Introduction
In 2026, vision-language-action (VLA) models have indisputably become the mainstream architecture for embodied AI, but the focus of competition has subtly shifted: leading players no longer compete on "can it be done," but on "can it be done in real time, can it be done cheaply, and can it be transferred to a different embodied platform without retraining." The VLA paper list at ICLR 2026 features a large number of works on token pruning, quantization, cross-embodiment soft prompting, memory retrieval, and unified diffusion—a typical signal that a technology has entered the engineering phase. Once the architectural dividend has been flattened, efficiency and generalization become the new dividing lines, and this is precisely the stage where demands on data quality are highest.
Technical Highlights
At the architecture level: unification has been established; divergence lies in the details. The mainstream approach remains "pretrained VLM + action expert," but several important variants emerged in 2026:
- •Unified diffusion and autoregression: Works such as Unified Diffusion VLA at ICLR 2026 attempt to unify discrete diffusion and autoregression into a single objective, balancing multimodal action-distribution modeling with inference speed; FASTer proposes a learnable action tokenizer and chunked decoding to improve the efficiency of the autoregressive route.
- •Cross-embodiment generalization: X-VLA uses a soft-prompt Transformer as a scalable cross-embodiment interface, attempting to adapt a single set of weights to multiple robots; MetaVLA uses unified meta-collaborative training to improve adaptation efficiency.
- •Memory and retrievalMemoryVLA introduces "perception-cognition" memory into VLA, while Scaling up Memory for Robotic Control via Experience Retrieval expands memory capacity through experience retrieval, directly addressing the weakness in long-horizon tasks.
- •Efficiency engineering: SP-VLA jointly performs model scheduling and token pruning; AutoQVLA points out that "not all channels are equivalent" in VLA quantization; Action-aware Dynamic Pruning dynamically prunes based on action relevance—together, these three works demonstrate thaton-device deployment is already a clear engineering objective.
Reflections at the evaluation level are beginning to emerge. The VLM4VLA system re-examines the impact of different VLM backbones on VLA performance; "Actions as Language" explores pathways for fine-tuning VLMs into VLAs without catastrophic forgetting. This kind of "looking back" work typically appears after a technical paradigm has stabilized.
Accelerated engineering deployment. Xiaomi open-sourced Xiaomi-robotics-0 (arXiv 2026), focused on real-time execution; NVIDIA and the community released VLA Foundry as a unified training framework; π0.7 (Physical Intelligence, 2026-04) emphasizes controllability and emergent capabilities, and integrates a world model as a subgoal image generator.
Key Breakthroughs
- •Efficiency is no longer a secondary metric: The three concurrent works SP-VLA, AutoQVLA, and Action-aware Dynamic Pruning all point to "inference budget," indicating that the industry has broadly recognized that the bottleneck of VLA has shifted from "effectiveness" to "latency × effectiveness."
- •Cross-Embodiment Moves from Vision to a Trainable Goal: X-VLA's soft-prompt approach and MetaVLA's meta-collaborative training turn "adapting one set of weights to multiple embodiments" from a slogan into an optimizable objective function. This has profound implications for the data side—the stronger the cross-embodiment capability, the higher the aggregation value of heterogeneous data.
- •Memory is explicitly modeled: The RoboDojo leaderboard shows that "memory" is a key dimension distinguishing top models (after the 2026-08-25 update, the top-ranked model scored 47.74 on the memory dimension, while the second-place model scored only 13.37), corroborating the large number of memory-related papers at ICLR 2026.
Impact on the Industry
First,The importance of data quality is amplified, not diminished, by architectural advances. When the model side pushes "effective information density" to the extreme through pruning, quantization, and retrieval, the signal-to-noise ratio of the input data directly determines output quality. Temporal synchronization errors across cameras, calibration drift, and blurry frames—previously merely "minor degradation"—now directly translate into policy failures.
Second,Improvements in cross-embodiment generalization will increase the premium on "multi-configuration data". Datasets that cover multiple embodiments, multiple tasks, and include failure and correction samples have higher marginal value than an equal volume of data from a single configuration.
Third,The on-device deployment trend raises requirements for "data–model co-design". If the customer's target model is a small model after pruning and quantization, then the distribution of training data should match that model's capacity—data suppliers need to understand the customer's deployment form, not merely deliver raw video.
Image Suggestions
Figure 1 (placement: after the introduction)
- •Figure description: A horizontal timeline of VLA architecture evolution, showing four generations from left to right: (1) VLM + simple action head (2023–2024) → (2) VLM + action expert / flow matching (2025) → (3) + memory retrieval and world-model subgoals (first half of 2026) → (4) + token pruning / quantization / cross-embodiment soft prompts (second half of 2026). Below each generation, the representative work names and conferences are annotated.
- •Figure caption: Figure 1 | The four-stage evolution of VLA architectures: from "able to output actions" to "real-time, cross-platform, deployable."
Figure 2 (placement: after the efficiency engineering paragraph in "Technical Highlights")
- •Screen description: A three-column comparison of efficiency optimization methods. Left column "Token Layer": lit patches across the feature map decrease layer by layer, labeled Pruning / Merging; middle column "Parameter Layer": weight matrix divided into blocks, some blocks marked as low-bit (INT4/INT8), labeled Mixed-precision Quantization; right column "Scheduling Layer": schematic of an early-exit inference path, labeled Model Scheduling. A bottom bar compares "latency before optimization" with "latency after optimization".
- •Figure caption: Figure 2 | Three categories of efficiency optimization for VLA: token pruning, mixed-precision quantization, and model scheduling (latency comparisons are illustrative values).
Figure 3 (placement: after the first item under "Impact on the industry")
- •Image description: A propagation chain from "data quality" to "policy success rate": synchronization error / calibration drift / motion blur → noisy geometric supervision signals → degraded representation quality → amplified information loss after pruning and quantization → increased policy failure rate. Each stage is connected by arrows, and the noise-injection stage is highlighted in red.
- •Figure caption: Figure 3 | How data quality is amplified through model efficiency optimization: on-device compression amplifies the impact of input noise.
Source and Date
- •List of ICLR 2026 VLA-related papers (MemoryVLA, PixelVLA, X-VLA, SP-VLA, AutoQVLA, Unified Diffusion VLA, FASTer, WholeBodyVLA, etc.): open-source paper collection repository Embodied-AI-Paper-TopConf, updated 2026-03.
- •π0.7 "a Steerable Model with Emergent Capabilities": Physical Intelligence, 2026-04.
- •Xiaomi-robotics-0, VLA Foundry, Being-H0.5/H0.7, Fast-WAM: arXiv 2026, cited as compiled in the NVIDIA technical blog "Pretrained to Imagine, Fine-Tuned to Act: The Rise of World-Action Models" (2026).
- •RoboDojo leaderboard memory dimension 47.74 vs 13.37: RoboDojo leaderboard update, 2026-08-25.
Annotation Notes: ICLR 2026 papers are subject to the official conference acceptance list; the final versions and metrics of individual papers may differ from the preprints (To be verified). The architecture generation scheme is the author's synthesis based on the public timeline and is anSpeculative compilation, not an official classification.