SciencePedia: A Scientific Encyclopedia Built on Inverse Knowledge Search and Verifiable Long Chain-of-Thought

Forum topic · ✨步子哥 · 2025-11-13

Summary

SciencePedia is a scientific encyclopedia system developed by 23 researchers from the Institute of Theoretical Physics (CAS), DP Technology, Lanzhou University, and other Chinese institutions, submitted to arXiv in October 2025. It addresses two problems in existing knowledge bases: compressed reasoning and difficult fact verification. The system uses four components: a Socratic agent that decomposes ~200 subject curricula into ~3 million first-principles questions; a long chain-of-thought (LCoT) generation pipeline with cross-model answer verification to suppress hallucinations; a Brainstorm search engine performing Inverse Knowledge Search to trace concepts back to foundational principles and reveal cross-disciplinary links; and a Plato synthesizer that converts verified reasoning chains into accessible, Feynman-style encyclopedia articles. The system currently covers ~200,000 entries across seven domains, 3 million reasoning chains, and 100,000+ practice questions. Evaluations show Plato articles outperform no-retrieval baselines in knowledge density, factual accuracy, and interdisciplinary coverage, with particular value for Chinese-language scientific content.

Overview

SciencePedia is a scientific encyclopedia system designed to fix two core weaknesses of traditional knowledge bases (textbooks, papers, Wikipedia): reasoning processes are compressed (only final conclusions are shown) and facts are hard to verify. This compression hinders verification, suppresses cross-domain knowledge association, and limits interdisciplinary innovation.

The project was developed by 23 researchers including lead authors Yu Li (Lanzhou University / CAS Institute of Theoretical Physics) and Yuan Huang (DP Technology), with corresponding authors Linfeng Zhang (DP Technology), Zhiyuan Yao (Lanzhou University), and Kun Chen (CAS Institute of Theoretical Physics). The paper was first submitted to arXiv on October 30, 2025, revised November 7, spanning 43 pages.

Architecture

The end-to-end system has four components:

1. Socratic Agent — Based on ~200 university course syllabi, it applies Socratic questioning to decompose high-level knowledge points into first-principles questions, generating ~3 million questions grounded in reductionist methodology.

2. Long Chain-of-Thought (LCoT) Generation and Verification — Multiple independent solver models generate LCoTs in parallel; strict cross-validation with answer consensus and prompt sanitization keeps only verifiable content, reducing AI hallucinations.

3. Brainstorm Search Engine — The core innovation: Inverse Knowledge Search. Given a concept, it traces backward to identify underlying concepts, priors, and misconceptions, retrieving diverse first-principles reasoning paths that end at the target concept and revealing cross-disciplinary applications.

4. Plato Synthesizer — Narrates verified reasoning chains into coherent, Feynman-style popular-science encyclopedia articles.

Key Innovations

Inverse Knowledge Search over an LCoT corpus (rather than human-written text), enabling discovery of long-range, cross-domain knowledge connections
Verifiable reasoning chains via multi-model cross-validation
Endpoint-driven, first-principles (reductionist) construction strategy
A reasoning-first corpus compensating for internet text that emphasizes results over derivations

Scale

~200,000 fine-grained entries across seven domains (math, physics, chemistry, biology, engineering, computing, etc.)
3 million scientific reasoning-chain networks
100,000+ practice questions

Evaluation

Across six subjects, Plato articles synthesized from retrieved LCoTs significantly outperformed no-retrieval baselines (judged by an external LLM) in knowledge-point density and factual error rate, and surfaced interdisciplinary knowledge missing from traditional sources. The system also addresses the gap where Chinese Wikipedia has roughly two orders of magnitude fewer entries than English Wikipedia with much lower quality.

A dual-engine update mechanism combines an AI engine that ingests new papers, textbooks, and databases in real time with an expert community of domain scientists who verify and iterate on key entries.

Significance

SciencePedia represents a paradigm shift toward reasoning-oriented, verifiable, and scalable scientific knowledge management, supporting a move from vertical specialized problem-solving toward general-purpose "AGI for Science."