Automating AI Safety Research
We build benchmarks, datasets, and infrastructure that help speed up AI safety research.
Our Mission
We are a research group focused on building open infrastructure for AI safety research. Our work is motivated by the belief that AI systems will become increasingly capable and autonomous, and that we need to develop the infrastructure, now, to keep pace with capabilities
What We Build
Benchmarks
We build benchmarks to evaluate how AI agents can contribute to AI safety research.
Datasets
1.1M enriched papers. 129K research repositories. 778K code functions. The raw material for studying how AI systems interact with real scientific work.
Infrastructure
Runtimes for structured agent workloads. Orchestration topologies for recursive improvement loops. The scaffolding to run experiments at scale.
Research
Epsilon: Infrastructure for Structured Agent Workloads
An open-source runtime for structured agent workloads with seven orchestration topologies, ZeroMQ-backed task brokering, deterministic...
S2ORC CS Enriched: 1.1 Million Computer Science Papers with Structured Metadata
A filtered and LLM-enriched version of Allen AI's S2ORC corpus containing 1.1 million computer science papers with structured extraction...
Study Failure: AI-driven GPU Kernel Optimization
A retrospective on 131,520 GPU kernel optimization attempts that were invalidated when agents were found to be substituting high-level...
Learning to Rank Architectures: A Small Model That Guides Neural Architecture Search
A tiny recursive reasoning model trained to rank architectures by predicted performance achieves 8-10x sample efficiency over random...
ARIA Benchmark: How Much Machine Learning Do AI Models Actually Know?
A suite of five closed-book benchmarks probing the ML knowledge that frontier language models have internalized during training.
ArXiv Research Code Dataset: 129K Research Repositories
A collection of 4.7 million code files from 129K research repositories linked to arXiv computer science papers.
ArXivDLInstruct: 778K Research Code Functions for Instruction Tuning
A dataset of 778,152 functions extracted from arXiv-linked research code, each paired with instruction prompts, for training...
DeltaMLBench: Can AI Agents Improve on Published ML Research?
A benchmark of 50 tasks drawn from real Papers With Code repositories where agents must achieve measurable improvement over published baselines.
Teaching Models to Bluff: Measuring Deception, Belief, and Coordination in LLM Secret Hitler
We implemented five LLM agents playing the social-deduction game Secret Hitler with structured logging to quantify deception, belief...
ML Research Benchmark: Can AI Agents Do Real ML Research?
A benchmark suite of 7 competition-level ML challenges for evaluating whether AI agents can perform genuine research iteration beyond...
About
Algorithmic Research Group builds tools and infrastructure for AI safety research. Benchmarks for evaluating autonomous agents. Datasets for studying how models fail. Runtimes for running agent workloads at scale.