Why it is worth attention

It demonstrates a novel 'write-the-paper-first-then-make-it-true' approach that enforces honesty through LaTeX macros and automated audits, and it won first place ($10,000 in OpenAI credits) in a workshop judged by 11 expert reviewers, with the entire paper and all artifacts (except the README) generated by AI.

Who it is for

AI researchers interested in automated scientific discovery and agentic systems
ML engineers designing integrity workflows for autonomous agents
Hackathon participants and organizers seeking proven patterns for auto-research tracks
Educators demonstrating research integrity and reproducibility in AI-driven experiments

Use cases

Automated end-to-end research paper generation with verifiable claims
Designing multi-agent systems with file-based communication and integrity gates
Learning from a transparent run log of 136 timestamped decisions and pivots
Reusing the paper quality guidelines and verification scripts for other projects

Strengths

First-place win at ICML workshop with $10,000 prize, judged by 11 experts
Novel ph/phm macro system that mechanically prevents unfounded claims from shipping
Complete transparency: every decision, result, and pivot is logged in the repository
Demonstrates agentic autonomy with human co-scientist feedback without compromising honesty

Considerations

Not a reusable framework—the harness is specific to this run and requires internal infrastructure (GPU fleet, Overleaf bridge)
The research result (Depth-AR) is a hackathon artifact produced in three hours on small models, not a vetted scientific claim
Credentials and live tokens are redacted; paths are machine-specific and may not work out-of-the-box

README quick start

🏆 1st Place at Ralphthon @ICML "Auto Research": $10,000 in OpenAI Credits

Writing-Driven Autoresearch

By Woomin Song and Byungjun Yoon, Team WooandB (KAIST · ALIN Lab, advised by Prof. Jinwoo Shin) 🙏

This repository is the harness, tooling, and complete run record behind that winning entry: a hackathon where agents produce a research paper end to end, with only light high-level human guidance. The papers were judged in ICML workshop format by a panel of 11 expert reviewers, and ours took first.

Every file in this repository except this README was written by the AI scientists themselves: the plan, the experiment code, the decision ledger, the integrity tooling, and the paper. This README is the only human-curated document here; everything else is the artifact.

"The paper is true because the system was built to make untruth expensive." — the experiment agent's closing reflection, ralph/DECISIONS.md

This is not a framework to install. It is the run itself — secrets redacted, everything else exactly as the agents left it:

The harness — three agent personas (master.md, experiment.md, writing.md) and the paper-quality specs they were built around
The skills — alin-skills/skills/: the only two external skills, imported from our lab's internal skill repository — one for coordinating the multi-agent panes via herdr, one for applying edits to Overleaf via git; every other skill in this repo was invented for this run
The ledger — ralph/: 136 timestamped decisions, 53 result JSONs, gate calls, pivots
The integrity tooling — verify-phm.py, writing-audit.sh: 795 checks green at submission
The paper — writing-main.pdf: "Depth-AR: Skipping Transformer Layers Without Dropping Their Updates", submitted 33 minutes early. The winning paper and its own git history — 45 commits, 39 of them made by the AI — live in the paper repo: [happyhappy-jun/

Description

Multi-agent harness + complete run record of the 1st-place entry at Ralphthon@ICML2026 — three AI agents wrote a workshop paper in 3 hours

Related repositories

Similar projects matched by category, topics, and programming language.

lopopolo

Featured

harness-engineering

Harness Engineering is a methodology for improving coding agent outputs by carefully crafting the environment around them—providing curated context, tools, and executable constraints that encode an organization’s nonfunctional requirements and cumulative lessons.

AI & Machine LearningAI Agents

2,390

slvDev

Featured

esp32-ai

A 28.9 million parameter language model runs on an $8 ESP32-S3 microcontroller entirely on-device, generating simple stories at about 9.5 tokens per second.

AI & Machine LearningLarge Language Models

1,960

littledivy

Featured

mimic

mimic captures traffic from any iOS or web app and automatically generates a Python client library that lets you call the app's API like a regular library.

AI & Machine Learning

1,482