Lab Notes
Research explainers, benchmark notes, project retrospectives, and technical essays from DeCLaRe Lab.
2026
Paper explainer · May 2026
Toward Efficient Data-Centric Training, Part II
PODS asks how much data a model should see over training, turning selection ratio into a dynamic scheduling signal.
Read note
Paper explainer · May 2026
Toward Efficient Data-Centric Training, Part I
Data Agent learns a training-aware policy for selecting useful data as the target model evolves.
Read note
Paper explainer · May 2026
δ-mem: Efficient Online Memory for Large Language Models
A note on why LLM memory should be more than long context, and how a compact online state can directly modulate attention.
Read note