---
title: 'FiqhQA: Islamic Jurisprudence QA Research'
url: https://www.emergentmind.com/topics/fiqhqa
type: topic
---

# FiqhQA: Islamic Jurisprudence QA Research

FiqhQA denotes question answering over Islamic jurisprudence (*fiqh*), a research area in which systems are expected to answer legal and normative questions from authoritative Islamic sources while preserving jurisprudential nuance, source fidelity, and, in many settings, school-sensitive interpretation. Recent work treats the area as broader than ordinary fact retrieval: some tasks are multiple-choice benchmarks over legal knowledge, some are extractive or retrieval-grounded systems over Qur’an, Hadith, fiqh manuals, and fatwa corpora, and some are rule-constrained reasoning problems such as inheritance and zakat computation. A central theme across the literature is that fiqh is not a simple single-answer domain, because valid rulings may differ across madhāhib, legal answers may require multi-step reasoning, and incorrect or weakly grounded answers are considered high-stakes in religious use [2603.23750].

## 1. Domain definition and jurisprudential structure

FiqhQA is defined in practice by the kinds of questions it must answer. Some systems target “basic Islamic questions” about worship law, halal/haram practical matters, muamalah, and contemporary legal questions; others target fatwa-style guidance with citations; still others handle structured calculations such as inheritance or zakat. The common thread is that answers are expected to be grounded in authenticated religious material and to remain conservative about legal authority [2512.16644].

The domain differs from generic Islamic QA because the legal object is often not a single textual fact. Inheritance reasoning, for example, requires identifying eligible heirs, resolving blocking rules (*hajb*), assigning fixed and residual shares, applying *ʿawl* and *radd*, and producing a coherent final distribution. The literature repeatedly emphasizes that this is simultaneously juristic and arithmetical, and that rule application can vary across Sunni madhāhib and civil-law codifications. A model therefore cannot safely assume a single implicit legal configuration [2603.24012].

A second defining property is jurisprudential pluralism. The IslamicMMLU Fiqh track explicitly states that evaluation must distinguish between a wrong answer and a different valid madhhab answer. Its bias-detection task operationalizes this by posing general fiqh questions for which all four options are correct, each corresponding to one of the four Sunni madhāhib. In that setting, the model is not judged for correctness in the usual sense; its choice reveals which school it defaults to when unconstrained. This formulation makes school-of-thought preference a first-class evaluation target rather than an annotation nuisance [2603.23750].

The domain is also closely related to non-factoid Islamic QA. A large context-enriched benchmark over Quranic Tafsir and Ahadith argues that many religious questions require contextual evidence, cross-referencing multiple sources, interpretation of legal or ethical implications, and a scholarly verdict rather than a short span answer. This suggests that FiqhQA spans a continuum from extractive answer selection to long-form, scholar-aligned response generation [2409.09844].

## 2. Benchmarks, corpora, and source construction

The most explicit jurisprudence benchmark in the recent literature is the Fiqh component of IslamicMMLU. It contains 4,000 Arabic four-option MCQs within a 10,013-question benchmark and is built from *Jurisprudence According to the Four Schools* by `Abd al-Rahman al-Jaziri, commissioned by Al-Azhar University. The source spans 2,221 pages, 1,043 sections, and eight fiqh categories. About 60% of sections were found to contain extractable rulings, while 40% were meta-commentary, philosophical discussion, or cross-references. From the extractable material, the authors constructed a structured corpus of 2,163 rulings across 797 topics, with near-uniform madhhab coverage of roughly 25% per school. Extraction was performed in parallel by GPT-4.1, GPT-4o, and GPT-5.1, aligned through ensemble voting, and adversarially verified by o1; 85% of extractions passed immediately, 12% entered iterative refinement via o3, and 3% were rejected. Human validation on a stratified sample of 213 questions yielded 207 approvals, or 97.2%, with three minor revisions and three rejections for factual issues [2603.23750].

MAWARITH extends the field from benchmarked recognition to full legal reasoning. It provides 12,500 Arabic inheritance cases, split into 12,000 training and 500 test instances, covering 36 legally recognized heir categories grouped into 8 kinship levels. The dataset includes 11,535 simple cases, 616 *ʿawl* cases, and 349 *radd* cases, and it evaluates whether a model can perform the full sequence of heir identification, blocking, share assignment, adjustment, and final allocation rather than merely choosing among options [2603.07539].

