FinLifeBench Evaluates Longitudinal Financial State Reconstruction in Banking Dialogues
September 2, 2026
FinLifeBench tests LLM ability to reconstruct 24 event types and 34 financial state paths from 6,000 banking dialogue sessions. Testing across eleven models shows event-anchor recall drops from 0.591 to 0.445 as context increases from 15 to 300 sessions.
HOW THIS AFFECTS YOU
●
researcherThe benchmark provides a deterministic way to evaluate how context length affects long-term event recall.
●
healthThis methodology for longitudinal state reconstruction could be adapted to tracking complex patient histories.