{"resourceId":"lee-scala-proactive-tutor-acl-2026","versions":[{"version":"external-1564d7fdec9f9da695f0dea4c4cb1772306ece8f1bf15f99ab8c0b12b93e167b","resource":{"id":"lee-scala-proactive-tutor-acl-2026","title":"ACL Industry: SCALA proactive LLM tutor in 1,508-student Python course boosts help-seeking vs prior channels","organization":"KAIST / Korea University / ClassDuo (Lee, Youn, Lee, Im, Kim, Yoo)","sector":"Higher education computer science / AI teaching assistants","geography":"South Korea (large undergraduate Python course; transferable design lessons with limits)","publishedAt":"2026","publicationDate":null,"eventDate":null,"sourceName":"ACL 2026 Industry Track (Anthology)","sourceLabel":"Peer-reviewed industry deployment paper on proactive LLM tutoring","sourceUrl":"https://aclanthology.org/2026.acl-industry.107.pdf","evidenceClass":"academic-research","outcomeClass":"emerging","topics":["knowledge-work","developers-agents","operating-model"],"finding":"SCALA, a proactive student-centered AI learning assistant, was deployed for a semester in an undergraduate Python course with total enrollment of 1,508 students. Predictive query management pre-generates likely lecture questions; students also chat interactively. The system processed on the order of 2,600 sessions—reported as a 124× increase over prior support channels—and SCALA achieved a 68.5% win rate versus alternatives such as GPT-4o in student preference comparisons. The paper emphasizes engagement and preference, not graded learning RCTs.","sledRelevance":"Transferable large-course tutoring operations pattern for U.S. HE: proactive, course-timed question surfacing can massively raise help-seeking relative to reactive office-hour channels—useful alongside causal learning evidence from other sources.","evidence":"Semester deployment metrics, predictive-query overlap analyses, and preference win-rate table in ACL 2026 Industry paper PDF; Korean university CS context.","architectureImplications":"Generate lecture-aligned predictive Q&A ahead of class; unify proactive cards and multi-turn tutoring in one UI; fine-tune or ground on course QA; measure session depth, not only unique users.","governanceImplications":"Disclose AI TA scope to students; keep human TAs for edge cases; avoid equating preference wins with learning gains in public claims.","securityPrivacyImplications":"Large query volumes create sensitive coursework data; retain under institutional control and limit vendor secondary use.","caveats":"Not a randomized learning-outcomes study; South Korea CS sample; preference vs GPT-4o is not mastery; transfer to U.S. advising/retention untested.","streamIds":["student-success"],"roles":{"sales":"For large-enrollment STEM student-success and academic-technology buyers, Lee et al.’s ACL 2026 Industry deployment of SCALA in a 1,508-student Python course shows proactive predictive queries plus interactive tutoring can multiply help-seeking sessions versus prior channels and win student preference against general GPT-4o responses. Sell course-timed, proactive AI TAs that surface likely questions before lecture—not ambient ChatGPT widgets. Be clear that this evidence is engagement/preference, not grade RCTs, and pair with local learning evaluation.","engineering":"Engineering should implement predictive query generation grounded in upcoming lecture materials, a shared UI for proactive cards and chat, and telemetry for session counts and multi-turn depth. Prerequisites include course content pipelines and TA escalation paths. Constraints: Korean CS deployment; no causal grade estimates here; model preference may shift with foundation-model upgrades. Security: institutional logging of 2,600+ sessions with FERPA-equivalent controls for U.S. ports.","delivery":"Delivery is a large-course AI TA pilot: pre-generate weekly predictive Q&A, staff human escalation, monitor session growth and preference samples, and add an outcome study before claiming learning gains. CS faculty own content; academic technology owns platform; student success watches equity of access. Acceptance criteria include documented proactive coverage and human fallback SLAs. Risks include mistaking 124× session growth for mastery and under-serving students who never open the tool."},"roleDetails":{"sales":{"customerProblem":"Large intro courses cannot staff enough personalized help, so students either stall or turn to ungrounded consumer GenAI.","stakeholders":["CS/STEM faculty","academic technology","undergraduate education","tutoring center"],"discoveryQuestions":["Can we generate lecture-aligned predictive questions before each module?","What human TA path handles out-of-scope or integrity-sensitive queries?","Which learning metrics will complement session and preference KPIs?"],"valueHypothesis":"Proactive, course-grounded AI TAs can raise help-seeking at scale versus reactive channels when content is lecture-aligned.","potentialEngagement":"One-term large-course pilot with predictive Q&A, chat, escalation, and a parallel mini learning evaluation.","unsupportedClaims":["Guaranteed grade or retention lifts from session growth","That preference over GPT-4o equals deeper learning","Automatic transfer from Korean CS to U.S. gen-ed courses"]},"engineering":{"fit":"Fit for high-enrollment courses with stable weekly content and capacity for TA escalation.","architectureIntegration":"Content pipeline → predictive multi-agent Q&A → student UI → tutoring model → analytics warehouse.","prerequisites":["Lecture/material feed","Identity-aware course roster","Human escalation queue"],"constraints":["Deployment evidence without RCT grades","International context","Heavy inference cost at 1,500-student scale"],"security":"Course-tenancy isolation; retention limits; academic-integrity mode for graded contexts.","proofOfValue":["Session and multi-turn rates vs prior support baseline","Predictive-query selection/overlap metrics","Optional blinded preference or learning subsample"]},"delivery":{"work":"Launch proactive AI TA for one large course; operate weekly content refresh; staff escalations; report engagement and a learning subsample.","dependencies":["Faculty content cadence","Cloud inference budget","TA staffing"],"ownership":"Course coordinator with academic technology co-owner.","skillsAdoption":"Orient students to predictive cards vs chat; train TAs on takeover protocol.","governanceCheckpoints":["Pre-term privacy/integrity review","Monthly engagement equity check","End-term learning evaluation decision"],"acceptanceCriteria":["Predictive Q&A live before lectures","Escalation SLA met","Public messaging distinguishes engagement from mastery"],"risks":["Hallucinated proactive answers","Unequal uptake by first-gen/low-access students","Overload from 124× interaction growth without TA capacity"]}},"retrievedAt":"2026-09-26T03:14:25Z","enrichedAt":"2026-09-26T03:21:39Z","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"Proactive cards may help students reluctant to initiate help; ensure WCAG-conformant UI and equivalent non-AI paths for students who opt out.","procurementImplications":"Procure course-grounded proactive tutoring with content-refresh SLAs and evaluation hooks—not generic chatbot seats alone.","operatingModelImplications":"Operating model must fund weekly content generation and human escalation alongside the model.","sourceVerification":{"openedUrl":"https://aclanthology.org/2026.acl-industry.107.pdf","referenceExcerpt":"total enrollment of 1,508 students","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}