A third line of resource construction targets free-form faithfulness rather than MCQ accuracy. A 3,810-item bilingual Arabic/English generative benchmark with atomic single-gold answers was introduced to measure hallucination and abstention directly under a three-way label scheme of Correct, Incorrect, and Not_Attempted. Although Qur’an-centric, it was assembled from high-quality Islamic QA sources such as Hajj-FQA, QIAS, and PalmX and is methodologically relevant to fiqh because it measures whether a system should answer at all when evidence is missing [2601.07528].

Large non-factoid corpora broaden the source base further. One benchmark reports over 73,000 question-answer pairs enriched with context from Islamqa.org, English Tafsir from Al-Tafsir.com, and English translations of over 33,000 Ahadith from the six major Hadith books. Each example pairs a question and answer with topic-matched religious context, narrowed to the top 3 Ayahs and top 3 Ahadith by soft cosine similarity. Its explicit aim is to support scholar-like answers over larger contexts rather than isolated factual retrieval [2409.09844].

A deployment-oriented system description adds another important evaluation resource: Fanar-Sadiq reports using FatwaQA, a 2,000-example Arabic fatwa-style QA benchmark, alongside IslamicFaithQA and QIAS. This reflects a trend in which FiqhQA is evaluated not only as knowledge recall but also as grounded open-ended legal assistance [2603.08501].

| Resource | Coverage | Main characteristic |
|---|---|---|
| IslamicMMLU Fiqh | 4,000 Arabic MCQs | 3,200 knowledge items and 800 madhhab-bias items |
| MAWARITH | 12,500 Arabic inheritance cases | Full reasoning chain with step-by-step legal solutions |
| 3,810-item bilingual generative benchmark | Arabic and English | Measures hallucination and abstention |
| Context-enriched non-factoid Islamic QA dataset | Over 73,000 QA pairs | Long-form answers with Tafsir and Hadith context |
| FatwaQA | 2,000 Arabic QA pairs | Fatwa-style jurisprudence and finance evaluation |

## 3. Methodological paradigms

Methodologically, FiqhQA research is heterogeneous because the task family itself is heterogeneous. Closed-domain consultation systems use retrieval and ranking over curated question-answer corpora. One Sharia chatbot is built under CRISP-DM and combines Sentence-Transformers, specifically `paraphrase-multilingual-MiniLM-L12-v2`, with Q-Learning. User questions are cleaned, normalized, tokenized, embedded, and matched to stored questions by cosine similarity; the Q-learning reward is derived from semantic similarity, and the action space is the set of candidate answers. The system retrieves and ranks an existing answer rather than generating a new fatwa, and may apply TF-IDF-based extractive summarization to return 2–3 representative sentences [2512.16644].

More recent systems adopt retrieval-augmented generation with stronger control mechanisms. Fanar-Sadiq begins with intent-aware routing across nine intent classes, including fiqh rulings, zakat calculation, and inheritance calculation. Fiqh queries are sent to a “Fiqh Documents Retriever” and a reasoning agent that is prompted to state ruling scope and assumptions, separate legal conclusion from evidential support, and assign deterministic citation tags such as \([CITE:N]\). The broader retrieval layer searches more than 500,000 documents drawn from Qur’an, major Hadith collections, classical fiqh texts, contemporary fatwas, Islamic history, and scholarly articles, using Qwen3-Embedding-4B, cosine similarity, top-\(k\) retrieval with default 12, optional cross-encoder reranking, source diversity enforcement, and citation normalization [2603.08501].

Inheritance reasoning has pushed the field toward hybrid neuro-symbolic design. The QIAS 2026 winning system is a RAG pipeline for *ʿilm al-mawārith* built from a deterministic inheritance calculator, roughly 100,000 unique synthetic cases, hybrid dense-plus-BM25 retrieval with weighted Reciprocal Rank Fusion, cross-encoder reranking, Qwen3.5-9B generation, and a deterministic validation layer that enforces required keys, types, legal labels, and approximate mass conservation with \(\varepsilon = 5\). The natural-language Q&A rendering of symbolic solutions is used as the primary retrievable document, which allows the generator to imitate intermediate legal traces rather than only final distributions [2603.24012].

Agentic retrieval has been proposed for broader faithful Islamic QA. An end-to-end grounded suite combines 25K Arabic text-grounded SFT reasoning pairs, 5K bilingual preference samples, and a verse-level Qur’an retrieval corpus of 6,236 atomic ayat. Dense retrieval uses mE5-base with ChromaDB; agentic RAG adds explicit tool calls such as `search_quran`, `get_surah_info`, `read_ayah`, and `search_surah` in a two-turn evidence-seeking and answer-finalization process. Although this framework is Qur’an-centered, its stated purpose is to reduce free-form hallucination and unsupported confident answering in Islamic QA, a problem directly relevant to fiqh assistance [2601.07528].

