CohortFlow reruns only the steps affected by a user correction
CohortFlow represents clinical cohort selection as a reusable execution graph. Across 100 constructed scenarios and 126 corrections, localized edits reduced waiting time without a significant improvement in requirement-parsing accuracy. The study does not establish clinical cohort validity.
When a user changes an eligibility rule halfway through an agent task, how much work should restart? CohortFlow, posted on October 26, studies that question in clinical cohort curation. Its contribution concerns inspectable task state and selective recomputation, rather than evidence that an agent can replace clinical judgment.
- Translate eligibility rules into a task specification
- Preserve conditions and dependencies
- Interpret a correction as a local change
- Reuse valid results and recompute affected steps
- Inspect logic and outputs
Separate the correction from the whole request
Regenerating every requirement after an edit can waste computation and accidentally change conditions the user wanted to retain. CohortFlow first converts natural language into a typed task specification, including requirements, Boolean composition and temporal relationships. It compiles this into a persistent execution graph. A correction becomes a local semantic change, while unchanged requirements retain stable identities.
Reusing a node requires more than finding an old result. Its identity and semantic fingerprint must be unchanged, its artifact must remain available, and every parent must also be reusable. A changed dependency invalidates affected downstream computation. Corrections can be interpreted while work continues, but state changes are committed in order so that an older interpretation cannot overwrite a newer request.
Less waiting, without a significant accuracy gain
The authors constructed 100 scenarios with 126 correction turns from CohortX trial criteria and compared both strategies using the same quantized Qwen3-8B model. Across 110 jointly scoreable states, requirement-field micro-F1 was 77.0% for local correction and 76.4% for full regeneration; p=0.493 does not establish an accuracy improvement. Reported mean correction latency fell from 48.6 to 17.1 seconds after completion, and from 69.7 to 38.8 seconds during an active run. The principal benefit was avoiding repeated work.
What backend execution does not establish
A separate MIMIC backend experiment used manually frozen, valid specifications rather than the first-stage model predictions. All 13 states, comprising 10 scenarios and three structured corrections, executed successfully. That supports the reuse mechanism, not end-to-end clinical accuracy. The first-stage workload also excluded clinical-note retrieval and reasoning, so its timing cannot be treated as a measurement of complete medical-record processing.
Ambiguity remained a weakness: local correction clarified just one of eight cases, versus four for full regeneration. Constructed trajectories and a single quantized model limit generalization, and local changes can still damage global logic. The useful engineering lesson is to preserve state and track invalidation explicitly; clinical validity and human review remain separate requirements.