DART: Debating agentic RESEARCH ToOL

Aim: A multi-agent system that uses lightweight, open‑source LLMs to process research papers, interactively answer questions, generate hypotheses, and design research proposals.

Potential impact: This system streamlines literature review by extracting key insights and enabling targeted, interactive querying. It enhances rigor through structured multi‑agent debates and generates novel hypotheses and research plans grounded in evidence.

Methodology: Raw PDFs are processed into overlapping text segments and embedded using BioBERT for vector indexing. Compact, open‑source models, including Mistral‑7B (Mistral AI), Phi‑2 (Microsoft) were orchestrated to run a RAG‑powered pipeline capable of answering questions with source‑backed reasoning, staging multi‑round agent debates to improve answer quality, and proposing testable hypotheses with drafted experimental designs.

The system preserved short‑term conversational context to support follow‑up queries and extracted figures and tables using OCR and Vision Transformers to enrich understanding and improve output relevance.

Previous work: Our earlier single‑round RAG prototypes showed high answer accuracy in blind tests, with a GPU‑optimized pipeline reducing query latency by over 50% in lab benchmarks.

Current progress: The PDF processing and embedding pipeline has been built and validated. Q&A and debate modules have been successfully demonstrated in internal trials, and the hypothesis‑generation prototype is already producing plausible experimental outlines.

Next Steps: Next steps include benchmarking the hypothesis module, adding human feedback and debate scoring, and integrating larger reasoning models. A domain expert pilot is planned for Q3 2025 to assess technical performance.