Space-X lets spatial agents evolve their own harness: a control agent rewrites the execution structure, and all four base models gain
Synopsis
Space-X makes building a complete, continuous evidence chain the target of spatial reasoning: an Execution Agent acquires, associates and judges evidence while solving, and a Control Agent reads execution traces offline, diagnoses breaks in that chain and jointly revises tools, interfaces, skills, memory organization and decision rules. Across ViewSpatial, SPAR, MMSI, VSI and BLINK, all four base models improve once equipped; for GPT-6-sol, answer accuracy rises by 50.39% on ViewSpatial and 23.07% on MMSI.
Interpretation
The improvement target for spatial agents shifts from stacking tools and calls to constructing a continuous evidence chain that starts from the question, stays grounded in observations and ends at the answer. Tool-augmented agents previously re-detected scenes while missing a required target, observer position or facing landmark, joined results from different instances, views, scales or reference frames, and answered before the necessary relations were established. Across ViewSpatial, SPAR, MMSI, VSI and BLINK, all four base models improve once equipped; for GPT-6-sol, answer accuracy rises by 50.39% on ViewSpatial and 23.07% on MMSI, and Gemini-3.8-Flash gains 47.50% and 25.22% on the same two benchmarks.
The framework splits solving from improvement: the Execution Agent handles evidence acquisition, association and judgment for the current task, while the Control Agent reads execution traces offline, locates the broken evidence link and rewrites the execution structure. The Control Agent's edits go beyond tool implementations to interface semantics, memory destinations, context delivery, review triggers and closure rules, and a candidate change is kept only after passing comparison at both the evidence-chain and complete-task levels. An ablation under the same development data, validation split and optimization budget compares four conditions: tool-only evolution reaches 68.15 on ViewSpatial, rule-only 61.52, and joint evolution 79.99, with SPAR and MMSI following the same order.
The Execution Agent first decomposes the question into target, reference, observer position, facing landmark, reference frame and output units, then selects tools for unresolved premises and stores each output with its source, semantic binding and status across five memory roles. Evidence is no longer just a return value: it records which object it refers to, which original image it came from, which frame it sits in and which claim it supports, with working, observation, spatial, episodic and procedural memory holding premises and gaps, original views, geometry, execution history and reusable guidance. With tools, skills and answer judgment fixed, flat records score 67.25, 63.03 and 59.91 on ViewSpatial, SPAR and MMSI; explicit association raises these to 73.26, 66.96 and 63.73; adding role-specific memory continuity reaches 79.99, 70.39 and 66.33, while semantic mismatch falls from 21.70 to 1.05 and evidence delivery rises from 71.06 to 98.24.
Perspective
The demonstrated scope is multimodal model evaluation on five spatial reasoning benchmarks, ViewSpatial, SPAR, MMSI, VSI and BLINK, with the harness built using GPT-6-sol and then used unchanged across four base models, and VSI fed only 10 initial frames rather than dense video. The intended readers are researchers and engineers building tool-augmented multimodal agents.
VSI uses only 10 initial frames, so the reported numbers describe behavior under sparse visual input; the repair and damage rates appear with different values in different tables and must be read together with their setting; after cross-model transfer, a further fifteen X-SFT iterations are needed for additional gains.
