Schema-Guided Scene-Graph Reasoning (SG²)

Citation

Authors: Zhuoqi Chen et al. Year: 2024 Venue: URL:

Abstract

LLMs struggle with spatial reasoning over scene graphs due to distraction by redundant information. This paper presents a multi-agent “Reason-while-Retrieve” framework with schema-guided graph query generation.

Summary

A multi-agent framework for spatial reasoning that uses schema-guided queries to filter relevant information from scene graphs.

Key Contributions

1. Two-module Reasoner/Retriever architecture
2. Schema-guided query generation
3. Reason-while-Retrieve strategy
4. Reduced hallucination through filtering

Core Concepts & Definitions

SG² Framework

• Reasoner: Decomposes task, generates natural language queries
• Retriever: Translates queries into executable graph programs

Scene Graph Schema

Abstract structure that:

• Guides schema-aligned query generation
• Enables structure-aware reasoning
• Provides API for graph database operations

Main Results

1. 12% improvement over ReAct baseline
2. Reduces hallucination by filtering irrelevant data
3. Effective on numerical Q&A and planning tasks

Relevance to Project

Low — Application paper, less relevant to core theory:

• Multi-agent reasoning patterns could inform skill coordination
• Schema concept distantly related to our ontology work
• Graph reasoning applicable if we implement skill graphs

Related Papers

• (Application-focused, less connected to main skill literature)