October 21, 2026
To:
Editors-in-Chief
Neurosymbolic Artificial Intelligence
Subject: Submission of Original Research Manuscript, “Hierarchical Neuro-Symbolic Knowledge Tracing: From Attention Maps to Transferable Production Rules”
Dear Editors-in-Chief,
On behalf of my co-authors, Upendra Singh Jadon and Bineet Kumar, I am pleased to submit our research manuscript entitled “Hierarchical Neuro-Symbolic Knowledge Tracing: From Attention Maps to Transferable Production Rules” for consideration as an original research article in Neurosymbolic Artificial Intelligence.
Our research addresses a fundamental challenge in neuro-symbolic artificial intelligence: how to transform patterns learned by high-performing neural networks into interpretable, verifiable, and transferable symbolic knowledge without compromising predictive performance or introducing prohibitive computational overhead.
We introduce Hierarchical Neuro-Symbolic Knowledge Tracing (HNSKT), a framework integrating attention-based deep learning, hierarchical inductive logic programming, formal hypothesis revision, and symbolic knowledge transfer. Although evaluated in educational knowledge tracing, the underlying contributions address broader questions of neural-symbolic integration, explainable reasoning, efficient rule induction, and human-guided knowledge refinement.
The manuscript makes four principal contributions:
1. Scalable symbolic knowledge discovery: We introduce a two-stage hierarchical induction mechanism that transforms transformer attention patterns into human-readable first-order logic rules. By clustering related skills before rule induction, the approach substantially reduces computational search complexity. In our experiments, hierarchical induction completed in approximately eight minutes, compared with a flat ILP approach that did not complete within 48 hours.
2. Formally constrained human-in-the-loop reasoning: We model expert modifications to discovered rules as constrained hypothesis revision, defining permissible mutation operations that preserve the specified ILP hypothesis-space constraints. This enables educators to inspect, refine, and reject symbolic knowledge while avoiding the need to retrain the underlying neural model following each rule modification.
3. Transferable symbolic knowledge: We demonstrate how discovered rules can be represented as symbolic priors and incorporated into neural predictions in a target domain. In a synthetic physics transfer experiment using only 20% of the target training data, the proposed approach improved AUC from 0.65 to 0.73, representing a 12.3% relative improvement. Additional experiments with synthetic biology provide further evidence of the potential for cross-domain knowledge reuse.
4. Empirical assessment of interpretability and attention faithfulness: Through experiments involving ASSISTments, EdNet, and synthetic prerequisite structures, we examine predictive performance, rule recovery, scalability, and robustness. HNSKT achieves an AUC of 0.78 on ASSISTments, comparable to the underlying attention-based neural model, while recovering prerequisite relationships with precision of 0.74 and recall of 0.68 in synthetic validation. A preliminary evaluation involving five teachers provides encouraging evidence of rule interpretability. We additionally investigate temporal recency as a confounding factor in attention-derived knowledge and identify limitations requiring further research.
We believe this work is particularly relevant to Neurosymbolic Artificial Intelligence because it contributes to the principled integration of neural representation learning and symbolic reasoning. Rather than treating symbolic logic solely as a predefined constraint on neural systems, HNSKT investigates how learned neural representations can serve as a source of explicit knowledge, how that knowledge can be refined through human expertise, and how it can subsequently inform neural predictions.
The manuscript also examines important limitations of attention-derived explanations. In particular, we acknowledge that temporal proximity may influence inferred dependencies, that cross-domain transfer has so far been evaluated using synthetic target domains, and that the initial teacher evaluation requires replication with a larger and more diverse participant group. We believe these findings contribute to a balanced discussion of the opportunities and remaining challenges in building interpretable and trustworthy neuro-symbolic systems.
Given the journal's emphasis on advances in neural-symbolic integration, knowledge representation, learning, reasoning, and explainability, we believe our study will be of interest to researchers working on neuro-symbolic architectures, inductive logic programming, explainable AI, human-in-the-loop learning, and knowledge transfer.
We appreciate the journal's commitment to rigorous, open, and transparent peer review and would welcome the opportunity to engage constructively with reviewers and the research community.
Thank you for considering our manuscript for publication. We look forward to your evaluation and feedback.
Sincerely,
Ankit Anand
Corresponding Author
ORCID: 0009-0003-5540-5566