New LoRA Skills Should Read but Never Write to Preserve Old Abilities
A method adds new AI skills to a model without erasing or rewiring what the model already knows, aiming for a simpler, safer way to grow capabilities over time.
- Publication
- arXiv
- Stage
- Preprint
- What we read
- Summary of the paper
- Authors
- Zeyan Li, Panqi Yang, Qirong Guo, Shengda Zhuo, SIyuan Qiu, Hu Xu, Chun Li, Jianfeng Xu
- Universities and research institutions
- Not yet supplied in verified metadata; the Brief does not guess.
What they did and found
A study shows that new AI skills can be added in a read-only way, preserving existing abilities while folding the new skill into the base model, avoiding extra routing or task-specific heads.
Why it matters
Practically, teams can scale capabilities without frequent retraining, yet some tasks, especially tougher ones, may still challenge the approach.
The method rewrites each adapter into a balanced form that preserves its effect exactly, then links them with a fixed, one-way channel. Only the new skill’s input directions are trained, so older features stay intact.
A hypothetical company could add a sentiment-checking skill to a chat model; it would improve tone detection while keeping the model’s grammar and recall unchanged for existing tasks like scheduling or facts.
Original sources · 1
- New LoRA Skills Should Read but Never Write ↗arXiv · 2026-09-25
Check the original paper for its authors, methods, version and access terms.
What remains uncertain
The technique does not guarantee success on very hard tasks; some benchmarks may still underperform after a single addition.
What is your take?
Ask a question, add useful context or share a different perspective. Keep the conversation respectful and grounded.