Share story
AI Science

MemMux preprint reports memory controls for parallel coding agents

Researchers report in an arXiv preprint that MemMux, a local runtime for parallel coding agents, kept a Linux test fleet within a 7.5 GiB memory budget with zero swap. It admitted only a subset of agents and reclaimed resources under pressure; ungoverned tools ran every agent and spilled about 2 GiB into swap. MemMux reclaimed 100% of a terminated agent's process subtree and detected 10 of 10 escaped child processes. The researchers released the engine, benchmark and a one-command reproducer. Its memory-attribution scan ran at 1 Hz and consumed 2.7% CPU with ten agents, exceeding their 2% target.
Sources
cs.AI updates on arXiv.org
Published by Tech & Business, a media brand covering technology and business. This story was sourced from cs.AI updates on arXiv.org and reviewed by the T&B editorial agent team.
Back to Newswire
Keep reading
Full wire
AI Science
AI Science

Kurate preprint demonstrates evidence-quality scoring across 4,347 papers

Researchers report in an arXiv preprint that Kurate, a system using large language models, assessed study quality across 4,347 papers, including 3,913 reporting randomized trials. It scores eight dimensions of study design and rep...

cs.CL updates on arXiv.org
Robotics Science
Robotics Science

MIT flight planner avoids moving obstacles in 12 UAV tests

MIT researchers developed a flight planner that avoided all dynamic obstacles in 12 real UAV test flights, using onboard computing and sensors to replan trajectories. Called SANDO, the system also reached its goal faster than seve...

MIT News
AI Science
AI Science

SchemaFill preprint reports up to 4.05× throughput for AI tool calls

Researchers report that SchemaFill improved end-to-end throughput by up to 4.05× over token-by-token generation on the Glaive and BFCL benchmarks. The preprint presents a method for generating structured calls that let language mo...

cs.LG updates on arXiv.org
AI Science
AI Science

OpenAI publishes 722 math manuscripts from an unreleased model

OpenAI said it released 722 mathematical manuscripts organized into 372 families in a public GitHub repository on October 6. The release adds proof artifacts and 10 abridged reasoning summaries to results produced by an unreleased...

Unite.AI