GAPRS: Mapping Scientific Knowledge with Claim Dependency Graphs

Introduction

Most scientific papers are underread, peer review capacity is limited, and much innovation never reaches broad audiences. Current methods of scientific discovery are manual and slow.

GAPRS transforms papers into interactive epistemic graphs where both humans and LLMs can reason over claims, assumptions, invalidators, and dependencies, accelerating discovery and surfacing hidden insights.

Section 1: Core Idea

At the heart of GAPRS is the concept of claims as nodes, with edges representing support, dependency, or extrapolation between them.

Here’s an example using Attention Is All You Need (Vaswani et al., 2017):

ClaimStatementConfidence
C1Self-Attention achieves state-of-the-art performance on sequence modeling tasksHigh
C2
Transformer architecture enables more efficient parallelization than RNNs		High
C3
Self-Attention allows modeling of long-range dependencies better than RNNs		Medium
C4
Positional encoding is sufficient to provide sequence order information		Medium

C5
Transformer generalizes beyond translation tasks
Low

ConfidenceMeaning
High
Claim is well-supported, reproducible, and robust. Strong experimental/theoretical evidence exists.
Medium
Claim is plausible but has some untested assumptions, limited replication, or moderate evidence.
Low
Claim is speculative, poorly supported, or highly dependent on unverified assumptions.


Explore the Interactive GAPRS Epistemic Metadata Graph for C1-C5 Here: https://wiknwo.github.io/CT5129_AIProject/GAPRS_C1-C5_interactive_v2.html

Section 2: Research Potential for Scientific Discovery Teams

GAPRS enables new types of analysis that would be of interest to teams like SonyAI's Scientific Discovery Team:

  1. Weakest Claim Detection: Identify claims with the lowest confidence and map their downstream influence across the network. This highlights fragile parts of scientific knowledge.  
  2. Influence Mapping: Quantify which assumptions are most central. This reveals which parts of a field hold the most epistemic weight.
  3. Automated Hypothesis Generation: Using LLMs, GAPRS can suggest experiments or new directions by reasoning over claim dependencies and invalidators.
  4. Cross-Paper Reasoning: Expand DAGs to include multiple papers. Track knowledge evolution and detect inconsistencies or emerging consensus in a field.
  5. Interactive Exploration: Tools like pyvis allow researchers to explore DAGs dynamically, inspecting assumptions, invalidators, and theoretical arguments node-by-node.

Section 3: LLM-Assisted Workflow

  1. Input: JSON of paper metadata + claims
  2. LLM Claim Extraction: Identify assumptions, invalidators, and theoretical arguments
  3. Cross-Claim Verification: Suggest new edges, validate dependencies
  4. Graph Construction: Build interactive pyvis DAG
  5. Re-authoring: Produce GAPRS-native claim JSON
  6. Human Verification: Ensure epistemic rigor

Section 4: Why This Matters

This approach makes scientific knowledge computable, enabling faster discovery, better hypothesis generation, and insight into fragile or highly influential claims.

  • Rigorous claim analysis

  • Automated reasoning with LLMs

  • Human-in-the-loop verification

Section 5: Next Steps

  • Explore multi-paper DAGs for full-field epistemic analysis

  • Expand interactive visualization for team collaboration

  • Encourage discussion with research teams to refine methodology and share insights