Neighboring work on Quranic and Hadith QA supplies transferable design patterns. A two-stage Quranic QA system shows that retrieval can benefit from fine-tuned Arabic encoders and ensembling, while answer extraction can benefit from instruction-tuned LLMs with few-shot prompting when labeled data are scarce; the authors explicitly note that transfer to fiqh QA would require specialized legal corpora, stronger source attribution, school-specific interpretation, and multi-hop legal inference [2508.06971]. CGRA-DeBERTa, developed for Hadith QA, adds LoRA-based DeBERTa adaptation and a residual concept-gating mechanism driven by an Islamic Concept Dictionary of 12 core terms, with token-level scaling from \(1.04\times\) to \(3.00\times\), to prioritize semantically decisive theological tokens during span extraction [2602.15139].

## 4. Evaluation regimes and what they measure

Evaluation in FiqhQA is notably multi-paradigmatic. IslamicMMLU uses a zero-shot MMLU-style four-option MCQ setup, so the random baseline is 25%. Within that setup, however, the Fiqh track separates ordinary knowledge questions from the 800-item bias-detection task. Bias is defined relative to a neutral \(25/25/25/25\) school-selection distribution over Hanafi, Maliki, Shafi’i, and Hanbali options, and significance is tested with a chi-squared test after answer order randomization to control positional bias [2603.23750].

Inheritance benchmarks evaluate intermediate reasoning stages rather than only the final split. MAWARITH introduces MIR-E as a weighted multi-stage score over heir identification, share assignment, adjustment, and final allocation, with weights \(\alpha_h = \alpha_s = \alpha_f = 0.30\) and \(\alpha_a = 0.10\). Its design makes early-stage legal errors visible rather than letting them disappear behind an accidental correct final number [2603.07539]. The QIAS 2026 benchmark uses a related MIR-E formulation that scores heirs, blocked, shares, *ʿawl_or_radd*, *ʿawl_stage*, and final distribution, with stage weights \(0.30/0.20/0.20/0.10/0.20\). This stage-aware scoring reflects the view that legal structure and arithmetic coherence are equally essential [2603.24012].

Generative faithfulness introduces a different evaluation axis. The 3,810-item bilingual benchmark grades outputs as Correct, Incorrect, or Not_Attempted and uses GPT-4.1 as an LLM judge with human calibration. The goal is not merely whether a model can state something plausible, but whether it answers supportably and abstains when support is absent [2601.07528].

Long-form religious QA complicates metric choice further. The context-enriched non-factoid benchmark evaluates models with ROUGE and BERTScore but reports that human evaluation tells a different story: verdict consistency with expert scholars is low even when contextual understanding is much higher. In this line of work, automatic lexical-overlap metrics are treated as insufficient for jurisprudential correctness [2409.09844].

The field therefore evaluates at least four partially distinct properties: legal knowledge recognition, stage-wise reasoning correctness, faithfulness with abstention, and scholar-aligned verdict quality. One implication is that FiqhQA systems with high scores under one regime are not automatically reliable under another.

## 5. Empirical findings

IslamicMMLU shows that fiqh knowledge is challenging but tractable for frontier systems. Across 26 evaluated models, the Fiqh track spans 89.06% to 45.16%. On the 3,200 knowledge questions, Gemini 3 Flash reaches 89.1% Fiqh accuracy, while GPT-3.5-turbo is at 45.2%. The paper reports top-model accuracies of 89.7% for Hanafi-tagged items, 88.9% for Maliki, 92.2% for Shafi’i, and 87.1% for Hanbali, while Comparative Fiqh is the hardest subtask, with the lowest top accuracy reported as 80.0 in one breakdown and 76.7% in another summary discussion. Error analysis identifies three dominant failure modes: confusion between schools when a specific madhhab is specified, failure on multi-hop reasoning, and difficulty with comparative reasoning. The reported frontier-to-baseline contrast on multi-hop fiqh questions is 87.2% versus 41.1% [2603.23750].

The same benchmark also establishes that school preference can be measured empirically. Gemini 3 Flash is the closest to neutral on the bias-detection task, with \(\chi^2 = 0.32\) and \(p > 0.9\), whereas GPT-5.1 significantly favors Hanbali at 35.6%. Several Arabic-specific models show moderate Hanafi leaning in the roughly 32–33% range. After Bonferroni correction for 26 simultaneous tests, only Claude Opus 4.5 remains statistically significant with \(\chi^2 = 16.06\). Bias magnitude correlates negatively with model accuracy, with Pearson \(r = -0.38\) \((p = 0.049)\) and Spearman \(\rho = -0.53\) \((p = 0.005)\), though the authors caution that high capability does not guarantee neutrality [2603.23750].

