本文由 AI 分析生成
建立時間： 2026-05-30 來源： https://mbreuss.github.io/blog_post_iclr_26_vla.html
Summary
Moritz Reuss surveys VLA research at ICLR 2026, covering 164 submissions (up from 9 at ICLR 2025 — an 18x increase). The post provides benchmark calibration guidance, maps 9 active research directions, and identifies the hidden gap between simulation benchmarks and real-world open-world behavior. Key finding: benchmark saturation masks the gap between open-source VLAs and frontier models.
Moritz Reuss 調查 ICLR 2026 的 VLA 研究現況：從 2025 年的 9 篇成長到 164 篇（18 倍）。文章提供基準測試校準指南、九個研究方向，並指出模擬基準與真實世界零樣本行為之間被隱藏的差距。核心發現：開源 VLA 在模擬中能匹敵前沿模型，但在開放世界表現中仍存在顯著差距。
Key Points
- VLA definition: uses internet-scale vision-language pretraining as backbone (vs. LBMs which use robot-scale data but not VL pretraining)
- LIBERO nearly saturated (95%+); CALVIN ABC >4.0 is current standard; SIMPLER highly variable
- 9 trends: discrete diffusion, reasoning/ECoT, new tokenizers, efficient VLAs, RL for VLAs, video prediction, eval/benchmarking, cross-action-space learning, misc
- Discrete diffusion: generates 100-step action sequences in few forward passes vs. 100 sequential iterations
- ECoT: spatially-grounded predictions (bounding boxes, 2D trajectories) improve VLM-to-embodiment transfer
- Stage-aware RL: decomposes tasks into semantic phases (reach→grasp→transport→place) with per-stage rewards
Insights
The benchmark saturation problem is the central methodological concern — LIBERO at 98% vs. 95% carries no meaningful signal. The “hidden gap” observation is important: open-source VLAs appear to match frontier models in simulation but fail in zero-shot open-world tasks due to limited data diversity and poor evaluation scope. The two underrepresented problems — data quality curation and in-context learning — are both tied to the scaling bottleneck that isn’t about architecture.
Connections
Raw Excerpt
“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.”