Factual accuracy and hallucination reduction in language models for legal information processing

Develop reliable techniques to improve factual accuracy and reduce hallucinations in outputs from large language models used in legal information-processing applications, enabling dependable use in consequential legal settings.

Background

The authors identify hallucinations—incorrect or fabricated outputs—as a key limitation of LLMs in legal contexts, posing significant hurdles to adoption for consequential tasks.

Despite many efforts to improve factuality, the paper asserts that achieving robust factual accuracy remains an unsolved research problem, necessitating close verification of model outputs prior to use.

References

While there are many ongoing efforts to improve factual accuracy, it is as yet an unsolved research problem.

— Promises and pitfalls of artificial intelligence for legal applications  (2402.01656 - Kapoor et al., 2024) in Section “Information processing,” paragraph “Unresolved limitations make the adoption of language models challenging”

Our findings indicate that pre-arrest inference from incomplete evidence remains an open challenge.

— Before the Arrest: Benchmarking LLMs on Criminal Profiling from Incomplete Evidence  (2609.19965 - Yao et al., 17 Sep 2026) in Abstract

Finally, we evaluate retrieval independently of its effect on end-to-end legal QA. Future work should test whether the observed retrieval gains lead to more accurate and better grounded answers.

— GreekBarRetrieval: A Benchmark for Greek Statutory Retrieval  (2608.18752 - Beta et al., 19 Aug 2026) in Section 6, “Conclusions”; Section 7, “Limitations”

Work on domain-specific RAG [4], [5], [12] shows that retrieval quality and hallucination remain open challenges.

— Automated Extraction of Records of Processing Activities (RoPA) Using Hybrid RAG and Locally Deployed Large Language Models  (2609.27359 - Hinh et al., 23 Sep 2026) in Section II, Related Work

Automated interpretation of building codes is a recurring target: prior work covers LLM-based regulation interpretation [7], compliance checking against BIM data [5], and automated code review [9], while surveys [19, 20] confirm reliable regulatory reasoning remains open.

— K-OPSD: Verifiable On-Policy Self-Distillation for Post-Training Vision-Language Models on AEC Drawings  (2609.34082 - Bai et al., 28 Sep 2026) in Section 2, Related work, subsection “LLM applications in the AEC domain”