DuoDiscover lets coding agents co-discover an optimization target and a fast, fidelity-checked proxy evaluator, so search is no longer bottlenecked by hours-long true evaluations. It ranks first on 121 of 235 kernels on NVIDIA’s SOL-Exec B200 leaderboard, doubles human-expert speedups on SWE-fficiency, and reaches 1.92× on vLLM serving. Work done at Google.

SWE-Protégé is a post-training framework (SFT + agentic RL) that teaches a small language model to stay the decision-maker on SWE tasks while selectively asking a strong expert model for guidance. It brings Qwen2.5-Coder-7B-Instruct to 42.4% Pass@1 on SWE-bench Verified, +25.4% over the prior SLM state of the art, with Claude Opus 4.1 as the expert. Work done at Meta.

We propose Experiment-as-Code (EaC) Labs, a declarative stack that lets AI agents run experiments in physical autonomous labs by compiling declarative experiment configurations down to instrument-level APIs, with safety checks, resource assignment, and job orchestration in between.

EXP-Bench is the first benchmark to evaluate AI agents on research experiment tasks that are semi-autonomously constructed from top-tier ML research papers.

We propose remote direct code execution (RDX), which extends RDMA from memory access to code execution, enabling an agentless architecture for runtime extensions such as Wasm filters, BPF programs, and UDFs.

We explore the promise of LLM-powered AI agents for cloud infrastructure management and report early takeaways and research challenges from a preliminary study across common cloud interfaces.

Curie is the first AI-agent framework designed for automated and rigorous scientific experimentation. Curie helps answer your curiosity through end-to-end experimentation automation, ensuring that every step—from hypothesis formulation to result interpretation—is conducted with precision, reliability, and reproducibility.

Cloud environments are increasingly managed by Infrastructure-as-Code (IaC) platforms like Terraform, which let developers define infrastructure as configuration code. While IaC automates deployment, its update logic is error-prone, often introducing subtle yet impactful bugs. IaC updates are common because cloud infrastructures are long-lived but user requirements fluctuate over time. Testing updates is challenging due to the vast and evolving search space of infrastructure setups and resources. We introduce TerraFault, an efficient, LLM-guided system for discovering update bugs. Our prototype optimizes search and testing to systematically detect bugs, even in simple updates, improving Cloud reliability.

While effective for greenfield (new) cloud deployments, existing IaC platforms struggle with brownfield migration—translating existing non-IaC infrastructure into IaC programs. This limits Cloud adoption, as current tools rely on error-prone, rule-based reverse engineering. We introduce Lilac, a novel approach that automates IaC lifting by combining LLMs for rule extraction with symbolic methods for correctness assurance. Lilac aims to enable an automated, provider-agnostic lifting tool with broad coverage and high accuracy, streamlining IaC adoption.

While LLMs show potential in general code generation, their efficacy in IaC development remains unknown. To address this, we developed the first dataset and benchmark capable of evaluating IaC code generation. Our dataset comprises 458 human-curated scenarios spanning various AWS services, involving over 1,720 hours of human effort. Our results reveal significant performance gaps.

Zodiac automatically unearths complex cloud IaC semantic checks/rules that state-of-the-art IaC tools cannot easily capture, allowing us to reduce runtime error violations that can take very long to debug, into simple compile time checks.

SpotProxy is a censorship resistance system that uses cost-effective and high-churn cloud instances to maximize the circumvention utility of cloud-hosted proxies.

NetShuffle is a censorship resistance system that offers shuffle proxies, designed to engage a new class of support base–edge networks–which have received scant attention from existing work.

Cloudless Computing makes a case for simplifying cloud infrastructure management, by sinking these cloudy infrastructure management tasks down from the user’s perception and providing them as-a-service, analogous to serverless computing that relieves users of the burden of managing server instances.

Stargaze is a security-centric experimentation platform for Low-earth orbit (LEO) satellite constellations.