Abstract
This paper investigates temporal conflict resolution in legal retrieval-augmented generation systems. The central problem is that laws and regulations change over time, while retrieval systems may continue surfacing outdated or superseded material.
Authors
Dmytro Asieiev, Emilija Bareikaitė, Jadrine Kaburu, Jonas Urnėžius, Aušra Šubonienė, Kostas Ragauskas
What the study evaluates
The study ablates retrieval modules to understand which components help legal RAG systems remain grounded when newer and older legal sources conflict. The evaluation focuses on factual grounding, temporal relevance, and resistance to outdated answers.
Main findings
The results show that retrieval design directly affects whether a legal AI system can distinguish current authority from obsolete context. A RAG pipeline that retrieves relevant documents but ignores temporal conflict can still produce confident and incorrect answers.
The ablation approach clarifies which retrieval components are most important for keeping generated answers aligned with the legal state in force at the time of the query.
Why it matters
Legal AI systems operate in domains where outdated information can be materially harmful. Reliable retrieval must account not only for semantic relevance, but also for whether a source is still valid, superseded, or temporally constrained.