BARRED: Synthetic Training of Custom Policy Guardrails via Asymmetric Debate

Arnon Mazza, Elad Levi · 2026

Read BARRED: Synthetic Training of Custom Policy Guardrails via Asymmetric Debate by Arnon Mazza, Elad Levi (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-25203.txt.

Introduction

Mazza and Levi present BARRED, a framework for making synthetic training data for custom policy guardrails from a task description and a few unlabeled examples. Generic safety models miss task requirements, prompting gives uneven boundary-case performance at high cost, and custom classifiers need costly labeled data. BARRED splits the domain into dimensions for full coverage and uses multi-agent debate to verify label correctness. Small models fine-tuned on this data beat proprietary and dedicated guardrail models. Ablations confirm both dimension splitting and debate verification matter.

Catalog record

CREATORArnon Mazza, Elad Levi
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE46387
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2604.25203

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

BARRED: Synthetic Training of Custom Policy Guardrails via Asymmetric Debate Arnon Mazza, Elad Levi Abstract Deploying guardrails for custom policies remains challenging, as generic safety models fail to capture task-specific requirements, while prompting LLMs suffers from inconsistent boundary-case performance and high inference costs. Training custom classifiers achieves both accuracy and efficiency, yet demands substantial labeled data that is costly to obtain. We present BARRED (Boundary Alignment Refinement through REflection and Debate), a framework for generating faithful and diverse synthetic training data using only a task description and a small set of unlabeled examples. Our approach decomposes the domain space into dimensions to ensure comprehensive coverage, and employs multi-agent debate to verify label correctness, yielding a high-fidelity training corpus. Experiments across diverse custom policies demonstrate that small language models finetuned on our synthetic data consistently outperform state-of-the-art proprietary LLMs (including reasoning models) and dedicated guardrail models. Ablation studies confirm that both dimension decomposition and debate-based verification are critical for ensuring the diversity and label fidelity required for effective fine-tuning. The BARRED framework eliminates the reliance on extensive human annotation, offering a scalable solution for accurate custom guardrails. † † affiliationnotice: Equal contribution Figure 1 : Overview of BARRED . The pipeline consists of four stages: (1) decomposing the task into task-relevant dimensions based on the task description and few-shot samples, (2) sampling dimension instantiations and target labels, (3) generating boundary-challenging samples with reasoning, and (4) validating samples through multi-agent debate. Accepted samples are added to the training set; rejected samples undergo iterative refinement. 1 Introduction Large Language Models (LLMs) are increasingly deployed in high-stakes, user-facing applications spanning customer service, healthcare, finance, and education ( 4 ; 32 ) . Ensuring that these systems adhere to safety constraints and application-specific policies is critical, yet fundamentally challenging. The core difficulty lies in the context-dependent nature of safety: a response that is benign in one deployment setting may constitute a critical failure in another ( 10 ; 36 ; 12 ) . Guardrail models, specialized classifiers that moderate LLM outputs for harmful or policy-violating content, have emerged as the dominant approach to this challenge ( 14 ; 9 ; 8 ; 34 ; 20 ) . High accuracy is critical given deployment stakes; low latency is equally essential as guardrails gate every user interaction. Existing approaches present a fundamental trade-off between these requirements. Static guardrail models achieve strong accuracy on predefined harm categories through task-specific fine-tuning, but cannot adapt to novel policies without costly retraining ( 14 ; 9 ; 34 ) . Dynamic guardrail models offer flexibility by conditioning on arbitrary policies at runtime, but require larger models with increased latency and yield suboptimal accuracy compared to task-specific training ( 10 ; 36 ; 20 ) . Synthetic data generation is an approach that has gained significant attention for training and aligning LLMs ( 30 ; 29 ; 1 ; 28 ) . While synthetic data has been applied to guardrail training, existing approaches use it for augmentation by mixing synthetic samples with curated safety datasets ( 9 ; 10 ; 26 ) . A challenge still remains on how to train a task-specific guardrail purely from synthetic data, provided with only a policy definition and few-shot unlabeled examples. In this paper, we