← News

ResearchSources: arXiv · 19 Nov 2026

CRAM targets forgetting and parameter cost in multimodal continual instruction tuning

Researchers propose CRAM, a mixture-of-experts approach that isolates task-specific patterns, reuses existing experts, and adds only the parameters a new task needs.

In plain words

Multimodal AI models that handle both images and language tend to forget old skills when trained on new tasks. CRAM is a proposed method that stores each task's patterns in separate modules and reuses existing ones when possible, adding only the parameters a new task requires. The authors report it outperformed existing methods on their benchmarks.

What was announced

The arXiv paper frames the problem as a trade-off in Multimodal Continual Instruction Tuning, the practice of extending vision-language models with new tasks over time. Updating all tasks with one shared parameter set makes heterogeneous tasks compete, which the authors say causes forgetting of learned capabilities, while giving every new task its own dedicated modules avoids interference but limits parameter efficiency over long task streams.

CRAM, short for Centroid-Routing and Adaptive MoE, addresses this with three mechanisms described in the abstract. It isolates task-specific patterns into independent modules to mitigate catastrophic forgetting. Adaptive-rank instantiation measures the gap between what existing experts can do and what a new task demands, allocating only the necessary parameters. Centroid-guided routing activates existing experts' capabilities for reuse, and an orthogonality penalty keeps new updates in task-specific directions so general capability is not re-learned. The authors report extensive experiments across diverse benchmarks showing superiority over existing methods, and code is available on GitHub.

What it does not show

  • The evidence is the paper's abstract only; benchmark names, metrics and comparison figures are not given, and the reported results are the authors' own and not independently confirmed.

For security and IT teams

Access
The authors state code is available at the GitHub repository LAMDA-CL/EMNLP2026-CRAM.

Sources

Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 20 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.