The Archive

Windows 11 updates could soon include fixes for more security issues at once. Microsoft said in a blog post on Thursday that it's now using AI to "identify potential issues earlier," which means "customers will see a higher volume of security updates included in each security release." Hackers, even amateurs, have increasingly been using AI to quickly exploit security weaknesses over the past several months. Security researchers are also using AI to find issues faster, leading to more frequent high-severity vulnerabilities, like the "Copy Fail" exploit that impacted nearly every Linux distrib...

About two weeks after OpenAI's GPT-5.6 was caught up in regulatory drama - rolled out only to government-approved organizations during a "limited preview" period - the company has received the Trump administration's green light for a public rollout of the model. OpenAI CEO Sam Altman called it "the best model we have ever produced." To celebrate, OpenAI also unveiled a new AI agent on the same day: ChatGPT Work. It's billed as a combination of ChatGPT and Codex, allowing the everyday non-technical user to take advantage of Codex's capabilities for non-coding tasks, and it's powered by the GPT...

ProjAgent uses procedural similarity retrieval to improve repo-level code generation by matching functional logic across codebases.

Benchmark of training-free relaxed speculative decoding techniques shows speed-capability trade-offs for LLM sampling without retraining.

SolarChain-Eval benchmarks autonomous agents in decentralized energy markets with physics constraints and trustworthiness metrics.

Framework compares test-time resampling vs. rerouting for cost-aware LLM selection under realistic budget constraints.

WebSwarm orchestrates recursive multi-agent LLM web search with adaptive collaboration to improve search depth and evidence coverage.

EdgeRefine applies Jaccard sampling to balance privacy and utility in GNNs under edge differential privacy constraints.

Meta releases Muse Spark 1.1 with API access and improved agentic tool calling and computer use capabilities.

Marketplace simulation with DeepSeek-V3 agents tests formal mechanisms for maintaining trade stability against adversarial defection.

gspDAG-FL proposes decentralized federated learning via gossip consensus without global coordination, replacing central servers.

Multi-modal inverse RL framework for training robust reward functions across diverse environments and feedback modalities.

UltraX uses adaptive programmatic editing to refine pre-training data quality at scale, addressing efficiency and reliability limits of rule/LLM-based approaches as scaling laws diminish.

BiSCo-LLM introduces binary spherical coding for extreme low-bit LLM compression without explicit codebooks, enabling sub-2-bit weight quantization for memory-constrained deployment.

DominoTree combines tree-structured drafting with Domino's GRU-based causal correction for faster LLM speculative decoding via path-dependent token distributions.

Method steers neural network training via interpretable constraints based on partial dependence to enforce explanations faithful to prior domain knowledge.

llm-meta-ai 0.1 released, enabling CLI access to Meta's muse-spark-1.1 model via Simon Willison's llm tool.

llm 0.31.1 patches tool call JSON error in OpenAI Chat Completion endpoints with empty arguments.

Ben Bernanke joins Anthropic's Long-Term Benefit Trust, signaling governance structure for AI safety oversight.

Anthropic launches public Q&A initiative to address community concerns about AI safety and development practices.

Analysis of 2,053 real patient-chatbot conversations reveals wide variation in communication patterns; proposes patient simulator modeling clinical content, emotion, strategy, and style separately.

S²AE (Structured Sparse AutoEncoder) enforces semantic and spatial consistency in sparse autoencoders for vision-language models to learn modality-consistent mechanistic interpretability concepts.

HCC-STAR, a clinically-aligned LLM, reads EMR narratives to output risk-based staging, ranked treatments with rationales, and survival estimates for hepatocellular carcinoma.

MAESTRO prunes Mixture-of-Experts models using Markov-chain approximation of routing transitions, addressing deployment bottlenecks while maintaining inference efficiency.

Federated deep learning framework for cardiovascular disease risk prediction across institutions without sharing patient data, addressing heterogeneous dataset sizes and population characteristics.

Framework for sequential decision-aware experimental design under adversarial uncertainty, handling hidden/weakly-modeled effects that can alter decision optimality.

Spectral analysis of Extreme Learning Machines' numerical stability via pseudoinverse computation and singular value conditioning.

ImputeViz dashboard for interactive missing data diagnosis and imputation method comparison across MICE, Random Forest, XGBoost, kNN.

TreeSHAP-weighted multimodal fusion for emotion/sentiment recognition balances early/late fusion tradeoffs via sample-level expert weighting.

SMetric proposes session-centric LLM scheduling for agentic workloads with 80%+ KV-cache reuse, shifting optimization from latency to tokens-per-second.