SpatialCLI demonstrates that a vision‑language model can keep 73.8 % accuracy on the MindCube navigation suite even after every external spatial‑tool call is removed at test time. The authors report, “On MindCube, SpatialCLI raises Qwen3-VL-8B-Instruct from 29.3% to 84.6% with tools, surpassing GPT-5.6 Sol with tools (72.1%), while retaining 73.8% without tools after internalization.” [1] This overturns the prevailing belief that high‑fidelity perception must be streamed from specialist modules during inference.
Prior multimodal pipelines suffer from multi‑second pauses for each perception request, inflating latency and complicating deployment. In their long‑running reinforcement‑learning runs, “the observed mean latency per spatial‑tool call was 2.916 seconds.” [1] Such overheads have forced engineers to provision dedicated GPU clusters or cache results, limiting real‑time applicability.
When tool access is allowed, SpatialCLI lifts Qwen3‑VL‑8B‑Instruct’s success rate on MindCube from 29.3 % to 84.6 %, outpacing GPT‑5.6 Sol’s 72.1 % ceiling. This leap showcases how exposing a VLM to specialist perception APIs can instantly amplify embodied performance beyond what pure language grounding achieves.
After the internalization stage, the same model climbs to 72.7 % on SpatialCLI‑Bench without any tool calls and reaches a peak of 91.3 % when tools remain available. The paper states, “For example, SpatialCLI-8B improves from 35.3 to 72.7 on SpatialCLI-Bench w/o Tools and further reaches 91.3 w/ Tools.” [1] The distillation step therefore captures most of the tool‑derived advantage inside the model weights.
The approach does not eliminate the cost of training: during long‑running RL training each spatial‑tool call incurred a mean latency of 2.9 seconds, which can make large‑scale fine‑tuning expensive. Moreover, evaluation is confined to the curated SpatialCLI‑Bench suite; how well the internalized knowledge transfers to unseen embodied domains remains an open question.
Consequently, production teams can drop external spatial perception services from their deployment stack and still retain three‑quarters of the tool‑augmented performance, shaving seconds off latency and simplifying infrastructure. The immediate next experiment is to benchmark the zero‑tool model on a broader set of navigation environments such as Habitat or RoboTHOR to verify whether this internalization generalizes beyond the authors’ testbed.
Top comments (0)