All news

IBM releases granite-speech-5.0-2b and updated smaller speech models for transcription, translation and speaker attribution.

Writes continual updates forward-only into an LM native n-gram memory table.

NVIDIA Kumo Relational is a pretrained relational foundation model that predicts churn or fraud directly from database tables with no training.

How a city builder game was made to run on a 16 MHz CPU.

Stochastic-computing hardware accelerator for RoPE in LLMs.

Combined popularity and positivity bias makes recommenders unfair.

Strategies to handle silent data errors in workflows of moldable parallel tasks.

Controlled study of learned unmasking policies in diffusion LMs.

Studies compositionality in RAG.

Causal evidence of hidden latent communication between diffusion LM agents.

LLM-enhanced GNNs for CTR prediction.

Researchers remotely disabled a vulnerable updater in an abandoned Popcorn Time fork that exposed millions of devices to RCE.

Proposes serving RAG workloads through decentralized collaboration of edge devices instead of the cloud.

Deep reinforcement learning task scheduling for mobile edge computing that accounts for providers overselling capacity.

Uses curiosity-driven reinforcement learning to schedule requests collaboratively across mobile edge and cloud servers.

Argues software is slow because of design choices, not language, so AI rewrites in fast languages will not help most programs.

C++26 ballot was disapproved with 12 of 20 P-members in favor, short of the required threshold.

LitPivot helps researchers adjust ideas by placing them in the related literature as they read.

PingCAP, maker of TiDB, confirmed an extortion group stole source code and customer data from internal systems.

Looped uniform diffusion language models.

Proposes a compatibility ratio metric to guide hardware and DNN co-design for embedded accelerators.

Cloudflare WAF now inspects each request together with its response to spot successful exploitation, and adds response rules.

Schedules dependent tasks and manages container image layers together across device, edge and cloud tiers.

HyperDrone quadrotor simulator at millions of frames per second.

Disaster-recovery analytics where an AI agent generates constrained SQL while the query engine stays the same.

Recursive self-improvement for forecasting in open worlds.

Diffusion LMs find natural-looking black-box attack prompts that trigger overthinking.

A step-by-step guide to building Java ME feature phone apps with OpenJDK 8, MicroEmulator stub JARs and ProGuard for preverification, without the old Sun SDKs.

Speeds up out-of-core graph analytics on SSDs by preprocessing graphs based on which vertices stay active.

User simulators that stay on task and sound natural for multi-turn evaluation.