Develop adaptive forgetting budgets

Develop adaptive forgetting budgets that adjust dynamically based on task importance or context, rather than remaining fixed across all edits, to improve stability–adaptability trade-offs in gated continual self-editing.

Background

The paper introduces a gated continual self-editing framework that controls catastrophic forgetting by enforcing a budget on update-induced changes using one of three metrics: EM drop, bits increase, or KL divergence. In all experiments, this budget is fixed per run and applied uniformly across sequential LoRA merges.

The authors note that fixed thresholds can be either overly rigid or permissive depending on the edit, task importance, and context. They therefore flag the need for dynamically adjusting budgets to better reflect varying risk tolerances and priorities across edits, which could enhance performance while maintaining reliability.

References

Building on this foundation, several extensions remain open. One direction is the development of adaptive forgetting budgets that adjust dynamically based on task importance or context, rather than remaining fixed across all edits.

— STABLE: Gated Continual Learning for Large Language Models  (2510.16089 - Hoy et al., 17 Oct 2025) in Section 6.3 (Future Work)

Several directions remain open. Most importantly, the information budget is currently specified by a simple global schedule; learning or adapting this budget from the model’s uncertainty, task progress, or observed forgetting could provide finer control over the stability–plasticity trade-off.

— iSDFT: Information-Proximal Self-Distillation for Continual Learning in LLMs  (2609.24646 - Khamis et al., 21 Sep 2026) in Section 6, Conclusions & Future Work