The Archive

Defense-as-Skill: runtime guards implemented as inspectable skills to protect skill-augmented agents from malicious skill-based exfiltration and steering.

Harness-of-Harness: framework for multi-day autonomous software development with continual improvement via iterative planning-coding-testing loops.

Vercel AI SDK, Astro, Flue, and tldraw replace community PRs with agent-based software factories for managing large contributor bases.

Live trace model: append-only ledger compiled into typed run state and per-consumer views, reducing monitoring token cost 14–15× vs. full traces.

John Deere is testing a new "JD" AI assistant that it says can help farmers make more money, with answers about best practices and historical trends that are based on their own data. It uses their "field, machine and operational data" to answer questions on topics like equipment settings, fuel usage, or harvest timing. The press release didn't mention which AI technology is behind this platform. But after fighting with farmers and the FTC for years over the right to repair, the news of Early Access Program testing for this bot comes with a mention of its ten-point Farmer Data Commitment sayin...

Training-free reranker for composed image retrieval using metadata-based concept alignment and filtering.

Google launches Pics, an image creation/editing tool built on Nano Banana model, integrated into Google Workspace.

Google wants Workspace users to edit and generate their business imagery with Pics. | Image: Google Google has a new suite of creative design tools for Workspace users called Google Pics, which aims to make editing and generating "professional-grade" AI images less cumbersome for businesses. Built around Gemini and the Nano Banana generative AI model, Google Pics is designed to give more granular control over prompt-based image making and manipulation, allowing users to tap on specific objects or text, and describe what they want to change - features that may avoid the ugly results that chatb...

Fisher geometry analysis reveals safety alignment fragility in LLMs under benign fine-tuning via selective output-side MLP resharpening.

Empirical study comparing expert and imitation-learned dexterous manipulation policies across task execution speeds.

Timestep-free diffusion denoiser with persistent hidden state enables anytime reasoning, achieving 99.90% on Sudoku-Extreme.

Amazon is adding a new Alexa-powered feature called “Update Me When” that can send personalized alerts about product launches, tours, books, shows, and other events that could trigger a purchase.

Edge-girth structural descriptor for GNNs addresses Weisfeiler-Leman limitations without precommitting to subgraph size.

AIR's platform can discover agents running at a company, continuously vets any skills and add-ons they use, and blocks any unwanted behaviour.

HarnessDev benchmark evaluates LLM agents on infrastructure generation, shifting evaluation from task outputs to runnable code.

Masked-citation study shows LLMs cite less critically and with weaker rhetorical intent than human-generated scientific citations.

Power-Law Entropy Search (PLES) uses multi-fidelity Bayesian optimization to efficiently estimate optimal LLM hyperparameter scaling laws.

TREVIS: Tree Transformer VAE approach jointly optimizes decision tree predictive performance and structural sparsity.

TRIAGE three-level routing framework reduces LLM agent token consumption by reusing historical trajectories as reusable skills.

Vision Transformer method for cancer grading combines nuclei classification with histopathology images via semantic-guided multimodal preprocessing.

CATeye framework detects e-commerce voucher fraud by learning invariances across attribute-topology distribution shifts without retraining.

Self-Routing post-training framework routes LLM samples to GRPO, distillation, or regularization based on rollout correctness without external teachers.

Provably safe sim-to-real RL transfer method addresses simulator-reality mismatch with safety guarantees for robotics and healthcare applications.

Physics-Informed Neural Networks applied to Ground-Penetrating Radar data prediction for subsurface anomaly forecasting.

Wasserstein-based analysis of generative data augmentation shows risk distortion controlled by augmentation strength and class-conditional distribution discrepancy.

datasette-mcp 0.2 releases with structured SQL result format (objects instead of arrays) to improve weaker model reasoning over tabular data.

DaEdiTikZ dataset leverages scientific revision trajectories to train VLMs for iterative scientific figure editing without synthetic supervision.

NeuSOGA framework extracts explicit symbolic mathematical representations from geometric observations using neuro-symbolic learning.

Contribution-aware bandwidth allocation optimizes split learning for multimodal edge training by allocating uplink budget based on modality importance.

DroneCATS benchmark evaluates MLLMs as generalist vision-language-action agents for drone control with full action-space prompting.