State of Vision-Language-Action (VLA) Research at ICLR 2026

Author: Moritz Reuss Date: October 2025

What is a Vision-Language-Action Model?

A VLA is a model that uses a pretrained backbone, which was trained on large-scale vision-language data and is subsequently trained with generating control commands.

This definition emphasizes internet-scale pretraining as the crucial differentiator, setting VLAs apart from standard multimodal policies that merely combine separate pretrained encoders.

Large Behavior Models (LBMs): trained on large-scale robotic demonstrations without requiring vision-language pretraining. All VLAs trained on massive robotic data are LBMs, but not vice versa.

The Explosive Growth of VLA Research

  • ICLR 2024: 1 rejected submission
  • ICLR 2025: 9 total submissions
  • ICLR 2026: 164 submissions (18x increase year-over-year)

Practitioner’s Guide to Benchmarks

LIBERO: Nearly saturated (95%+ standard). Minimal discrimination between 95–98% results.

CALVIN:

  • ABC (cross-setup) >4.0 standard, >4.5 state-of-the-art
  • D (fine-tuning) >3.75 standard, >4.0 very good
  • ABCD (multi-setup) >4.5 relevant

SIMPLER: Highly variable (40–99% on Bridge alone). Google Robot ~70–80% for current SotA.

1. Discrete Diffusion VLAs

Parallel action sequence generation vs. sequential autoregressive. Generates 100-step sequences in few forward passes rather than 100 iterations. Combined with ECoT for simultaneous sub-goal and reasoning generation.

2. Reasoning VLAs and Embodied Chain-of-Thought (ECoT)

Spatially-grounded predictions (bounding boxes, 2D trajectories) + subtask decomposition. Helps align representations between static VLM pretraining and embodied control.

3. New Discrete Tokenizers

Residual Vector Quantization + frequency/time-domain losses, spline-based parameterization, DCT-inspired objectives for physically plausible motion.

4. Efficient VLAs

Smaller architectures, better tokenizers, quantization, distillation, hypernetworks. Task-specific policy generation conditioned on instructions + initial observations.

5. RL for VLAs

Residual RL with small correction policies on top of frozen VLAs. Stage-aware RL decomposing tasks into semantic phases (reach→grasp→transport→place) with stage-level rewards.

6. VLA + Video Prediction

Initializing from video foundation models + future-frame and action prediction. Limitation: diffusion/flow-based video models face slow inference.

7. Evaluation and Benchmarking

Real-to-sim translation for automatic environment construction. World models as evaluation environments.

8. Cross-Action-Space Learning

Soft-prompting tokens for different datasets, unified vision-motion representations via shared codebooks, hierarchical mixture-of-experts for embodiment adaptation.

9. Other Notable Directions

  • Memory modules aggregating prior context for robustness
  • Policy composition via score-summing (diffusion/flow formulations)
  • VLM backbone selection uncorrelated with standard VLM benchmarks

The Hidden Gap Between Frontier and Research VLAs

Despite open-source VLAs matching frontier performance on simulation, significant gaps emerge in zero-shot open-world behavior. Root causes: benchmark saturation, limited open-source data diversity, insufficient understanding of high-quality demonstration data, narrow evaluation scope.

Two Underrepresented Problems

  1. Data quality: Few submissions address demonstration data curation despite acknowledged OXE data quality concerns
  2. In-context learning: Language alone insufficient for complex physical tasks; limited VLA work explores this despite LLM/VLM success