Zero-Token 5.6x Faster Spatial Intelligence: MJev-Spatial Tops HF Daily Papers Alongside CUHK JIA-Lab's CarryOn Dual-Loop Robot Harness

Core Background & Industry Pain Points

Deploying multimodal agents that perceive and act reliably in the physical world faces two major bottlenecks: excessive perception latency and brittle physical execution. Standard Vision-Language Models (VLMs) answer spatial queries—whether selecting a spatial relation, estimating metric distance, or predicting a 2D placement coordinate—via free-form autoregressive text generation and regex parsing, paying multi-token decoding latency for a compact output space. Meanwhile, generalist Vision-Language-Action (VLA) models lack persistent user personalization across sessions and remain prone to hazardous collisions and unrecoverable physical stalls during open-world manipulation.

Architectural Highlights & Core Mechanisms

Two complementary open-source frameworks—MJev-Spatial (arXiv:2610.22276, #1 on Hugging Face Daily Papers) and CUHK JIA-Lab's CarryOn (arXiv:2610.21453)—address System-1 spatial perception and dual-loop execution reliability:

• MJev-Spatial: Zero-Parameter Native Attention Matching: Reformulates multiple-choice relations, numerical distance estimation (via coarse-to-fine value bins), and 2D point grounding (via coarse-to-fine spatial image grids) into a unified classification paradigm. Candidates are scored directly using the VLM backbone's pretrained Query and Key projections with zero added parameters. At inference, each decision completes in a single forward pass with zero generated answer tokens. During training, joint teacher-forced supervision across coarse-to-fine stages combined with length-aware global batching and cross-GPU load balancing boosts training throughput by 30.8%.
• CarryOn: Dual-Loop Harness for Frozen VLA Policies: Wraps a frozen VLA backbone in two coordinated loops. The Planning Loop maintains persistent user-level preferences and scene-level state across multi-turn interactions for personalized subtask grounding; the Execution Loop shields policy rollouts from predicted hazardous contacts and triggers reusable control-based recovery skills refined via post-failure reflection.

Authoritative Benchmarks & Empirical Comparisons

Both systems demonstrate substantial empirical gains across spatial and embodied benchmarks:

• MJev-Spatial Across 12 Spatial Benchmarks: Built on Qwen3.5-4B and trained on 12.3M spatial QA pairs, MJev-Spatial achieves the highest overall score (65.64%) among 8 evaluated models, ranks #1 on 7 of 12 benchmarks, and outperforms its unmodified autoregressive base model on all 12—while running 2.5x faster than the fastest autoregressive baseline and 5.6x faster than the base model.
• CarryOn Real-Robot & Simulation Evaluations: Across multi-user and multi-scene manipulation tasks, CarryOn slashes the safety violation rate of an already safety-finetuned VLA from 52.6% down to 6.1%, and boosts physical failure recovery success from 13% to 80% when reusing reflection-refined skills.

Developer Practical Guide & Quickstart

Both MJev-Spatial (Fr0zenCrane/MJev-Spatial) and CarryOn (JIA-Lab-research/CarryOn) are open-sourced on GitHub. Embodied AI developers can deploy MJev-Spatial as a real-time, zero-decoding spatial grounding front-end while wrapping existing VLA policies in the CarryOn harness to enforce collision shielding and automated skill recovery without retraining the underlying manipulation weights.