Get Started
Home
Topics
Search
Library
Research questionHow can practitioners train compact code embeddings effectively without reliable documentation or costly execution traces?Compact code encoders need semantic supervision, but human-written docstrings can be labor-intensive and inconsistent while execution traces can be costly and dependent on the execution setting.
AI
Information Retrieval
Machine Learning
Natural Language Processing
Research Paper
Small / On-device Models
Technology
Latest papersRecent research connected to this question, newest first.Synthetic Semantic Supervision for Contrastive Code Representation Learning in Small Transformers: An Empirical StudyThe empirical evidence covers contrastive dual-encoder training for C, C++, and Java across eight retrieval, classification, and generation tasks. It compares compact models with pretraining-based, generalist, embedding-specific, and execution-aware alternatives, including matched inference size and pretraining-data comparisons; the reported results do not establish performance beyond these tasks and settings.research paper · Sep 3, 2026
Related questions
How can compact text embedding models improve retrieval and generalization through better training and data quality?How can large language models cut training and inference costs without materially harming accuracy?How can pathology foundation models be compressed to reduce inference cost without sacrificing classification accuracy or reliability?How can we distill capable small language models with fewer training tokens without losing teacher behavior?