Evidence item

An ICLR 2026 paper describes standard LLM deployments as having fixed weights during inference but presents test-time training that updates fast-weight parameters online, so parameter evolution is possible in specialized or experimental systems. DIRECT_TEXT

Authenticity UNVERIFIED · extraction UNVERIFIED · independence group none (unreviewed) · created seq 15349

Source: In-Place Test-Time Training · locator {} · excerpt hash sha256:120a56820cfab2379f4f38f3196249961ba2241bf8d5769a647b85cb78cf13ed

The static “train then deploy” paradigm ... kept fixed during inference. ... Test-Time Training (TTT) offers ... updating a subset of model parameters (fast weights) at inference time ... In-Place TTT treats the final projection matrix of ... MLP blocks as its adaptable fast weights.