Show HN: PonderTTT – Training-free adaptive compute for LLMs via TTT gating

Hi HN,

This is my first project/paper (independent work; I’m a high school student in Korea).

PonderTTT explores when to run Test-Time Training (TTT) updates instead of updating on every segment.

The gate is training-free: it uses the TTT layer's self-supervised reconstruction loss to decide UPDATE vs SKIP.

A single threshold is calibrated on unlabeled data and adapted online via EMA to maintain a target update rate.

In GPT-2 code language modeling experiments (124M–1.5B), this reached 82–89% Oracle Recovery and improved over Random Skip baselines on some OOD languages (up to 16% lower loss).

Note: v1 experiments are JAX/Flax on GPUs. I'm working on a v2 scale-up to Gemma 3 (TPU).

I'd really appreciate feedback on:

1) whether reconstruction loss is a sensible gating signal in practice,

2) evaluation/baselines you think are important for this setting.