Language model agents for clinical cohort curation need to handle evolving eligibility criteria without re‑processing unchanged requirements. CohortFlow treats natural‑language corrections as localized updates to a structured cohort specification and reconciles them with a persistent execution graph. Explicit semantic dependencies are recorded so that only affected computation nodes are invalidated while prior valid results are preserved.
Across 100 scenarios derived from clinical‑trial eligibility criteria, encompassing 126 correction turns, localized steering achieves a micro‑F1 of 77.0% versus 76.4% for full regeneration using the same 8B model, a difference that is not statistically significant ($p=0.493$). The approach reduces mean correction latency by 64.8% after completion and 44.3% during execution, with 38.1% fewer model calls.
In a separate proof‑of‑concept evaluation on MIMIC, all 13 cohort states execute successfully, and the correction trajectories demonstrate selective preservation of unaffected computations and recomputation of changed or dependent nodes. Together, these findings show that localized semantic steering can substantially cut correction latency and model calls while maintaining strong requirement‑level accuracy.
Review