Research
Publications
Peer-reviewed papers, technical reports, and notes from the Userlite research team. We publish the methods behind our models — the wins and the open problems both.
2026 · 06
Legible Chains: supervising reasoning without collapsing it
A method for keeping intermediate reasoning faithful and human-auditable.
Interpretability
2026 · 05
Lumen 3: technical report
Architecture, training data, and evaluations for our flagship model.
Models
2026 · 03
Constitutional preference tuning at scale
Aligning model behavior to written principles with minimal human labels.
Alignment
2026 · 01
Calibrated refusals: knowing when not to answer
Teaching models to express uncertainty instead of confident error.
Safety
2025 · 11
Sparse retrieval memory for million-token context
An attention-efficient memory that scales recall without scaling cost.
Systems
2025 · 09
Evaluating uplift: measuring what a model does for a person
A benchmark suite for learning, decision quality, and creative output.
Evaluation
2025 · 06
Lumen Nano: distillation for private, on-device inference
Recovering frontier behavior in a model small enough to run at the edge.
Models