Skip to content

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

🕒 Published (v1): 2026-01-01 · Source: ICLR · Venue: ICLR 2026 · link

Why this paper was selected

Neuro-symbolic KG reasoning agent with dynamic learnable context; latency-aware

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

CLAUSE is a three-agent neuro-symbolic framework for multi-hop knowledge graph question answering (KGQA) that treats context construction as a constrained sequential decision process. It jointly learns to edit subgraphs, navigate reasoning paths, and curate evidence under explicit per-query budgets on graph edits, interaction steps, and selected tokens. On MetaQA-2-hop it achieves +39.3 EM@1 over GraphRAG with 18.6% lower latency.

Problem

Static \(k\)-hop graph expansions over-retrieve, inflate prompt token mass, and introduce distractor triples that degrade accuracy. Existing agentic LLM pipelines (ReAct, AutoGen) offer no principled control over per-episode resource costs (interaction steps, token budget), yielding unpredictable latency. Neither path/rule learners (MINERVA, NeuralLP) nor graph-guided RAG (GraphRAG) jointly optimize subgraph construction, traversal, and evidence selection under explicit deployment constraints.

Method

CLAUSE frames multi-hop KGQA as a Constrained Markov Decision Process (CMDP):

\[\max_\pi \mathbb{E}_{\tau\sim\pi}[R^{acc}(\tau)] \quad \text{s.t.} \quad \mathbb{E}[C_{edge}]\le\beta_{edge},\; \mathbb{E}[C_{lat}]\le\beta_{lat},\; \mathbb{E}[C_{tok}]\le\beta_{tok}\]

Three cooperative agents act on an evolving symbolic subgraph \(G_t=(V_t,E_t)\):

  1. Subgraph Architect — scores frontier-adjacent edges via a fused signal \(s(e|q,G_t)=w^\top[\phi_{ent},\phi_{rel},\phi_{nbr},\phi_{deg}]\) and issues ADD/DELETE/STOP, maintaining a frontierized, compact subgraph.
  2. Path Navigator — maintains a path prefix \(p_t\), issuing CONTINUE/BACKTRACK/STOP based on whether expected shaped value exceeds the current step price \(\lambda_{lat}\).
  3. Context Curator — performs listwise, redundancy-aware selection from a textualized candidate pool \(P_t\) with an explicit STOP conditioned on the token dual \(\lambda_{tok}\).

Training uses LC-MAPPO: centralized training with decentralized execution (CTDE) where a multi-head centralized critic estimates one task head \(Q^{task}\) and three cost heads \((Q^{edge}, Q^{lat}, Q^{tok})\). Per-step shaped returns are \(r_t' = r_t^{acc} - \lambda_{edge}c_t^{edge} - \lambda_{lat}c_t^{lat} - \lambda_{tok}c_t^{tok}\). Dual variables are updated by projected stochastic ascent \(\lambda_k \leftarrow [\lambda_k + \eta(\hat{\mathbb{E}}[C_k]-\beta_k)]_+\), optionally stabilized with PID control. At inference, operators choose cap mode (hard \(\beta\)) or price mode (fixed \(\lambda\)) from a single checkpoint.

Key Contributions

  • Formulation: casts KGQA as requirements-conditioned context assembly with three deployment-relevant costs (edge edits, interaction steps, selected tokens) exposed as per-query budgets or prices.
  • Framework: CLAUSE — three-agent neuro-symbolic controller with auditable symbolic actions (EDIT/TRAVERSE/CURATE/STOP) scored by lightweight neural modules.
  • Algorithm: LC-MAPPO — Lagrangian-constrained CTDE-PPO with a multi-head critic (task + three cost heads) and separate dual variables per resource type, replacing single-penalty RCPO.
  • Deployment controls: single checkpoint supports both hard budget caps and smooth price-based trade-offs without retraining; symbolic traces provide step-level provenance.

Results

  • HotpotQA (distractor): CLAUSE 71.7 EM@1 vs. KG-Agent 68.7, GraphRAG 50.1, Hybrid RAG 66.0; latency 1.48× vs. AutoGen 2.43×.
  • FactKG: CLAUSE 84.2 vs. KG-Agent 82.1, Hybrid RAG 80.2.
  • MetaQA-1/2/3-hop: CLAUSE 91.0/87.3/85.5 vs. KG-Agent 87.3/78.0/75.4; MetaQA-1-hop latency 0.98× (below Vanilla RAG baseline).
  • vs. GraphRAG on MetaQA-2-hop: +39.3 EM@1, 18.6% lower latency, 40.9% lower edge growth.
  • LC-MAPPO vs. MAPPO / fixed-penalty PPO / RCPO: LC-MAPPO achieves higher EM and lower resource consumption across all datasets; ablations removing any agent or constraint handling degrade accuracy and/or efficiency.

Limitations

  • Evaluation is limited to three KGQA datasets (HotpotQA, MetaQA, FactKG); generalization to open-domain or heterogeneous KGs is untested.
  • Lightweight neural scorers use frozen encoders; end-to-end fine-tuning of encoders jointly with RL is not explored.
  • LC-MAPPO requires per-dataset training; the single checkpoint adapts at inference only to budget/price settings, not to entirely new domains without retraining.
  • Constrained RL convergence guarantees (Appendix H) apply under standard CMDP assumptions; non-stationarity from multi-agent interactions may affect practical convergence speed.
  • Edge budget normalization to Vanilla RAG conflates retrieval costs with graph construction costs, making cross-dataset efficiency comparisons approximate.

Relevance to Agentic AI / LLM Agents

CLAUSE demonstrates a concrete architecture for budget-constrained multi-agent orchestration over structured knowledge: each agent has a specialized role (build, navigate, curate), a well-defined action space, and a learned STOP condition — a pattern directly applicable to tool-use and retrieval-augmented agents. The LC-MAPPO training scheme addresses a key open problem in agentic systems: enforcing per-episode resource constraints (latency, cost) without hard-coding stopping rules or hop limits. The Lagrangian dual approach to accuracy–efficiency Pareto frontiers is a deployable mechanism that complements prompt-based "think-longer" strategies by controlling what evidence is visible rather than only how long the model reasons. This work advances the line of research on making agentic reasoning both auditable (symbolic traces) and economically controllable (price-mode inference dials).