UniScientist 30B: Open-Source Research Model Redefines Autonomous AI Science

Model Overview

UniScientist is a 30‑billion‑parameter open‑source model that implements a full research loop – hypothesis generation, evidence collection, reproducible inference, and iterative validation – and claims performance comparable to or exceeding closed‑source models that are an order of magnitude larger.

Data Construction Pipeline

The training data pipeline solves two bottlenecks (high manual annotation cost and low‑quality synthetic data) by using a “large‑model‑generated + human‑expert‑verified” division of labor. The model generates interdisciplinary research questions and draft solutions at scale; human experts perform high‑precision verification. This process has produced more than 4,700 research‑grade instances covering over 50 disciplines and 400+ research directions. Each instance includes 20+ evaluation criteria, and a single expert annotation now takes 1–2 hours.

Technical Architecture

UniScientist focuses on two core operations: active evidence integration and model‑level causality tracing. Three core components implement the research intelligence:

Evolutionary, human‑AI collaborative multidisciplinary synthesis that mass‑produces research problems and multidimensional test standards.

Intelligent‑agent research loop that incorporates academic retrieval, code interpretation, and the cycle “obtain evidence → derive conclusion → update hypothesis”.

Report‑aggregation module that merges multiple research reports to enable self‑evolving quality.

An “evidence state” system classifies evidence as either verifiable or derivable, and atomic check items guarantee reproducibility.

Benchmark Results

On the FrontierScience‑Research leaderboard UniScientist‑30B‑A3B scores 28.3, surpassing Claude Opus 4.5 and GPT‑5.2 xhigh. Using the aggregation mode raises the score to 33.3. After enabling the FrontierScience‑Olympiad tool, the score reaches 71.0, matching Claude Opus 4.5, and the model performs on par with leading closed‑source models on the DeepResearch Bench. Even without tool assistance, the model shows notable gains in retrieval, deduction, verification, and writing across the entire research pipeline.

Open‑Source Release

The model is fully open‑source, supporting local deployment, inference, and result aggregation. External tools require Serper‑type API keys. Complete inference traces on authoritative benchmarks have been released.

Current capabilities focus on reproducible reasoning and simulation; real‑world resource scheduling is not yet covered. Future work aims to extend the system to genuine experimental environments and controlled execution of compute infrastructure, advancing large language models from pure text generation toward concrete problem solving.

GitHub:

https://github.com/UniPat-AI/UniScientist