We treat a model's weights as frozen the moment training ends, so we adapt the
only way that allows — stuffing examples into the prompt and paying the K/V Cache
Tax in memory and throughput. The 2025–2026 research moves adaptation into the
inference loop instead.

I'll demo four ways to adapt a frozen LLM at inference, live on a single Mac:
three primitives from the literature — Test-Time Training (a discarded in-flight
gradient step), TAO (reward-based tuning, no labels), and δ-mem (online memory,
zero added context) — plus a fourth of my own, Reinforce-or-Retrain (RoR): a
confidence-gated loop that reinforces a confirmed answer and spends a gradient
step only when a human corrects it. Every demo runs at roughly [your measured ms]
per step, off the critical path, with nothing leaving the box. I'll close on why
the cloud's multi-tenant bottleneck makes local adaptation inevitable.

KEY TAKEAWAYS
- Frozen weights are a deployment default, not a limit — gradient, reward, and
memory each adapt a shipped model mid-session.
- Most "adaptation" in production is in-context learning in disguise; moving it
into the loop removes the memory and throughput tax.
- Reinforce-or-Retrain (RoR) — reinforce when right, retrain when wrong — is a
control loop you can lift into your own stack, shared in full.