analiza aiqbench
The paper tackles catastrophic forgetting in sequential LLM fine-tuning by freezing model weights and learning low-rank interventions on hidden representations, with a KL penalty to protect earlier tasks. Its reported gains over LoRA-based baselines come from benchmark results across several model scales, so practitioners should check how well they hold under their own task orderings and data. The approach is worth watching as a parameter-efficient alternative to adapter-stacking for continual learning, though it still needs independent replication.
CRAFT: Forgetting-Aware Intervention-Based Adaptation for Continual Learning
arXiv:2605.05732v3 Announce Type: replace-cross Abstract: Large language models (LLMs) can learn new tasks through fine-tuning, but learning tasks sequentially often leads to catastrophic forgetting of earlier ones. This paper proposes CRAFT, a continual learning framework that keeps model weights frozen and instead learns low-rank interventions on hidden representations. Similar tasks share an intervention to help knowledge transfer, with similarity judged by how each task alters the model's output distribution rather than by a learned gating network. Within a shared intervention, a Kullback-Leibler (KL) penalty keeps each new task close to the intervention's existing behavior, which limits forgetting of earlier tasks, while a small, task-private intervention in the deepest layers preserves dedicated capacity that later tasks cannot overwrite. The learned interventions also make it possible to identify the task at test time, so task identity labels are not needed. Across multiple benchmarks and model scales, CRAFT improves overall performance and reduces forgetting compared with strong LoRA-based methods, while remaining robust to task ordering.
Abstract:Large language models (LLMs) can learn new tasks through fine-tuning, but learning tasks sequentially often leads to catastrophic forgetting of earlier ones. This paper proposes CRAFT, a continual learning framework that keeps model weights frozen and instead learns low-rank interventions on hidden representations. Similar tasks share an intervention to help knowledge transfer, with similarity judged by how each task alters the model's output distribution rather than by a learned gating network. Within a shared intervention, a Kullback-Leibler (KL) penalty keeps each new task close to the intervention's existing behavior, which limits forgetting of earlier tasks, while a small, task-private intervention in the deepest layers preserves dedicated capacity that later tasks cannot overwrite. The learned interventions also make it possible to identify the task at test time, so task identity labels are not needed. Across multiple benchmarks and model scales, CRAFT improves overall performance and reduces forgetting compared with strong LoRA-based methods, while remaining robust to task ordering.
Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
arXiv:2605.05732 [cs.LG]
(or arXiv:2605.05732v3 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2605.05732 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Text integral sub licența deschisă a publicației · 252 cuvinte
Subiecte conexe
What's Continual Learning, and Why Might We Expect To See It In Advanced LLM Agents?
LessWrong · 12 iun. 2026
Reflections on unlearning and inoculation
LessWrong · 20 sept. 2026
Game Design Creativity evals
LessWrong · 12 oct. 2026
2026.5: We're on the same frequency now 📡
Home Assistant · 6 mai 2026