Research map / Large language models / Instruction and fine-tuning
BARRED: Synthetic Training of Custom Policy Guardrails via Asymmetric Debate
Arnon Mazza, Elad Levi
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| fine-tune a small model on BARRED synthetic data instead of prompting a frontier LLM | Qwen2.5-14B · four guardrail tasks | accuracy | Improved |
| fine-tune 3B model on BARRED data vs generic guardrails | Qwen2.5-3B · four guardrail tasks | accuracy | Improved |
| add debate-based verification vs no verification (ablation) | BARRED · Plan verification (human test) | accuracy: 0.58 -> 0.85 | Improved |
| add debate-based verification vs no verification (ablation) | BARRED · Plan verification (synthetic test) | accuracy: 0.65 -> 0.99 | Improved |
| use multi-agent debate instead of single-agent Self-Refine (ablation) | BARRED · Plan verification (human test) | accuracy: 0.53 -> 0.85 | Improved |
| use Self-Refine instead of no verification (ablation) | BARRED · Plan verification (human test) | accuracy: 0.58 -> 0.53 | Hurt |
| increase number of dimension instantiations (ablation) | BARRED · test set coverage and accuracy | accuracy / coverage | Improved |
| scale fine-tuned Qwen2.5 from 1.5B to 14B on simple task (ablation) | Qwen2.5 fine-tuned · Repetition | accuracy | No clear effect |
| scale fine-tuned Qwen2.5 from 1.5B to 14B on complex tasks (ablation) | Qwen2.5 fine-tuned · Privacy and Health | accuracy | Improved |
| use iterative refinement with judge feedback on rejected samples | BARRED · qualitative examples | sample quality | Improved |
Closest papers by meaning
- LPG: Balancing Efficiency and Policy Reasoning in Latent Policy Guardrails (NeurIPS 2026 · similarity 0.88)
- SafeWatch: An Efficient Safety-Policy Following Video Guardrail Model with Transparent Explanations (ICLR 2025 · similarity 0.87)
- Building a Foundational Guardrail for General Agentic Systems via Synthetic Data (ICLR 2026 · similarity 0.87)
- Towards Policy-Adaptive Image Guardrail: Benchmark and Method (CVPR 2026 · similarity 0.87)
- LiSA: Lifelong Safety Adaptation via Conservative Policy Induction (NeurIPS 2026 · similarity 0.87)
- On-Policy Consistency Training Improves LLM Safety with Minimal Capability Degradation (NeurIPS 2026 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
