1 of 9

Progress update presentation: Team 10

Sicheng Lyu, Xinyu Wang

2 of 9

Recap: Improving Low-Bit Quantization for VLA Models

  • Motivation: VLA models are expensive to deploy
  • Low-bit quantization reduces memory and compute
  • But quantization noise can accumulate over long-horizon action generation
  • Existing PTQ methods such as QuantVLA still depend on static calibration scales
  • Static calibration may not generalize well across diffusion timesteps, since different phases can have different activation distributions.

​

3 of 9

Recap: Our Method

  • Start from QuantVLA
  • Improve the DiT action head calibration
  • Key idea: Timestep-Grouped Interface Calibration
  • Split diffusion timesteps into G groups
  • Compute separate trainable (new) ATM / OHB scales for each group
  • Also add trainable Clipping (It was static in original paper)

​

4 of 9

Recap: Experiments

  • Evaluation setup: compare QuantVLA, our method, and full precision (and other baselines) on LIBERO and LIBERO-Plus under W4A8.
  • Test for ablation on our method.
  • Test for efficiency.

5 of 9

What we have completed so far

  • Reproduced the QuantVLA baseline in our setup
  • Built the evaluation pipeline
  • Implemented our timestep-grouped calibration variant
  • Ran the first comparison on LIBERO
  • Collected baseline robustness results on LIBERO-Plus

​

6 of 9

Results on LIBERO

  • Our method matches QuantVLA on average
  • There are category-level trade-offs
  • LIBERO may be too saturated to clearly reveal robustness differences
  • So we use LIBERO-Plus as a more informative benchmark

7 of 9

Results on LIBERO-Plus

  • LIBERO-Plus is a more challenging benchmark
  • Baseline already shows a noticeable gap from full precision
  • Our method on LIBERO-Plus is still in progress because of compute limits

8 of 9

Challenge

  • The released QuantVLA implementation mainly uses fake quantization, making real efficiency gains difficult to measure directly. We therefore focus on accuracy and robustness.
  • Limited time and compute resources restrict the scale of our experiments and ablations.
  • Benchmarks such as LIBERO may be relatively saturated in our setting, making method differences harder to observe.

9 of 9

Updated Timeline

  • Week 1: Complete LIBERO-Plus evaluation for our method; Finalize our method.
  • Week 2: Run ablations on our method. Analyze which perturbation types benefit the most
  • Week 3: Summarize final results and main findings; Prepare final presentation and poster
  • If time allows, run one additional robustness or cross-dataset experiment

​