[Paper Review] Explainable Innovation Engine: Dual-Tree Agent-RAG with Methods-as-Nodes and Verifiable Write-Back

Renwei Meng|arXiv (Cornell University)|Mar 10, 2026

Scientific Computing and Data Management0 citations

TL;DR

Proposes an Explainable Innovation Engine that uses methods-as-nodes with a dual-tree provenance and abstraction structure, plus a strategy-driven synthesis and verifiable write-back to enable controllable, auditable RAG-based innovation.

ABSTRACT

Retrieval-augmented generation (RAG) improves factual grounding, yet most systems rely on flat chunk retrieval and provide limited control over multi-step synthesis. We propose an Explainable Innovation Engine that upgrades the knowledge unit from text chunks to methods-as-nodes. The engine maintains a weighted method provenance tree for traceable derivations and a hierarchical clustering abstraction tree for efficient top-down navigation. At inference time, a strategy agent selects explicit synthesis operators (e.g., induction, deduction, analogy), composes new method nodes, and records an auditable trajectory. A verifier-scorer layer then prunes low-quality candidates and writes validated nodes back to support continual growth. Expert evaluation across six domains and multiple backbones shows consistent gains over a vanilla baseline, with the largest improvements on derivation-heavy settings, and ablations confirm the complementary roles of provenance backtracking and pruning. These results suggest a practical path toward controllable, explainable, and verifiable innovation in agentic RAG systems. Code is available at the project GitHub repository https://github.com/xiaolu-666113/Dual-Tree-Agent-RAG.

Motivation & Objective

• Address the limitations of flat chunk retrieval and limited multi-step control in RAG systems by introducing methods-as-nodes as the indexing unit.
• Develop a dual-tree knowledge representation (method provenance tree and clustering abstraction tree) for explainable navigation and auditing.
• Design a closed-loop inference pipeline with strategy operators, provenance-aware synthesis, score-based pruning, executable verification, and write-back for continual growth.
• Demonstrate improvements over vanilla baselines across multiple domains and backbone models, with ablations validating component contributions.

Proposed method

• Introduce two complementary structures: a Method Provenance Tree that records weighted derivations among reusable method units, and a Clustering Abstraction Tree that enables hierarchical, top-down navigation.
• Offline multimodal normalization and context-preserving segmentation to extract prior and derived methods with edge attributions and weights.
• Deduplicate and canonicalize methods using vector embeddings and cosine similarity; maintain auditable provenance traces.
• Online funnel retrieval on the abstraction tree with a decaying budget to descend to leaf methods, augmented by ancestor provenance backtracking to form a context for synthesis.
• A strategy agent selects from a library of synthesis operators (e.g., induction, deduction, analogy) to generate candidate method nodes, with explanations, novelty, applicability, and a plan for validation.
• A scoring and verification loop (human-interpretable rubric and executable verification in formal domains) prunes low-quality candidates and writes validated nodes back to the repository.
• Optional unsupervised evolution with safety constraints, including falsification mechanisms and provenance logging.

Figure 1: The complete algorithm process

Experimental results

Research questions

• RQ1 How does indexing methods-as-nodes with a dual-tree structure affect explainability and controllability in RAG-based innovation tasks?
• RQ2 What is the impact of strategy-guided synthesis and provenance backtracking on achievement of higher-quality, verifiable derivations across domains?
• RQ3 How do pruning and executable verification affect reliability and continual growth of the method repository?
• RQ4 What are the efficiency and scalability characteristics of the proposed dual-tree Agent-RAG under multimodal inputs?

Key findings

• Agent-RAG yields consistent improvements over vanilla baselines across six domains and multiple backbones.
• The largest gains occur in derivation-heavy domains (e.g., Mathematics).
• Ablations show complementary roles for provenance backtracking and pruning in performance.
• Timeliness experiments reveal a favorable cost–quality trade-off with near-optimal quality under moderate budgets.
• Backbone LLM quality largely governs absolute performance, with Agent-RAG improving domain-averaged scores across all tested backbones.

Figure 2: A schematic diagram of Clustering Abstraction Tree and Method Provenance Tree