TUSI-Bench: a Persian legal benchmark for citation-grounded RAG evaluation
Researchers introduce TUSI-Bench, a structured Persian legal benchmark with two tasks requiring answers tied to explicit legal citations, plus a graph-based retrieval baseline that reportedly improves recall.
In plain words
Retrieval-augmented generation (RAG) means AI systems that answer questions by pulling in source documents. The TUSI-Bench authors built a Persian-language legal test set so such systems can be checked on whether their answers cite real legal sources. It includes about 2,000 open questions and roughly 500 multiple-choice questions drawn from Iranian Bar exam material.
What was announced
The benchmark rests on a legal knowledge framework of 7,261 articles spanning codified statutes, commentaries, doctrinal writings, and Supreme Court unification rulings. The authors argue existing Persian legal resources fall short: some are generated synthetically from statutes and lack natural information needs, while others come from informal question-answer sources without links to authoritative documents. TUSI-Bench also includes a knowledge graph connecting articles and concepts, and a graph-based retrieval baseline that uses those relationships. In the reported experiments, adding graph-based relations improved retrieval recall by up to 8 percentage points over standard baselines.
The benchmark is positioned as a standardized resource for evaluating grounded legal reasoning in Persian, meaning systems must justify answers with explicit citations rather than free-form text. The results and the benchmark itself come from the paper's authors; the figures above are their own reported measurements, not independent evaluations.
What it does not show
- Evidence is limited to the paper's abstract; the full methodology, dataset construction details, and evaluation protocols were not available for review.
- The recall improvement and all other results are the authors' own reported figures and have not been independently confirmed.
- The abstract does not state whether the benchmark or baseline code is publicly released.
For security and IT teams
- Scope
- The benchmark targets Persian-language legal RAG evaluation, covering Iranian statutes, commentaries, doctrinal writings, and Supreme Court unification rulings.
Sources
Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 19 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.