All news

Shows much apparent rare-word TTS error is really ASR verifier error.

Audits the hand-written deny sets and registries behind learned access control.

Text-query-guided video summarization with a self-attentive conv network.

Language-guided state representations improve policy learning in world models.

Trains small LMs to select tools across corpora of over ten thousand tools.

Neural transistor models speed up SPICE and cell characterization.

Multi-token prediction and alternative tokenization help LMs learn latent graph shortest-path search.

Reward models for when LLM assistants should use memory.

Homomorphic encryption built on coset codes.

Arista patches an exploited zero-day in VeloCloud Orchestrator.

EcoxPU orchestrates disaggregated LLM inference across heterogeneous accelerators with agents to cut energy use.

Description of the Belarusian BelT block cipher.

Live tracker of currently exploited vulnerabilities.

Neural approximation for reachability of vision-controlled systems.

Microsoft says China-aligned attackers used AI agents and stolen ConnectWise credentials to hit 2,250 cloud tenants.

A network flow scheduler for AI clusters that penalizes contention so latency-sensitive flows meet deadlines alongside bulk training traffic.

Cost-aware abstention for LLM IT ticket routing.

Evidence-aware retrieval evaluation for a clinical RAG system.

A Stackelberg game plus diffusion-based auction for offloading agentic AI tasks in IoT.

LLM-aided RTL timing optimization via logic metamorphosis.

Survey of low-precision training for LLMs, including FP8 and lower formats.

More metadata can mislead LLM museum descriptions.

Engineering framework for multi-cloud cost optimization focused on hidden egress from service placement.

Mix-attention fuses multimodal LLM features into sequential recommendation.

Geometry-aware feature lifting for RGB-D object pose.

Continual deep RL on dynamic graphs for UAV-assisted multi-user 360 VR video streaming.

DWI speeds up on-device LLM decoding by picking the width of a nested model per step.

Distills reliable recommendation rationales from LLMs for conversational recommenders.

Full bypass of eMudhra one-factor OTP login.

Michigan professor builds Jac, a Python-like language designed for AI agents.