Research map / Architectures and efficiency / Attention and transformers
Reflex: Real-Time Vision-Language-Action Control through Streaming Inference
Yuanchun Guo, Bingyan Liu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Partitioned Attention (cache static/sliding/dynamic regions) (ablation) | Reflex on Pi0.5 · LIBERO | inference latency (ms): 135.2 -> 61.5 | Improved |
| add async execution and future-conditional overlap scheduling (ablation) | Reflex on Pi0.5 · LIBERO | reaction latency (ms): 151.9 -> 82.5 | Improved |
| use timestep-aware partitioned caching instead of naive KV caching (ablation) | Flow matching VLA · LIBERO | success rate (%): 12.5 -> 85.4 | Improved |
| use Partitioned Attention instead of full recomputation (ablation) | Flow matching VLA · LIBERO | success rate (%): 85.2 -> 85.4 | No clear effect |
| add AdaRMSNorm to BF16 inference (ablation) | Reflex · BF16 stress test | max stable steps: 120-220 -> >2000 | Improved |
| use AdaRMSNorm instead of FP32 norm-only (ablation) | Reflex · BF16 stress test | added latency (ms): +0.8 -> +0.4 | Improved |
| add fused FlashNorm and FusedAdaLN kernels (ablation) | Reflex action expert · Action expert latency | inference latency (ms): 22.8 -> 18.6 | Improved |
| add future-state predictor at 4 delay steps (ablation) | Reflex · Kinetix controlled delay | success rate (%): +11pp over removing the predictor | Improved |
| increase context window K from 10 to 50 (ablation) | Reflex · LIBERO | speedup: 2.58x -> below 1x | Hurt |
| use K=1 context window (ablation) | Reflex · LIBERO | accuracy | Hurt |
| use Reflex instead of synchronous inference on dynamic Kinetix | Pi0.5 · Kinetix | success rate: +7.4% | Improved |
| use Reflex instead of synchronous inference on LIBERO-Long | Pi0.5 · LIBERO-Long | success rate (%): 68.8 -> 72.4 | Improved |
Closest papers by meaning
- FASTER: Rethinking Real-Time Flow VLAs (NeurIPS 2026 · similarity 0.90)
- ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models (NeurIPS 2026 · similarity 0.89)
- VLA-Cache: Efficient Vision-Language-Action Manipulation via Adaptive Token Caching (NeurIPS 2025 · similarity 0.89)
- EfficientVLA: Training-Free Acceleration and Compression for Vision-Language-Action Models (NeurIPS 2025 · similarity 0.88)
- Generative Control as Optimization: Time Unconditional Flow Matching for Adaptive and Robust Robotic Control (NeurIPS 2026 · similarity 0.88)
- VLA-Corrector: Lightweight Detect-and-Correct Inference for Adaptive Action Horizon (NeurIPS 2026 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
