@inproceedings{sidaoui-2026-ags,
    title = "{AGS}-{KSU} at {QIAS} 2026: A Comparative Study of Prompting and {LLM} Approaches for Structured Islamic Inheritance Reasoning",
    author = "Sidaoui, Hicham Ghazi",
    editor = "Al-Khalifa, Hend  and
      El-Haj, Mo  and
      Ezzini, Saad",
    booktitle = "The 7th Workshop on Open-Source {A}rabic Corpora and Processing Tools ({OSACT}7) with 5 Shared Tasks",
    month = may,
    year = "2026",
    address = "Palma, Mallorca (Spain)",
    publisher = "Association for Computational Linguistics",
    url = "https://preview.aclanthology.org/revision-previews/2026.osact-1.26/",
    doi = "10.63317/2irfqjkfpyyn",
    pages = "209--212",
    abstract = "This paper describes our submission to the QIAS 2026 shared task on structured Islamic inheritance reasoning, based on the MAWARITH benchmark (Bouchekif et al., 2026). The task requires multi-step structured prediction for Arabic inheritance cases, including heir identification, blocking, share assignment, adjustment detection, and final distribution, evaluated with the MIR-E metric. We compare four system configurations: a QLoRA fine-tuned Qwen2.5-3B baseline, a multi-stage Fanar-Sadiq pipeline with deterministic validation and post-processing, and two GPT-5.4 prompting setups. On the official test set, the best result was achieved by GPT-5.4 with explicit inheritance rules and development examples used as in-context demonstrations, reaching a MIR-E score of 0.84, compared with 0.76 for a minimal-prompt GPT-5.4 variant. These results suggest that explicit rule conditioning and in-context demonstrations can improve performance in this setup. Since the compared systems vary in model family and prompting strategy, the findings should be interpreted as a comparison of task configurations rather than a controlled model-only comparison."
}