Layers important to LoRA adaptation varied across tasks
Source check: Source checked. Not independent scientific confirmation.
Finding: In an intervention study on fine-tuned Qwen2.5-0.5B, the seven layers most important to adaptation had a mean pairwise overlap of 39.5% by the Jaccard measure across four tasks. Only the first three layers appeared in every top-seven list.
Technology: Qwen2.5-0.5B; low-rank adaptation; activation path patching
Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.
- AI contribution
- Path patching intervened on learned LoRA activations to measure each layer’s contribution on four datasets.
- Test and result
- The intervention compares layer-level learned contributions, then ranks the layers separately for the four task datasets.
- Limitation / still unknown
- The overlap depends on selected datasets, model, rankings and layer cutoff. It does not establish identical sparsity across architectures.
Evidence sources and annotation context
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
Evidence sources: https://arxiv.org/abs/2610.34756v1 · https://arxiv.org/html/2610.34756v1
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: linked separately.
Review date: 21 Nov 2026. Review scope: Qwen2.5-0.5B four-dataset path-patching Methods and Figure 3.
No follow-up recorded; this is not a claim that no follow-up exists. Follow-up review scope and date are not recorded. No source-supported relationships recorded.
Editorial assessment: Impact 2/5 · Validation 3/5 · Novelty 2/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.