CLAUSE: Agentic Neuro-Symbolic Knowledge Graph Reasoning via Dynamic Learnable Context Engineering
Summary: arXiv:2509.21035v2 Announce Type: replace
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.
Key Features of CLAUSE
CLAUSE employs a novel approach to optimize the balance between accuracy, latency, and cost. Key features include:
- User-Specified Budgets: Latency (interaction steps) and prompt cost (selected tokens) are exposed as user-specified budgets, allowing dynamic adaptation for each query.
- Three-Agent Coordination: The framework coordinates three agents—Subgraph Architect, Path Navigator, and Context Curator—to optimize subgraph construction, reasoning-path discovery, and evidence selection.
- Lagrangian-Constrained Multi-Agent Proximal Policy Optimization (LC-MAPPO): This algorithm facilitates joint optimization under resource constraints, ensuring efficient processing.
Performance Evaluation
CLAUSE has been rigorously evaluated across several benchmark datasets, including HotpotQA, MetaQA, and FactKG. The results demonstrate its effectiveness in improving performance metrics:
- Higher EM@1: CLAUSE yields higher Exact Match at one (EM@1) scores while maintaining lower latency and subgraph growth compared to static methods.
- Reduced Latency: On the MetaQA-2-hop task, CLAUSE achieves a remarkable +39.3 EM@1 with 18.6% lower latency.
- Compact Contexts: The contexts produced by CLAUSE are not only compact but also preserve provenance, which is critical for transparency in AI systems.
Conclusion
The introduction of CLAUSE marks a significant advancement in the field of knowledge graph reasoning for multi-hop question answering. By treating context construction as a dynamic decision-making process, CLAUSE not only enhances answer accuracy but also ensures that systems remain efficient and cost-effective. Its unique approach allows for adaptable queries without the need for retraining, making it a valuable tool for real-world applications where resource constraints are paramount. The promising results across various datasets highlight the potential of CLAUSE to set new standards in the AI landscape.