1 of 1

LoRA on the Go:

Instance-level Dynamic LoRA Selection and Merging

Seungeon Lee, Soumi Das, Manish Gupta and Krishna P. Gummadi

Research Question

Why training & data-free?

Why combine LoRAs?

  • No single LoRA fits all real inputs�- Real-world queries are unseen or span multiple domains.

  • One input often needs several LoRAs�- combining related adapters helps more than any single one.

Given a pool of pre-trained LoRAs, how can we select and combine the right ones for each input, without any training or per-task data?

Key Observation

LoGo Workflow

  1. Run one probe pass with all adapters attached
  2. Score each adapter by its activation signal
  3. Select top-k adapters and merge by signal-weighted sum

Experiments

  • Evolving LoRA pool�- Adapters are added or deprecated constantly, so any trained router must be re-trained.
  • Per-LoRA data maintenance is inefficient�- Selection by retrieval/similarity needs each adapter's training data on hand.
  • Similar tasks activate adapters similarly: clear block-diagonal pattern (red boxes).

  • Simple signals (norm / entropy) capture this relevance.
  • We train 260 LoRAs* per model on FLAN-v2 tasks, across 3 models (LLaMA-3.1-8B, Qwen-2.5-7B, DeepSeek-7B).
  • We evaluate on 27 datasets spanning 5 task categories.

* Trained LoRAs released on HuggingFace (huggingface.co/archon159/LoGo-loras-collection)

Evaluation Setting

NLP Task Results

Results on LLaMA-3.1-8B backbone. The reported results are averaged over each dataset category. The best results are in bold, the second-best are underlined.

  • LoGo outperforms baselines on most categories, and remains comparable on the rest.

Github

Arxiv

Analysis

  • LoGo assigns larger weights to more similar tasks (left).

  • Probe-pass overhead is amortized over long generations (right).