DynaTokens accepted at EMNLP: dynamic adaptation tokens for continual video-language learning

flosalim · x · 2026-09-22

The author announced that DynaTokens was accepted to the EMNLP 2026 main conference, tackling a pain point in continual learning for video-language models: adapting to new tasks without storing a separate set of adaptation tokens per task.

Core idea: instead of storing task-specific prompts, the model learns to dynamically generate the right adaptation tokens when needed, controlling token dynamics.
Implementation: frozen LLaMA-2-7B backbone + CLIP ViT-L/14, with LLaMA-Adapter for efficient video-language adaptation — no fine-tuning of the LLM backbone.
Why it matters: avoids accumulating task-specific prompts over time, offering a new path for continual multi-task video-language adaptation.

Original post →

DynaTokens accepted at EMNLP: dynamic adaptation tokens for continual video-language learning

More from Multimodal