#1 ·
VFlash: Fast, Accurate Visual Planning via Flash Primitive Decoding
Yu-Cheng Chou, Shihao Wang, Jin Huang, Chi-Pin Huang
Embodied AIInference EfficiencyVision-Language Models
Why it matters
Any system where a vision-language model must emit spatial output — manipulation, driving, UI grounding — pays for it twice: autoregressive decoding spits out coordinates one small numeric token at a time, which is slow and where sampling error compounds. Predicting an entire primitive (points, boxes, trajectories) in a single forward pass removes that bottleneck, and the result is a rare speed-AND-accuracy win rather than a trade-off. The measured gains are practical: roughly 25x faster generation for 3D point motion forecasting, and a world action model conditioned on VFlash plans that lifts success from 62.6% to 70.2% across four MolmoSpaces task suites.
VFlash introduces a lightweight drafter that predicts an entire visual primitive — all coordinates jointly — in one forward pass, with a layout that adapts to the primitive's type, dimensionality, and length (compact boxes to long variable-length trajectories). Instead of copying an autoregressive model's exact token sequence, it trains on annotated primitives and directly accepts any structurally valid prediction. Across 2D and 3D robot trajectory tasks it consistently reduces prediction error relative to autoregressive decoding, and preliminary single-rollout measurements show about 25x faster generation for 3D point motion forecasting.
arXiv | HuggingFace
#2 ·
SLA: Scaling Spatial-Language-Action Model via Continual Pre-training
Fuhao Li, Wenxuan Song, Qizhou Huang, Ziyang Zhou
RoboticsVision-Language-ActionPre-training
Why it matters
VLA models inherit their visual representations from 2D vision-language pre-training, so depth, occlusion, and 3D geometry have to be relearned at every task-specific adaptation. SLA's bet is that spatial understanding belongs in the foundation layer, and the design is production-realistic: all the spatial machinery — alignment with a frozen 3D perception model, feature-dependent gates across VLA layers, use of partial depth and camera calibration — is training-only, so the shipped model reuses the original VLA weights and inference architecture unchanged. Teams already running pi0-family policies get a concrete recipe for making manipulation generalize across embodiments and scenes without per-task geometry fine-tuning.
SLA scales spatial understanding into continual pre-training: initialized from a pretrained pi05 policy, it is trained on curated, unified robot datasets spanning diverse embodiments, tasks, and scenes, jointly optimizing action prediction and alignment between intermediate visual embeddings and spatial representations from a frozen 3D perception model. Feature-dependent gates aggregate visual embeddings from multiple VLA layers for the spatial supervision, the teacher model exploits partially available depth and camera calibration alongside RGB, and probabilistic state dropout reduces overfitting to demonstration trajectories. Critically, every auxiliary component is removed after pre-training, preserving the original inference pipeline.
arXiv | HuggingFace
#3 ·
Harness-S: Learning from Sibling Harnesses for Generalizable Agent Training
Kangning Zhang, Wenxiang Jiao, Bowen Fang, Shuai Shao
AI AgentsReinforcement LearningAgent Training
Why it matters
The harness — the scaffolding around the model: prompts, tools, execution framework — is where agent teams spend most of their iteration time, but scaffolding improvements typically die with the scaffold itself. Harness-S shows a policy trained through scaffolds it will never use at inference time keeps the behavior: the model is rescued from its own failures by sibling harnesses, and the rescued trajectories are folded back into the same policy, which then performs substantially better under the original base harness. For practitioners, that decouples training infrastructure from deployment — you can train with expensive, elaborate scaffolding and still ship the plain system. The router even helps single-harness inference at the same rollout budget.
Harness-S is built on the observation that harness variants evolved from a common base are local, complementary, and nonstationary: each repairs different failures, and their value shifts as the policy learns. Directional Harness Evolution grows a fixed pool of specialized sibling harnesses; when the base harness fails a task, rescue training invokes the pool — the same policy generates fresh rollouts under a selected auxiliary harness, and those trajectories replace the failed group for reinforcement learning. An adaptive router picks the top-k siblings using harness behavior profiles and voting across multiple LLMs, with profile re-anchoring as the policy improves. On AppWorld, SWE-Gym, and DBBench, the resulting policy substantially outperforms policy-only training when returned to the original base harness.
arXiv | HuggingFace
Outlook
The theme is the gap between model capability and deployable physical action: where signal is lost (token-level coordinate decoding, 2D visual priors) and how to recover it without re-architecting (draft-based primitive decoding, training-only spatial alignment, scaffolds that train but never ship). Expect follow-ups as VLA labs adopt spatial pre-training as a standard stage, and as fast primitive decoding gets ported to driving and UI grounding.