TRISM NeuroSymbolic AI for Legal: RASOR RAG and symbolic KBs

TRISM NeuroSymbolic AI for Legal: RASOR RAG and symbolic KBs

The authors Deepa Tilwani, Yash Saxena, Ankur Padia, Srinivasan Parthasarathy and Manas Gaur submitted a position paper titled "NeuroSymbolic AI for Legal AI-TRISM: Trustworthy, Reliable, Interpretable, Safe Models" to arXiv on 5 Apr 2026 (arXiv:2606.15646, 1,487 KB). The paper proposes the TRISM framework, which integrates NeuroSymbolic AI principles with large language models and centers Retrieval-Augmented Generation to ground outputs in verified legal sources.

What is TRISM?

TRISM is a framework that combines neural learning and symbolic reasoning to produce interpretable, verifiable legal-model outputs. The paper formalizes extraction of symbolic knowledge from legal textual documents, proposes a symbolic legal knowledge base, and positions RAG as a core component to ground LLM outputs in verified sources.

The authors list four contributions: (1) an analysis of limitations of AI in law; (2) introduction of RASOR RAG, which creates foundations for neurosymbolic RAG by generating explicit interpretable rationales that could be formalized into symbolic representations; (3) a formalized methodology for creating symbolic legal knowledge bases that support interpretable reasoning and output verification in LLMs; and (4) the TRISM framework for integrating symbolic legal knowledge with LLMs.

How do RASOR RAG and the symbolic KB work?

RASOR RAG generates explicit, interpretable rationales that the authors say can be formalized into symbolic representations, while RAG more generally is used to ground model outputs in verified legal sources for citation and precedent verification. The paper positions RAG as a core component of TRISM for grounding LLM outputs.

The authors motivate this by noting key problems with LLMs in legal contexts: lack of interpretable reasoning, a tendency to hallucinate, and struggles with accurate citation attribution and precedent verification. They identify two limitations of current approaches: inadequate integration of structured legal knowledge during training or fine-tuning, and insufficient verification mechanisms for generated legal content. TRISM aims to address those gaps by extracting symbolic knowledge from legal documents and combining that with retrieval-grounded neural generation.

Why it matters

Legal work depends on precise citations and correct precedents; the paper stresses that a single incorrect precedent can jeopardize a case. By producing explicit, formalizable rationales and a symbolic knowledge base, TRISM targets interpretable decision pathways and verifiable outputs, which are direct responses to the paper's stated failure modes of current LLM-based systems in law.

If the ideas in TRISM translate into implemented systems, they would change how models surface their reasoning in legal workflows and how generated legal content is verified against primary sources.

What to watch

Look for follow-up work that implements RASOR RAG and the TRISM pipeline and that publishes empirical evaluations comparing verification rates or citation accuracy. The arXiv entry is the position paper (arXiv:2606.15646, submitted 5 Apr 2026) and any subsequent code, datasets, or formal evaluations would be the next concrete signals of progress.

Authors and source

The paper is authored by Deepa Tilwani, Yash Saxena, Ankur Padia, Srinivasan Parthasarathy and Manas Gaur and was submitted to arXiv (cs.AI) on 5 Apr 2026. The PDF and TeX source are available through the arXiv entry.