A slot-based interface with a parameter-free memory links recurring object observations into persistent narratives and feeds a frozen LLM just 144 visual tokens.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Persistent Object Narratives for Token-Efficient Video Language Models
A slot-based interface with a parameter-free memory links recurring object observations into persistent narratives and feeds a frozen LLM just 144 visual tokens.