mlp-weights

This paper introduces In-Place Test-Time Training, a framework that updates MLP weights in real-time during inference, allowing LLMs to dynamically adapt and handle long contexts up to 128k tokens.

This paper investigates the internal mechanisms of knowledge editing methods ROME and MEMIT, revealing that edits rely on a common functional subspace of weights and suppress rather than overwrite knowledge, explaining why edits fail to propagate to related facts.