GRAID: Synthetic Data Generation with Geometric Constraints and Multi-Agentic Reflection for Harmful Content Detection

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Rad, Melissa Kazemi, Purpura, Alberto, Kumar, Himanshu, Chen, Emily, Sorower, Mohammad Shahed
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866908500092780544
author Rad, Melissa Kazemi
Purpura, Alberto
Kumar, Himanshu
Chen, Emily
Sorower, Mohammad Shahed
author_facet Rad, Melissa Kazemi
Purpura, Alberto
Kumar, Himanshu
Chen, Emily
Sorower, Mohammad Shahed
contents We address the problem of data scarcity in harmful text classification for guardrailing applications and introduce GRAID (Geometric and Reflective AI-Driven Data Augmentation), a novel pipeline that leverages Large Language Models (LLMs) for dataset augmentation. GRAID consists of two stages: (i) generation of geometrically controlled examples using a constrained LLM, and (ii) augmentation through a multi-agentic reflective process that promotes stylistic diversity and uncovers edge cases. This combination enables both reliable coverage of the input space and nuanced exploration of harmful content. Using two benchmark data sets, we demonstrate that augmenting a harmful text classification dataset with GRAID leads to significant improvements in downstream guardrail model performance.
format Preprint
id arxiv_https___arxiv_org_abs_2508_17057
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle GRAID: Synthetic Data Generation with Geometric Constraints and Multi-Agentic Reflection for Harmful Content Detection
Rad, Melissa Kazemi
Purpura, Alberto
Kumar, Himanshu
Chen, Emily
Sorower, Mohammad Shahed
Computation and Language
Cryptography and Security
Machine Learning
We address the problem of data scarcity in harmful text classification for guardrailing applications and introduce GRAID (Geometric and Reflective AI-Driven Data Augmentation), a novel pipeline that leverages Large Language Models (LLMs) for dataset augmentation. GRAID consists of two stages: (i) generation of geometrically controlled examples using a constrained LLM, and (ii) augmentation through a multi-agentic reflective process that promotes stylistic diversity and uncovers edge cases. This combination enables both reliable coverage of the input space and nuanced exploration of harmful content. Using two benchmark data sets, we demonstrate that augmenting a harmful text classification dataset with GRAID leads to significant improvements in downstream guardrail model performance.
title GRAID: Synthetic Data Generation with Geometric Constraints and Multi-Agentic Reflection for Harmful Content Detection
topic Computation and Language
Cryptography and Security
Machine Learning
url https://arxiv.org/abs/2508.17057