Mrinmaya Sachan

ETH Zurich · Switzerland

Unclaimed · AI-structured profileORCID OpenAlex

NLP researcher at ETH Zurich focused on LLM evaluation, machine translation benchmarking, and LLM-based educational AI

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

h-index · OpenAlex

Bibliometric figures come from OpenAlex (2026-10-16) and cover the author's entire public record, not just the papers listed here. Reproducibility signal comes from Evolve's own reviews.

Research profile

Output—

Not scored: total publication counts and venue-level productivity are not available in this snapshot, so cumulative output cannot be computed. This axis would measure career-wide production volume and how it is distributed across venues and authorship roles.

Impact—

Not scored: citation totals and field-normalized citation percentiles are unavailable here (only raw h-index 24 and i10-index 62 are recorded), so citation influence is not estimated. This axis would measure how widely and how favorably the work is cited relative to field and year.

Focus0/100

Scores 0/100 for topical concentration, and the raw spread bears this out: the field distribution splits evenly at one third each across machine translation evaluation, educational AI with LLM-based student simulation, and LLM evaluation, while topic counts span five clusters from Topic Modeling (317) and NLP Techniques (225) down to Intelligent Tutoring Systems (83), AI in Education (70) and Explainable AI (43) — breadth within NLP rather than one narrow thread.

Collaboration83/100

Scores 83/100. The snapshot records 62 distinct coauthors in the recent window and a recurring core of twelve frequent coauthors (Heejin Do, Vilém Zouhar, Niyati Bafna, Mukund Choudhary, Maike Züfle, Sara Rajaee, Pinzhen Chen, Jannis Vamvas, Sara Papi, Ona de Gibert, Bhavitvya Malik, Eliya Habba). In the three cataloged papers he is last author twice and middle author once, a placement typical of a senior supervising author in a wide network.

Activity100/100

Scores 100/100. Recent activity is the standout signal: 251 works in the last three years, with annual indexed counts climbing from 29 (2021) and 63 (2022) to 78 (2023), 86 (2024) and 111 (2025), and 54 already indexed in 2026 — a steep, sustained ramp consistent with running a large, continuously producing group.

Reproducibility70/100

Scores 70/100, the mean over three reproducibility-assessed papers. Readiness in the catalog is 'high' for SWIM (proficiency-conditioned student writing simulation) and 'medium' for both the Knowledge Space Theory study and the Last Translation Benchmark, whose verdicts flag annotation noise, absent significance tests, and unquantified LLM-based verifiers as limits on replication confidence.

Research summary

Mrinmaya Sachan is a natural language processing researcher at ETH Zurich whose recent work concentrates on evaluation in three registers: evaluating large language models themselves, evaluating machine translation, and building educational AI in which LLMs simulate learners. The OpenAlex snapshot (keyed to ORCID 0000-0001-8787-8681) records an h-index of 24, an i10-index of 62, 251 works over the last three years, and 62 distinct coauthors. Annual indexed counts rise steeply — from single digits each year across 2010–2020 to 29 (2021), 63 (2022), 78 (2023), 86 (2024) and 111 (2025), with 54 already indexed in 2026 — a ramp consistent with a large active group. Topical load is broad: Topic Modeling (317) and NLP Techniques (225) lead, followed by Intelligent Tutoring Systems and Adaptive Learning (83), AI in Education (70) and Explainable AI (43), while the field distribution splits evenly, one third each, across machine translation evaluation, educational AI with LLM-based student simulation, and LLM evaluation. The three papers Evolve has reviewed — all 2026 arXiv preprints — cast him as last author on the Knowledge Space Theory study and on SWIM, and a middle author on the Last Translation Benchmark. Reproducibility assessments average 70 across those three papers.

Collaboration pattern

The recent window lists 62 distinct coauthors, and a recurring core — Heejin Do, Vilém Zouhar, Niyati Bafna, Mukund Choudhary, Maike Züfle, Sara Rajaee, Pinzhen Chen, Jannis Vamvas, Sara Papi, Ona de Gibert, Bhavitvya Malik and Eliya Habba — spans both the translation-evaluation and educational-AI strands of the catalog. Across the three reviewed papers Sachan sits last author twice and middle author once, a pattern consistent with a senior supervisory role within a broad, lab-scale collaboration network rather than small two-person partnerships.

Research keywords

LLM evaluationMachine translation evaluationLLM-based student simulationKnowledge Space TheoryMathematical reasoningKnowledge tracingBenchmark constructionLLM-as-a-judgeAutomated essay scoringControllable text generationMultilingual evaluationIntelligent tutoring systems

Highlights

Testing whether LLM mathematical knowledge has human-like structure
A 2026 arXiv study with Sachan as last author applies Knowledge Space Theory to LLM responses in mathematical reasoning and finds that the models' knowledge lacks the prerequisite structure characteristic of human learners. Evolve judges the KST framing novel and persuasive, while cautioning that annotation noise and the absence of significance tests temper the conclusions.
Community-built 'Last Translation Benchmark' stress-tests frontier MT
This community-constructed machine translation benchmark (2026, arXiv) pairs 3456 hard translation exceptions with rule-based verifiers that succeed in separating frontier models from human rankings. Evolve calls it solid despite relying on LLM verifiers whose reliability is not quantified; Sachan appears as a middle-position author.
Simulating student writing at controllable proficiency levels
SWIM (2026, arXiv, last author) generates student essays conditioned on trait scores, finding plain prompting weak, supervised fine-tuning strong, and reinforcement learning with an automated-essay-scoring reward (GRPO) best. Evolve deems the approach credible but dependent on an AES proxy, and rates its reproducibility readiness high.

Timeline

2026Last-author arXiv study applies Knowledge Space Theory to LLM mathematical reasoning, finding no human-like prerequisite structure.
2026Co-authors the community-built Last Translation Benchmark, pairing 3456 hard exceptions with rule-based verifiers.
2026Last-author SWIM paper on proficiency-conditioned student writing simulation, with AES-rewarded RL (GRPO) as the strongest method.

Papers in the Evolve catalog

Frequent co-authors

Heejin Do×2 Vilém Zouhar×1 Niyati Bafna×1 Mukund Choudhary×1 Maike Züfle×1 Sara Rajaee×1 Pinzhen Chen×1 Jannis Vamvas×1 Sara Papi×1 Ona de Gibert×1 Bhavitvya Malik×1 Eliya Habba×1

About this profile

This profile is AI-structured by Evolve.Pub from public bibliographic data (OpenAlex) and Evolve's own paper reviews. It is not written or endorsed by the researcher unless the profile is claimed. Radar dimensions are computed from real bibliometrics with fixed formulas — raw figures are shown alongside; missing dimensions mean insufficient data, never zero. Errors? Claim the profile to correct or hide it.

Bibliometric figures (h-index 24, i10-index 62, annual counts, topics, coauthor counts) come from an OpenAlex snapshot keyed to ORCID 0000-0001-8787-8681 and are not independently verified.
Output and impact dimension scores are absent because publication and citation totals were unavailable; no values are estimated in their place.
The Evolve review set covers only three 2026 arXiv papers, and the reproducibility score (70) rests on those three assessments alone.
The 2026 annual count (54) is a partial-year figure, and indexed works may include items where Sachan is neither first nor corresponding author.