
Axiom is an advanced, multi-agent AI pipeline designed to automate academic systematic literature reviews following the rigorous PRISMA 2020 framework. Developed for the AMD MI300X hardware environment, its primary purpose is to accelerate evidence synthesis for researchers by autonomously searching, screening, extracting, and analyzing scientific papers to produce citation-ready APA-7 narrative drafts and executive reports. System Logic & Architecture Orchestrated by LangGraph, Axiom operates through a specialized multi-agent workflow: Searcher: Decomposes natural language questions into API-specific queries (PubMed, OpenAlex, arXiv, Scielo, Crossref). It asynchronously fetches papers, deduplicates them, and verifies Open Access availability. Screener: Employs a cascading evaluation. A fine-tuned 7B model makes the initial pass; uncertain or low-confidence cases escalate to a 32B reasoning model, creating an automated inter-rater reliability workflow. Extractor: Parses PDFs using PyMuPDF and strictly enforces structured JSON data extraction (variables, sample sizes, outcomes) using Pydantic, ensuring zero hallucinated schemas. Analysts & Reconciler: Parallel branches (7B and 32B) cluster findings semantically. A deterministic reconciler crosses these insights without LLM hallucinations to map scientific consensus versus contradiction. Gap Finder & Writer: Identifies unaddressed research gaps and drafts a fully structured academic review. Technologies Axiom leverages an optimized dual-model stack served via vLLM on ROCm. It uses a fine-tuned Qwen2.5-7B-Instruct for high-speed, structured tasks, and QwQ-32B for deep reasoning. The infrastructure integrates LangGraph for state management, Instructor for reliable JSON validation, BGE-M3/ChromaDB for multilingual semantic clustering, and Streamlit for the frontend.
10 May 2026