Skip to content

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

24 Commits
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

🔬 ARA-Demo

Two complete Agent-Native Research Artifacts, produced by AI agents doing real research

ARA Hub arXiv Toolkit

An Agent-Native Research Artifact (ARA) is a research knowledge package whose primary reader is an AI agent, not a human. Instead of a single paper, it stores the research as structured, machine-readable layers: the current best understanding (logic/), the full journey that produced it — dead ends included (trace/), and grounded proof (evidence/). Each demo here also ships a human-facing PAPER.md and a self-contained interactive trajectory.html.


🎮 Demo 1 — arc-agi3/: an agent plays Locksmith and writes its own manual

Interactive trajectory viewer for the ls20 Locksmith run — exploration tree with dead ends on the left, per-step drill-down on the right

Click the image to open the interactive trajectory viewer.

An AI agent (Claude Code / Opus) played the ARC-AGI-3 game ls20 (Locksmith) with no tutorial and cleared all 7 levels. The ARA in arc-agi3/ls20/ was built live, during play — and the agent then used it as a world model to crack the levels it could not solve cold (L4, L6, L7). An ablation on the unseen L7 showed a recipes-only agent never operated a single lock control, while the world-model agent characterized the whole lock: the ARA earns its keep as a world model, not a trajectory cache.


🚀 Demo 2 — nanogpt_ara/: an autonomous optimizer-search speedrun

Interactive trajectory viewer for the nanoGPT optimizer-search speedrun — exploration tree on the left, per-step reasoning, results, and recipes on the right

Click the image to open the interactive trajectory viewer.

An ARA compiled from an autonomous agent's optimizer search on the fixed-architecture modded-nanogpt track_3_optimization benchmark. Starting from the Muon baseline (3500 steps to 3.28 validation loss), the agent produced statistically-validated records at 3205 → 3037 → 2949 steps (v1/v2/v3), plus a documented terminal negative result from a hard-isolated novelty-constrained wave.


🔍 How to explore

Browse everything online on the ARA Hub page for this repo — the tables above link straight to the live viewers.

Everything is also static and self-contained — you can clone the repo and open the same HTML files directly in a browser, or read the ARAs as plain Markdown: start with a demo's PAPER.md, then dig into logic/ (what is known), trace/ (how it was learned, dead ends and all), and evidence/ (the proof).

🧬 Anatomy of an ARA

Layer What it holds
PAPER.md Human-facing summary of the artifact
logic/ Current best understanding: claims, concepts, problem, solution recipes
trace/ Append-only research journey: exploration tree, session records, raw notes
evidence/ Grounded proof: verbatim numbers, figures, tables
trajectory.html Self-contained interactive visualization of the whole process

🔗 Related ARA-Labs resources

Releases

Packages

Contributors

Languages