ai-coscientist

name	ai-coscientist
title	Training AI Co-Scientists Using Rubric Rewards
version	0.0.2
engine	skillxiv-v0.0.2-claude-opus-4.6
license	MIT
url	https://arxiv.org/abs/2512.23707
keywords	["research","reinforcement-learning","rubric-rewards","cross-domain"]
description	Train LLMs to generate high-quality research plans via rubric-based RL without requiring experimental verification. Extracts research goals and domain-specific rubrics from scientific papers, uses frozen model as grader with 12-22% relative improvements, achieves human-expert preference 70% of time with strong cross-domain generalization.

Overview

Automated research plan generation using extractable domain knowledge from papers.

Core Technique

Rubric Extraction and Grading:

# Extract from papers
research_rubrics = extract_rubrics_from_papers(papers)

# Grade via frozen model
grade = frozen_model.score(generated_plan, rubrics)

# GRPO training on grade signal

When to Use

Use when: Research automation, domain-specific planning, cross-domain generalization.

References

• Rubric extraction from scientific papers
• Frozen model as grader
• Self-reward GRPO training