Fully Local AI Cascade for Educational Dialogue De-Identification

Haocheng Zhang and four coauthors submitted a paper on 16 June 2026 proposing a fully local AI cascade for de-identifying educational dialogue and reporting a 0.958 macro F1 on math tutoring transcripts. The paper, arXiv:2606.18372, evaluates three reviewer configurations against same-family LLM-only baselines and a commercial API and runs entirely on a single laptop.

What did the authors build?

The authors built a fully local cascade that reframes de-identification as constrained privacy triage: a recall-first union proposer over-generates candidate spans, then a context-aware reviewer makes a binary Redact/Keep decision for each candidate. The proposer combines two lightweight encoders with deterministic rules to produce candidate spans; the reviewer uses surrounding dialogue and speaker role to decide whether to redact. The design aims to avoid sending student data to third parties while handling ambiguity where, as the paper puts it, "Riemann may refer to a real student or to a mathematical concept."

How did the cascade perform on transcripts?

The strongest local configuration reached 0.958 macro F1 on math tutoring transcripts drawn from two large platforms, while the same-family LLM-only baseline scored 0.767 and a commercial API scored 0.706. The paper also reports the system runs entirely on a single laptop. The authors further evaluated a targeted challenge set focused on curricular-personal name ambiguity: the strongest local configuration degraded by only 0.03 F1 on that set, whereas smaller reviewers degraded by 0.19 to 0.25 F1.

Why use a cascade instead of off-the-shelf NER or cloud LLMs?

Local NER systems preserve governance but tend to over-redact curricular terms, the paper notes, and commercial LLMs can handle ambiguity but require sending student data to third parties. The cascade seeks a middle path: use a recall-first proposer to capture all possible sensitive spans, then make a context-sensitive binary decision locally. The reported numbers show the strongest local configuration outperforming both a same-family LLM-only baseline (0.767) and a commercial API (0.706) while maintaining local execution, suggesting the problem formulation—recall-first candidate generation plus context-aware review—can matter more than simply scaling or outsourcing models.

What to watch

Look for published code, data, or replication artifacts tied to arXiv:2606.18372 and for any follow-up evaluations beyond the two large tutoring platforms used here. A clear signal that the approach generalizes would be replication on transcripts from different subjects or institutions and availability of the cascade components for local deployment.

Submitted on 16 Jun 2026, the paper appears under Computation and Language and Artificial Intelligence (cs.CL; cs.AI) on arXiv.