Tatiana Otbetkina
Tatiana Otbetkina has a total of 8 co-authors. Here are the top co-authors by number of shared publications:
- Aleksandr Volkov
Educational and Methodological Department, National University of Science & Technology (MISIS), 119049 Moscow, Russia
3 shared publications - Dmitriy Istomin
Hats Lab, 68 Circular Road, #02-01, Singapore 049422, Singapore
2 shared publications - Taras Pustovoy
Hats Lab, 68 Circular Road, #02-01, Singapore 049422, Singapore
2 shared publications - Egor I. Tuzharov
MSU, Moscow, Russian Federation
1 shared publication - Maksim Malykh
MSU, Moscow, Moscow, Russian Federation and MIRAI, Moscow, Russian Federation
1 shared publication - Mariia Kovalchuk
MSU, Moscow, Moscow, Russian Federation and MIRAI, Moscow, Russian Federation
1 shared publication
Latest Publications
- Open Access
- 9 October 2026
- Published by MDPI AG in Machine Learning and Knowledge Extraction
- Vol. 8 (10) , 322
From Retrieval to Evidence: A Task-Adaptive Two-Layer Knowledge Architecture for LLM Agents
Retrieval-augmented generation and structured agent memory are commonly treated as competing ways to ground large language models. Our results show that their relative value changes with the task and that neither retrieved similarity nor graph association is sufficient evidence for an answer. We present a task-adaptive two-layer methodology combining a compact, reviewed knowledge layer for orientation and cross-source synthesis with a versioned evidence layer for exact lookup and claim support. A router selects the representation, while an independent admission stage checks named entities, time, source role, contradiction, and provenance. A controlled course-grounding study reveals a task-conditional crossover within Course 2: the distilled Body of Knowledge (BoK) trails plain RAG by 1.07 points on lookup and leads by 0.62 points on synthesis. Across a 72-question three-course lookup panel, RAG leads by 2.17 points; Course 1 synthesis is near parity. Complementary, position-debiased LLM-judged mastery comparisons favor the BoK on both courses. The standing representation is reduced 22–35-fold. An observational production case over a 289,642-chunk evidence base illustrates why candidate discovery, packet selection, and evidence admission require separate gates. The contribution is a tested method for assigning retrieval, structure, and evidence admission distinct responsibilities within one grounded agent architecture.
- 9 October 2026
- Published by MDPI AG in Machine Learning and Knowledge Extraction
- Vol. 8 (10) , 322
Open AccessFrom Retrieval to Evidence: A Task-Adaptive Two-Layer Knowledge Architecture for LLM Agents
Retrieval-augmented generation and structured agent memory are commonly treated as competing ways to ground large language models. Our results show that their relative value changes with the task and that neither retrieved similarity nor graph association is sufficient evidence for an answer. We present a task-adaptive two-layer methodology combining a compact, reviewed knowledge layer for orientation and cross-source synthesis with a versioned evidence layer for exact lookup and claim support. A router selects the representation, while an independent admission stage checks named entities, time, source role, contradiction, and provenance. A controlled course-grounding study reveals a task-conditional crossover within Course 2: the distilled Body of Knowledge (BoK) trails plain RAG by 1.07 points on lookup and leads by 0.62 points on synthesis. Across a 72-question three-course lookup panel, RAG leads by 2.17 points; Course 1 synthesis is near parity. Complementary, position-debiased LLM-judged mastery comparisons favor the BoK on both courses. The standing representation is reduced 22–35-fold. An observational production case over a 289,642-chunk evidence base illustrates why candidate discovery, packet selection, and evidence admission require separate gates. The contribution is a tested method for assigning retrieval, structure, and evidence admission distinct responsibilities within one grounded agent architecture.
- Open Access
- 8 August 2026
- Published by Association for Computing Machinery (ACM)
- p. 13207-13212
Agent-Didactic Spaces: Mining Multi-Stakeholder Trade-Offs in Educational Pathway Data
- 8 August 2026
- Published by Association for Computing Machinery (ACM)
- p. 13207-13212
Open AccessAgent-Didactic Spaces: Mining Multi-Stakeholder Trade-Offs in Educational Pathway Data
- Open Access
- 13 July 2026
- Published by Springer Nature in Research Square
Mining AI-Assisted Course Design Workflows at Production Scale
We mine CourseFactory, a deployed AI-assisted course-design system, as a multi-surface workflow problem. From a production trace spanning 2023–2026 (197,858 AI calls, 14,598 course versions, 676,417 generated items, 6,323 users), we extract four bounded workflow surfaces using a privacy-preserving protocol that excludes raw prompts, generated text, and calendar timestamps. Version-level structural-quality triage reaches AUPRC 0.89 and saves approximately 20% review effort over random queuing. Item-to-assignment routing achieves a 2.74× AUPRC lift (0.28 vs. 0.10 baseline, AUROC 0.81) on unseen projects. Revision loops identify high-iteration projects for operational review. Task-state conformance against the designed flow reaches 66.8% over the multi-state tail. Feedback and rejected-request surfaces remain diagnostic; withheld input-prompt text adds no measurable signal under lexical, n-gram, and latent-semantic ablations. We formalize the evaluation protocol (Algorithm 1) and release the de-identified dataset, making this the first production-scale workflow-mining study of AI-assisted course design.
- 13 July 2026
- Published by Springer Nature in Research Square
Open AccessMining AI-Assisted Course Design Workflows at Production Scale
We mine CourseFactory, a deployed AI-assisted course-design system, as a multi-surface workflow problem. From a production trace spanning 2023–2026 (197,858 AI calls, 14,598 course versions, 676,417 generated items, 6,323 users), we extract four bounded workflow surfaces using a privacy-preserving protocol that excludes raw prompts, generated text, and calendar timestamps. Version-level structural-quality triage reaches AUPRC 0.89 and saves approximately 20% review effort over random queuing. Item-to-assignment routing achieves a 2.74× AUPRC lift (0.28 vs. 0.10 baseline, AUROC 0.81) on unseen projects. Revision loops identify high-iteration projects for operational review. Task-state conformance against the designed flow reaches 66.8% over the multi-state tail. Feedback and rejected-request surfaces remain diagnostic; withheld input-prompt text adds no measurable signal under lexical, n-gram, and latent-semantic ablations. We formalize the evaluation protocol (Algorithm 1) and release the de-identified dataset, making this the first production-scale workflow-mining study of AI-assisted course design.