Reproduce “Adapt-$\infty$: Scalable Continual Multimodal Instruction Tuning via Dynamic Data Selection” on OpenResearch

About this paper

Adyasha Maharana, Jaehong Yoon, Tianlong Chen, Mohit Bansal

Adapt-∞ selects a limited set from combined old and new multimodal instruction data, grouping examples by gradients from the current model and adapting selection to inferred skill groups. At a budget of 25K samples per timestep, it reports 102.3% Relative Gain and 0.9% Forgetting Rate, versus 95.3% and 2.1% for random selection. Relative Gain averages results against each skill’s best sequential-training result, not every skill’s outcome.

Key insights

• Gradient vectors from essential model layers are more effective for clustering samples by 'skill' than hidden layer outputs, leading to better pseudo-task identification.
• Dynamically selecting the optimal sample importance scoring function for individual pseudo-task clusters significantly outperforms relying on a single scoring function across diverse tasks.
• Proactive, semantic-based pruning of the entire data pool is crucial for preventing indefinite data growth and ensuring the long-term computational feasibility of lifelong instruction tuning.

Summary by alphaXiv.