← AI Techniques Radar
Training & fine-tuning

Demonstration-budget sweeping for VLA finetuning

Adapt a pretrained vision-language-action model by systematically sweeping the number of target-setting demonstrations and the finetuning steps, then select the best-performing demonstration/training budget for each suite.

Discovered 2026-11-10 · Sources checked 2026-11-10

This summary was generated with AI assistance.

Why it matters

Full-data or zero-shot comparisons can hide how many demonstrations are actually needed for effective adaptation; sweeping budgets exposes sample-efficiency differences between model variants and training settings.

What changed

Instead of comparing only zero-shot versus full-data finetuning, the study varies both the demonstration count and the number of training steps, and evaluates multiple VLA models across multiple suites. It also includes an ablation where the vision-language module (VLM) is frozen during finetuning to test sensitivity to joint training.

Where it helps

  • Choosing a target-demonstration budget for VLA adaptation in a new environment
  • Comparing sample efficiency across multiple pretrained VLA model families
  • Planning finetuning compute when labels/demonstrations are limited
  • Running ablations (e.g., freezing the VLM) to diagnose where adaptation signal is needed

Evidence & reproduction

Reproduction not evaluated. A linked paper or repository does not establish independent reproduction.

Limitations & open questions

Reported findings are specific to the stated VLA variants and benchmark suites, and the study notes that held-out validation loss can track within-suite success while not fully reflecting cross-model differences in sample efficiency.

Sources