Space-X
PAPER TITLE

Space-X: Self-Evolving Spatial Agents Weaving Clues into Evidence Chains

Weaving clues
into answers.

Seeing is a beginning.
Reasoning is making the connections.

Space-X turns images, tool outputs, and memory into a continuous evidence chain. Its agents evolve the tools and the way they use them.

Grounded Connected Traceable
EVIDENCE OBSERVATORYRECORDED CASE

How far apart are the closest surfaces?

Source view of the white step stool
Stool SOURCE A
Source view of the adult toilet
Toilet SOURCE B
PC1PC2 SAVED GEOMETRY / PCA PROJECTION
0.86msource-supported estimate

One evolving harness. Tools, memory, and evidence working together.

Scroll to connect the dots

THE MISSING CONNECTION

More tools. Better evidence?

A tool can return the right output and still leave a question unanswered.
The difference is what the agent does with it.

Sofa viewA second room viewA third room viewWhich view supports the question?

Looking without a target

Repeated observations can miss the object, observer position, or facing landmark that defines the question.

Acquire the missing premise
Stool segmentation candidate
Camera depthSurface distance
A number needs the right meaning.

Outputs without connections

A mask, depth map, and distance only become evidence when their objects, sources, and reference frames agree.

Bind and associate the evidence
Proposal A 0.86 mProposal B 0.10 mRevisit the sources. Resolve the premise.

Answers without enough support

Conflicting proposals need a source-based check of the disputed premise, with enough capacity left to conclude.

Judge with traceable support

Evolve the whole connection. Evidence needs guide tool development.
Usable tool results guide the next observation, association, and decision.

TWO AGENTS. ONE SHARED OBJECTIVE.

Learn to reason.
Then improve the way.

Each execution builds an evidence chain.
Each iteration improves the tools and rules that make the next chain possible.

Execution traces · broken links · evaluation feedback
AeTASK-TIME REASONING

Execution Agent

Use capabilities to solve the current question.

Execution robot working with tools inside its evidence workshop
Acquire Associate Judge
Reliable
evidence chains
THE EVOLUTION TARGET
AcCROSS-ITERATION IMPROVEMENT

Control Agent

Improve the system around the reasoning model.

Control robot inspecting and improving the tool workspace
Tools & interfaces Memory & skills Evidence-use rules
DiagnoseEditEvaluate
Retained capabilities · clearer interfaces · reusable reasoning practices
ONE DEVELOPMENT EXAMPLE

From forgotten observations
to source-linked memory.

OBSERVED FAILUREUseful views disappear across context windows.

STRUCTURAL CHANGERelay observations, restore cited images, reserve review capacity.

RECORDED OUTCOMEThe sofa case reaches a supported front-left answer.

THE RESEARCH FILM

See the connections.
Watch the reasoning unfold.

The core idea, spatial cases, and measured gains.
One connected story of self-evolution.

03:32 ENGLISH 1080P
NOW SHOWINGThe central idea00:00 / 03:32

A narrated overview of evidence chains, dual-agent self-evolution, spatial cases, benchmark gains, and X-SFT. English subtitles are included in the video.

A harness is more than a toolbox. Its components evolve together.

THE EVIDENCE LAB

Don't just see the answer.
See what supports it.

Explore four recorded development cases, from the first clue to a source-grounded decision.

METRIC SPATIAL REASONING Recorded evidence walkthrough

What is the closest surface-to-surface distance between the stool and the toilet?

01 / 03
EARLIER EXECUTION
IMPROVED EXECUTION

MEMORY THAT KEEPS THE LINKS

Remember the clue.
Keep its context.

Five complementary memory roles preserve what was seen, what it means, what remains uncertain, and what to do next.

A source reference travels with the claim, so later reasoning can return to the evidence.

Inside the complete system
01

ILLUSTRATIVE RECORD
RetrieveUpdate

A LOOK INSIDE SPACE-X

One system.
Connected capabilities.

Tools construct evidence. Memory maintains the connections.
The API reasons over what is actually supported.

View retrieval Segmentation Depth & geometry Spatial readouts Source restoration Reinspection
From input to tools to evidence construction, with five memory roles and source-linked review connecting successive API calls.

SAME MODEL. NEW POSSIBILITIES.

Same models.
Stronger spatial reasoning.

ViewSpatialSPARMMSIVSI 10 initial frames

Model aloneSpace-XBenchmark scores · labeled zoomed axes

THE BIGGEST LEAPS

See the difference.

Selected gains > 10%

SELF-EVOLUTION, MADE VISIBLE

Capabilities grow.
Connections grow with them.

Follow the tools, skills, memory and evidence links as Space-X takes shape.

EVOLUTION REPLAYOne shared objective. Five kinds of progress.
GPT-6-sol · construction run
Implemented candidates575of 575 in the recorded sequence
ViewSpatial79.99starting score 53.19
SPAR70.39starting score 57.15
Beyond adding tools66.26%of candidates address rules, memory,
evidence links or execution control
Capability & evidence organization cumulative candidates · 0–200
ViewSpatial SPARDashed lines · benchmark scores · 50–80
Integrate & refine

Combine capabilities, restore source evidence and close the reasoning chain.

575 optimization steps. Ten measured checkpoints.

WHAT THE EXPERIMENTS TEACH US

Not just more tools.
A better way to use them.

Better tools. Connected memory. A harness that adapts.

CAPABILITY × EVIDENCE USE

Build the tools.
Teach the connections.

Jointly evolving tools and evidence rules beats changing either side alone.

+11.84 ptsjoint evolution over tools alone
EVIDENCE THAT STAYS USABLE

Keep the clue.
Preserve the connection.

Explicit links align evidence. Role-specific memory carries it into the next decision.

SourceMemoryDecision
Evidence delivered to later calls98.24%71.06% with flat records
21.70 → 1.05%semantic mismatch, flat records → full continuity
X-SFT · ADAPT WHAT YOU ALREADY BUILT

Transfer the system.
Make it your own.

Let a new model improve the inherited harness, using the original development data.

Inherited
Space-X
15iterationsTarget-adapted
Space-X
One starting point.
Three successful adaptations.
Reuse structure. Refine model fit.

The improvement is in the coordination.
Tools, memory and decisions evolve around one continuous evidence chain.

See the mechanisms in action

THE EXTENDED CASE ATLAS

More than an answer.
A closer look at what changed.

Pick a scene. Follow the evidence. See what changes.

14distinct questions25thematic analyses

Explore a different angle.

WHAT THIS CASE SHOWS

REFERENCE

FROM SEEING TO SOLVING

Every answer has a chain.
Make every link count.

Space-X: Self-Evolving Spatial Agents Weaving Clues into Evidence Chains

Explore the evidence

Execution architecture

Source inspector