A new research paper posted on arXiv cs.CL outlines a retrieval-augmented generation model designed to run on smartphones for the legal context of Bangladesh. The system focuses exclusively on statutory interpretation, deliberately excluding queries that require case-law reasoning or judicial precedent. By compressing a 9-billion-parameter Gemma-2 teacher model down into a 2-billion-parameter student via a two-phase progressive knowledge distillation process, the developers aim to make statutory information functional on mobile devices with unreliable connectivity.
Addressing the Access Gap in Bangladesh
Accessing legal information presents substantial hurdles for the majority of citizens in Bangladesh. Statutory texts are published strictly in English, and legal professionals are largely concentrated within urban centres. Furthermore, relying on cloud-dependent artificial intelligence fails in regions with intermittent mobile networks, a limitation that can lead to hallucinated legal text and cause direct harm to users. The announcement did not specify the exact hardware specifications required for the smartphone execution, nor did it disclose deployment timelines or evaluation benchmarks against existing commercial systems.
Unknowns and Future Deployment
Several critical details regarding the practical rollout of this smartphone-compatible model remain unaddressed. The available documentation does not state which specific mobile operating systems are supported, how local language translation or comprehension is handled beyond the English statutory text, or whether the model will be released under an open-source license for public audit. It is also unclear how the system will manage updates when statutory laws are amended. Further updates from the researchers will be necessary to determine its real-world viability.
The project highlights ongoing academic efforts to adapt large language models for resource-constrained environments where cloud infrastructure is unavailable.
Lead image: FOX ^.ᆽ.^= ∫ / Pexels
