Research · MPP Labs
Research that ships into production.
MPP Labs is the applied R&D center behind MPP Insights’ products — research that ends up in production, not research for its own sake. Our founder’s published work on planning information-gathering actions for agentic systems is the idea behind Lizardata; the data-centric architecture behind MPP BI came out of the same lab.
Read the original paper at dl.acm.orgFig. 01 · Current programGrant-funded · active
Cross-distribution capability transfer in language modelsHow model capabilities transfer across languages — including Armenian — domains, and modalities, with direct application to low-resource environments such as the Armenian ASR case study.Read the research
Fig. 02 · Focus areas
What we study
5 areas · 1 program
01Focus areaVerifier & reward engineeringDesigning the checks that decide whether a model's output is actually correct. Our working hypothesis: that's the real limit on capability transfer, not model size.↳ Part of the current program02Focus areaMechanistic analysis of transferLooking inside the model to see whether a new skill lives in a few weights or spreads across the whole network — the answer decides which adaptation method wins.↳ Part of the current program03Focus areaComparative adaptationTesting fine-tuning, low-rank adapters, model merging, and reinforcement learning against each other, under the same compute budget.↳ Part of the current program04Focus areaEvaluation methodologySeparating four things most benchmarks blur together: did it learn the new task, did it keep the old one, does it sound right, and does it hold up.↳ Part of the current program05Focus areaData efficiencyGetting more out of every training example — synthetic data, generated tasks, and knowing when a small curated dataset beats a big noisy one.↳ Part of the current program
Fig. 03 · Innovation approach
How we work
Principle · 01Research-driven engineeringEvery production system traces back to a question we actually investigated first.
Principle · 02Production-first mindsetResearch that doesn't ship is a paper, not a product — we build for the second outcome.
Principle · 03Continuous improvementLab Notes exist because the work is published as we do it, not as a retrospective.
Principle · 04Cross-team collaborationThe researchers and the engineers shipping the product are the same people.