TLDR: CLAUSE is a new agentic neuro-symbolic framework that improves multi-hop knowledge graph question answering by treating context construction as a sequential decision process. It uses three coordinated agents (Subgraph Architect, Path Navigator, Context Curator) and a Lagrangian-Constrained Multi-Agent Proximal Policy Optimization (LC-MAPPO) algorithm to dynamically build compact, relevant contexts under user-specified budgets for graph edits, interaction steps, and selected tokens. This results in higher accuracy, lower latency, and reduced resource consumption compared to existing methods, offering predictable performance without retraining.
In the rapidly evolving landscape of artificial intelligence, large language models (LLMs) are becoming increasingly powerful, especially when paired with external knowledge sources like knowledge graphs (KGs). KGs provide structured context, which is crucial for answering complex questions that require multiple steps of reasoning and clear evidence. However, current systems often struggle to balance accuracy with practical deployment constraints such as speed and cost.
Traditional methods for using knowledge graphs in question answering, like expanding a fixed number of ‘hops’ or using lengthy ‘chain-of-thought’ prompts, frequently lead to problems. They can retrieve too much irrelevant information, inflate the amount of text an LLM needs to process (increasing cost), and result in unpredictable response times. This often means that systems are not only constrained by how much text they can handle but also by the number of interaction steps involved in finding an answer.
A new research paper, titled “AGENTIC NEURO-SYMBOLIC KNOWLEDGE GRAPH REASONING VIA DYNAMIC LEARNABLE CONTEXT ENGINEERING,” introduces a novel framework called CLAUSE. This innovative approach redefines context construction as a learning problem, where the system intelligently decides what information to expand, which paths to follow or backtrack within a knowledge graph, what evidence to keep, and when to stop. This dynamic process allows for per-query adaptation to balance accuracy, latency (speed), and cost without needing to retrain the entire system.
CLAUSE operates as an agentic neuro-symbolic framework, meaning it combines the strengths of symbolic reasoning (like traversing a knowledge graph) with neural networks for scoring and decision-making. It features three specialized agents that work together:
- Subgraph Architect: This agent is responsible for building a focused subgraph around the question. It makes careful, reversible edits to the graph, adding or deleting edges only when their utility outweighs a learned ‘price,’ ensuring the subgraph remains compact and relevant.
- Path Navigator: Once a subgraph is formed, this agent explores and revises reasoning paths within it. It decides whether to continue exploring, backtrack, or stop, all while respecting a predefined step budget. This leads to human-readable traces of its reasoning.
- Context Curator: The final agent selects a minimal set of textual snippets from the discovered paths. It uses a ‘listwise, redundancy-aware’ scoring system with a learned stopping mechanism to ensure that only the most relevant and non-redundant information is passed to the LLM, adhering to a token budget.
These three agents are coordinated using a sophisticated algorithm called Lagrangian-Constrained Multi-Agent Proximal Policy Optimization (LC-MAPPO). This algorithm allows for centralized training with decentralized execution, meaning the agents learn together but act independently during inference. LC-MAPPO jointly optimizes for task reward (answer accuracy) while enforcing per-query budgets on edge edits, interaction steps, and selected tokens. This is achieved by using ‘dual variables’ that effectively ‘price’ resources, allowing the system to learn when the marginal utility of an action no longer justifies its cost.
During deployment, users can specify either strict budgets (caps) for resources or set ‘prices’ for resources, allowing for flexible trade-offs between accuracy, speed, and cost. The symbolic nature of CLAUSE’s decisions also provides auditable traces, showing exactly what information was added, explored, and selected.
Empirical results across popular multi-hop knowledge graph question answering datasets like HotpotQA, MetaQA, and FactKG demonstrate CLAUSE’s effectiveness. It consistently achieves higher accuracy (EM@1) while significantly reducing subgraph growth and end-to-end latency, often at equal or lower token budgets compared to strong baselines like GraphRAG. For instance, on MetaQA-2-hop, CLAUSE achieved a +39.3 EM@1 improvement with 18.6% lower latency and 40.9% lower edge growth relative to GraphRAG.
The framework also shows superior constraint satisfaction, with LC-MAPPO achieving a 191% improvement in feasibility rate compared to standard MAPPO, and reducing latency violations and costs. Ablation studies further confirm that each of the three agents and the adaptive dual updates are crucial for CLAUSE’s ability to jointly optimize accuracy and efficiency under resource constraints.
Also Read:
- Dynamic Agent Coordination: Leveraging Knowledge Bases for Smarter Multi-Agent Systems
- ToMPO: Boosting LLM Strategic Decisions in Complex Social Games
In essence, CLAUSE offers a powerful and practical solution for knowledge graph question answering, moving beyond rigid heuristics to a learned, budget-aware control system. It delivers compact, verifiable context and predictable performance, making it a significant step forward for deploying LLM-based systems in real-world scenarios. You can read the full research paper here: AGENTIC NEURO-SYMBOLIC KNOWLEDGE GRAPH REASONING VIA DYNAMIC LEARNABLE CONTEXT ENGINEERING.
