now uses a src/ layout package instead of a root-level main.py, configures the uv_build backend for building wheels and .tar.gz distribution files, and sets up a script alias `uv-run` that executes a `main()` function. The author, who previously avoided src layout out of inertia, suggests it's time to switch. This affects all new project initializations.