TTT-Code: Test-Time Training for Local LLMs

Undergraduate research on test-time training: letting small, locally run language models adapt to long contexts as they read them.

Large language models usually freeze their weights after training. Test-time training (TTT) lets a model keep learning while it reads: it takes small gradient steps on the context in front of it, which can help with long documents and codebases.

This undergraduate research project, started in Spring 2026, explores TTT methods for small, locally run models, which are cheaper, more private, and use far less energy than frontier models in the cloud. We build on recent work such as In-Place TTT and end-to-end TTT, and evaluate on long-context benchmarks such as HELMET and LongBench.