introduce BARRED ( B oundary A lignment R efinement through RE flection and D ebate), a framework that leverages these minimal inputs to generate high-fidelity synthetic data and train fully customized guardrail models. Purely synthetic guardrail training is constrained by two fundamental challenges: diversity and faithfulness. Synthetic datasets frequently suffer from model collapse, failing to capture the full variance of the task domain ( 27 ) , while LLM-generated labels often contain significant noise that degrades downstream model performance ( 13 ) . Our framework addresses these issues directly. To ensure diversity, we employ dimension decomposition that systematically explores the challenging regions of the task domain space. To ensure reliable ground-truth labels, we introduce a multi-agent debate mechanism ( 5 ) that resolves ambiguous cases through structured deliberation. The resulting pipeline produces task-specific guardrails that combine the accuracy of fine-tuned models with rapid adaptation to novel policies. We study the effectiveness of the BARRED framework by evaluating it on three tasks spanning conversational policy enforcement, agentic output verification, and regulatory compliance. Experiments demonstrate that despite relying entirely on synthetic data, the resulting compact models deliver superior accuracy compared to both specialized guardrails and frontier-class LLMs with orders of magnitude more parameters. Ablation studies indicate that dimension decomposition and debate-based verification are critical for ensuring the label fidelity and domain coverage required for this performance. Ultimately, BARRED establishes a generalizable paradigm for low-resource classification, applicable to any domain requiring rapid, high-accuracy customization from minimal user input. Our code is publicly available 1 1 1 https://github.com/plurai-ai/BARRED . 2 Related Work Guardrail models. The deployment of LLMs in production systems has driven significant research into content moderation and policy enforcement mechanisms. Static guardrail models such as LlamaGuard ( 14 ) , WildGuard ( 9 ) , ShieldGemma ( 34 ) , and Aegis ( 8 ) fine-tune compact language models on curated safety taxonomies, achieving strong performance on predefined harm categories while maintaining low inference latency. These approaches typically train on mixtures of human-annotated and synthetically-augmented data to cover diverse attack vectors including adversarial jailbreaks ( 9 ) . However, static models cannot generalize to novel, user-defined policies without retraining. Dynamic guardrail approaches address this limitation by conditioning on arbitrary policies at inference time. DynaGuard ( 10 ) trains models to follow natural language policy descriptions through a combination of traditional safety data and synthetic policy-violation pairs. CoSAlign ( 36 ) enables inference-time adaptation to diverse safety requirements without retraining. Reasoning-enhanced approaches such as GuardReasoner ( 20 ) and R2-Guard ( 15 ) incorporate chain-of-thought traces to improve detection of nuanced violations. Despite these advances, dynamic approaches sacrifice accuracy compared to task-specific fine-tuning and require substantially larger models to achieve sufficient reasoning capabilities for handling arbitrary policies at inference time, resulting in increased computational cost and latency. These limitations motivate our work on rapid policy-specific model customization. Synthetic data generation. LLM-based synthetic data generation has emerged as a powerful approach for training and alignment. Foundational work on Constitutional AI ( 1 ) demonstrated that models can generate high-quality alignment data through critique-and-revise loops guided by natural language principles. Self-Instruct ( 30 ) and Alpaca ( 29 ) showed that instruction-following capabilities can be bootstrapped from minimal seed data. For safety applications, AART ( 25 ) introduced parametrized recipes for diverse red-teaming data, while SAGE-RT ( 17 ) proposed taxonomy-guided expansion to ensure coverage across violation modes. A key challenge in synthetic data generation is maintaining diversity; naive LLM generation suffers from mode collapse, where outputs cluster around typical responses ( 24 ) . Several approaches address this limitation: Persona Hub ( 7 ) leverages billion-scale persona collections to inject diverse perspectives, IntellAgent ( 19 ) employs policy-driven graph modeling to systematically explore scenario spaces. Verbalized Sampling ( 35 ) offers a principled solution by prompting models to generate probability distributions over responses rather than single outputs. Our work adopts this principle: we decompose the generation task along task-relevant dimensions and apply distribution-based sampling

Read the complete text