MAWARITH demonstrates that full inheritance reasoning remains difficult for most models. In zero-shot Arabic prompting, Gemini-2.5-flash reaches 0.918 MIR-E on validation and 0.901 on test, whereas Qwen3-32B, GPT-OSS-120B, LLaMA-3.3-70B, Fanar-Sadiq, and Fanar-C-2-27B remain between 0.437 and 0.328 on test. The major bottleneck is early-stage heir identification: Gemini reaches about 78.2% success at the first stage, while all other models stay below 25%. The error taxonomy includes scenario misinterpretation, false eligibility, false blocking, count mismatch, share-rule errors such as mishandling the mother’s conditional share under Qur’an 4:11, confusion over *ʿawl* and *radd*, and final normalization errors [2603.07539].

System-level evaluations show that tool use and routing improve practical performance. Fanar-Sadiq reports 90.1% routing accuracy on a manually annotated set of 700 user queries, 65.4 on IslamicFaithQA, 65.1 on FatwaQA, and 72.2% on QIAS Task 1. The authors note that the inheritance gap relative to stronger proprietary models is partly due to the extra difficulty of mapping a computed distribution into discrete multiple-choice answers [2603.08501]. A separate inheritance-focused RAG system reaches MIR-E \(= 0.935\) and ranks first on the official QIAS 2026 blind-test leaderboard, supporting the claim that retrieval-grounded, schema-aware generation is more reliable than unconstrained answering for this subdomain [2603.24012].

Closed-domain retrieval systems remain effective for lower-complexity consultation. The Sharia chatbot evaluates 100 test scenarios across worship, fiqh/jurisprudence, social ethics, muamalah, and contemporary Islamic issues, with 87 responses judged relevant, 9 fairly relevant, and 4 not relevant, yielding 87% semantic accuracy. The result is strongest for straightforward informational use rather than complex layered jurisprudential adjudication [2512.16644].

## 6. Limitations, controversies, and future directions

The literature is explicit about source and scope limitations. IslamicMMLU’s Fiqh corpus comes entirely from al-Jaziri’s encyclopedia, which may underrepresent minority opinions within schools, regional variation, and contemporary *ijtihad*. The benchmark is Sunni-only, excludes Shia jurisprudence, and the authors explicitly plan future extension to Shia schools. Human validation is based on a single expert reviewer, so inter-annotator agreement is unavailable. The multiple-choice format measures recognition rather than generative jurisprudential reasoning, and the authors caution that results should not be used to market systems as authoritative Islamic advisors [2603.23750].

Deployment-oriented systems have their own technical constraints. The Sharia chatbot is static after deployment, dataset dependent, lacks multi-turn dialogue support, and uses tabular Q-learning, which the paper notes is not ideal for very large or continuous semantic state spaces. For complex fiqh issues requiring layered reasoning, contextual awareness, or expert interpretation, the model may return a semantically close but jurisprudentially imperfect answer [2512.16644].

Generative faithfulness benchmarks also have structural blind spots. The 3,810-item bilingual benchmark is built around atomic single-gold answers, which underrepresents cases where multiple madhhab-based answers are valid. It is Qur’an-centric, depends on an LLM judge with limited human calibration, and the associated agentic RAG setup introduces latency and failure modes such as tool-use errors and citation laundering. The stated aim is research on faithfulness and abstention, not fatwa issuance [2601.07528].

Long-form Islamic QA exposes a broader evaluation controversy. In the context-enriched benchmark, human evaluation reports verdict consistency with expert scholars between 11% and 22%, while contextual understanding ranges from 53% to 90%. The discrepancy between these judgments and automatic metrics such as ROUGE is presented as evidence that surface overlap is inadequate for religious verdict assessment. Future work is directed toward broader Tafsir coverage, richer scholarly sources, and evaluation procedures that are more sensitive to jurisprudential correctness [2409.09844].

Multi-agent grounded assistants remain assistive rather than authoritative. Fanar-Sadiq states that routing errors may send a query to a suboptimal module, citations may still be incomplete, jurisprudential diversity can be difficult to summarize without oversimplification, and the deterministic calculators implement only a subset of schools and disputed cases. The paper therefore frames the system as decision support rather than a substitute for qualified scholarship [2603.08501].

Across these strands, a consistent future direction is visible: stronger source-grounding, explicit madhhab conditioning, better abstention, stage-aware reasoning evaluation, richer fiqh corpora beyond single-source encyclopedias, and interfaces that expose evidence and assumptions rather than masking them behind fluent prose.

Source: https://www.emergentmind.com/topics/fiqhqa