NASA’s new dark energy space telescope can also detect killer asteroids
technologyreview.com·3h ago
AI Research
arXiv CS.AI·May 26·5 min read
PALoRA: Projection-Adaptive LoRA for Preserving Reasoning in Large Language Models
TL;DR
Updating Large Language Models (LLMs) with new knowledge often compromises their reasoning abilities. PALoRA, a two-stage framework, uses Singular Value Fine-Tuning (SVF) to identify critical reasoning components and Low-Rank Adaptation (LoRA) to inject new knowledge without disrupting these components.
✦ Why It Matters
Engineers can leverage PALoRA to efficiently update LLMs while preserving critical reasoning skills.
Key Takeaways
- Use PALoRA for efficient knowledge updates in LLMs.
- Train SVF experts to identify critical reasoning components.
- Apply structural orthogonality to minimize interference.
- Achieve high reasoning retention with minimal parameter overhead.
How It Works
PALoRA operates in two stages. First, it trains a Singular Value Fine-Tuning (SVF) expert on a reasoning dataset, which helps identify the singular scaling vector that represents critical reasoning components.
In the second stage, it uses Low-Rank Adaptation (LoRA) to inject new factual knowledge while ensuring that updates do not interfere with the identified reasoning subspace, thus preserving the model's reasoning capabilities.