DYNAMIC-AUDIT is a novel multi-agent framework powered by local LLMs (Qwen2.5-7B) that features:
- Runtime dynamic agent composition via a meta-agent (2–6 agents, linear/parallel/tree structures)
- Explicit post-hoc auditing layer with dual independent verifiers, consensus adjudication, tri-level confidence (High/Medium/Low), and traceable audit trails
- Grounded generation using Serper Google Search API
- Designed for factual reliability and hallucination mitigation
This repository contains the core framework implementation as described in the paper:
DYNAMIC-AUDIT: A Self-Organizing Multi-Agent Framework with an Explicit Auditing Mechanism
Hao Wang, Li Zhu, Lihua Tian, Tao Xie
Expert Systems with Applications (ESWA), 2026 (submitted/under review)
- Meta-agent dynamically decides agent count (2–6), roles (12 specializations), and topology
- Supports linear, parallel, and tree-structured collaboration
- Explicit auditing with parallel verification + mandatory consensus
- Full audit trails with claim-level corrections and confidence scoring
- Local LLM inference via Ollama (no cloud dependency for core logic)
- GPU with ≥8GB VRAM recommended (RTX 4060 8GB or better)
- 16GB+ RAM
- Python 3.10 or 3.11
- Ollama running locally with Qwen2.5:7b-instruct-q4_K_M model pulled
- Serper API key is intentionally set to 'your-own-key' as a placeholder.
- Replace it with your own key (get free from https://serper.dev).
- Recommended: use environment variable instead: export SERPER_API_KEY='your-real-key'
- Never commit your real key to Git!