Micro-TTT: Weight-Level Evolution Explained

Micro-TTT is test-time training executed at inference scale. While the model generates each token, we run a lightweight optimization loop on a sliding window of the current conversation plus any attached Vidya Files. The update targets only the most sensitive parameters, identified by a per-layer Fisher diagonal computed on-the-fly.

The math is brutal and efficient. For a given input sequence x and target continuation y, we compute a masked gradient update where the mask is a binary tensor derived from the Fisher information threshold. Only 0.8% of the 109B parameters move on average. The resulting delta is quantized to 4-bit, sparsified, and stored as a .jdelta.pt file — 27 MB on disk, under 5 ms to merge at inference time.

This is not fine-tuning. Fine-tuning requires a full checkpoint, hours of GPU time, and a new deployment. Micro-TTT happens inside the forward pass, between two consecutive tokens, with zero downtime.

The Numbers

We validated the approach on a held-out benchmark of long-context technical documents. After a single pass of user data, perplexity dropped 75.8% compared to the frozen baseline. The same metric on RAG-augmented frozen models improved only 9%. The difference is not marginal. It is the difference between reading the manual and rewriting your own cortex.

Because the delta is low-rank and subspace-localized, inference cost remains O(1) even at 10M token effective context. The model does not re-read history. It has absorbed it.

What Developers Report

Developers who have used the alpha M.A.I. Code CLI report the same pattern: first session feels like any other LLM; third session feels like pair-programming with someone who has lived inside the codebase for months. That is not prompt engineering. That is weight-level memory.

The frozen models will keep scaling parameters and hoping size fixes the problem. We scale plasticity instead.