CLAUSE: Agentic Neuro-Symbolic Knowledge Graph Reasoning via Dynamic Learnable Context Engineering

2025-09-25

AI Summary

To address the challenge of balancing accuracy, latency, and cost in context construction for multi-hop question answering over knowledge graphs, this paper proposes a dynamic, learnable context engineering framework that models subgraph expansion as a sequential decision process under resource budget constraints. We introduce a novel tri-agent collaborative architecture—Subgraph Architect, Path Navigator, and Context Curator—that explicitly models query-level latency and prompt cost as configurable budgets. Furthermore, we propose LC-MAPPO, a resource-aware reinforcement learning algorithm enabling end-to-end, neural-symbolic, budget-constrained reasoning. Extensive evaluation on HotpotQA, MetaQA, and FactKG demonstrates significant improvements: on MetaQA-2hop, EM@1 increases by 39.3 points over GraphRAG, latency decreases by 18.6%, and edge growth reduces by 40.9%. Our approach markedly enhances context compactness, traceability, and performance predictability.

Technology Category

• Search and Optimization: Metareasoning and Metaheuristics
• Knowledge Representation and Reasoning: Computational Complexity of Reasoning
• Natural Language Processing: Question Answering

Application Category

• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

Knowledge graphs provide structured context for multi-hop question answering, but deployed systems must balance answer accuracy with strict latency and cost targets while preserving provenance. Static k-hop expansions and "think-longer" prompting often over-retrieve, inflate context, and yield unpredictable runtime. We introduce CLAUSE, an agentic three-agent neuro-symbolic framework that treats context construction as a sequential decision process over knowledge graphs, deciding what to expand, which paths to follow or backtrack, what evidence to keep, and when to stop. Latency (interaction steps) and prompt cost (selected tokens) are exposed as user-specified budgets or prices, allowing per-query adaptation to trade-offs among accuracy, latency, and cost without retraining. CLAUSE employs the proposed Lagrangian-Constrained Multi-Agent Proximal Policy Optimization (LC-MAPPO) algorithm to coordinate three agents: Subgraph Architect, Path Navigator, and Context Curator, so that subgraph construction, reasoning-path discovery, and evidence selection are jointly optimized under per-query resource budgets on edge edits, interaction steps, and selected tokens. Across HotpotQA, MetaQA, and FactKG, CLAUSE yields higher EM@1 while reducing subgraph growth and end-to-end latency at equal or lower token budgets. On MetaQA-2-hop, relative to the strongest RAG baseline (GraphRAG), CLAUSE achieves +39.3 EM@1 with 18.6% lower latency and 40.9% lower edge growth. The resulting contexts are compact, provenance-preserving, and deliver predictable performance under deployment constraints.

Problem

• Optimizing knowledge graph reasoning under strict latency and cost constraints
• Reducing context over-retrieval in multi-hop question answering systems
• Balancing answer accuracy with predictable runtime performance

Innovation

• Agentic neuro-symbolic framework for dynamic context construction
• Lagrangian-Constrained Multi-Agent PPO algorithm for coordination
• Three-agent system optimizing subgraph construction under resource budgets