Exploring Spatial Workspace in Multimodal Models

To understand space is to construct it within, to reason is to manipulate what we construct.

S-Space is an internal spatial workspace in multimodal models. We identify a three-dimensional linear subspace in intermediate activations that encodes continuous object positions along horizontal, vertical, and distance axes. This representation persists across prompts and multi-view contexts, integrates visual and linguistic information beyond visible objects, and can be causally manipulated to change spatial judgments.

Although models possess rich spatial representations, they do not always use them reliably: explicit computation over S-Space can outperform native reasoning. By examining how this workspace is shaped by supervision, transformed through reasoning, and refined by additional context, we provide a mechanistic perspective on spatial intelligence and a path toward more reliable spatial reasoning and action.

Code lives in sspace/; experiment settings live in configs/.

• [2026/09/07] MirroS Blog and Report is live.

The release includes pretrained S-Space axes and experiment commands for Molmo2-ER, MolmoAct2-Pretrain, MolmoAct2, Qwen3.5-4B and Qwen3.6-27B.

Use a CUDA host with uv installed; the installation command creates the three model environments and downloads the main models and InstructPart auxiliary models.

git clone https://github.com/Manchery/s-space.git
cd s-space

./scripts/install.sh

Run the remaining commands from this directory. Launchers select the right environment automatically; no manual activation is needed. Environments are stored under .envs/ and model weights under .cache/assets/models/.

Download and prepare COCO, SpatialTunnel, EmbSpatial, CV-Bench, SpinBench and HStar:

./scripts/prepare_data.sh

Downloads are stored under .cache/assets/ and benchmark-ready data under data/processed/. Preparation reuses existing valid files. The MMSI and InstructPart experiment commands below prepare their own data. Experiment results are written under outputs/.

After setup, choose an experiment below and copy its command. The repository already includes the required configurations; you do not need to write your own.

Evaluations reuse the released axes in sspace/core/artifacts/pretrained/; rebuilding axes is optional.

Learn S-Space axes on COCO-6000 and select readout layers on COCO-1800.

./scripts/run_experiment.sh --experiment sspace-construction

Plot the relationship between S-Space coordinates and physical positions using saved COCO projections.

./scripts/plot_object_coordinates.sh

Compare native answers with S-Space readout on CV-Bench, EmbSpatial and SpatialTunnel.

./scripts/run_experiment.sh --suite benchmarks-generalization \
  --output-root outputs/reproduction/benchmarks

Test whether S-Space readout generalizes across pairwise, single-object and non-spatial prompts.

./scripts/run_experiment.sh --suite prompt-generalization \
  --output-root outputs/reproduction/prompts

Test whether S-Space follows the viewpoint specified by the prompt on HStar.

./scripts/run_experiment.sh --suite hstar \
  --output-root outputs/reproduction/multiview

Test whether editing S-Space changes spatial answers while preserving non-spatial judgments on EmbSpatial.

./scripts/run_experiment.sh --experiment spatial-causality

Generate Qwen3.6-27B reasoning on MMSI and plot where its direction words lie in S-Space.

./scripts/run_experiment.sh --experiment direction-coupling

Compare part–whole spatial judgments across Molmo2-ER, MolmoAct2-Pretrain and MolmoAct2 on InstructPart.

./scripts/run_experiment.sh --experiment action_supervision

Compare direct and chain-of-thought answers with S-Space readout and rotation on SpinBench, with and without the premise.

./scripts/run_experiment.sh --suite spinbench-direct \
  --output-root outputs/reproduction/spinbench-native

./scripts/run_experiment.sh --suite spinbench-projection \
  --output-root outputs/reproduction/spinbench-sspace

Track how Qwen3.6-27B's spatial representations change across a supplied, gold-conditioned reasoning template on SpinBench.

./scripts/run_experiment.sh --experiment cot-evolution

Measure how adding the premise changes layerwise S-Space readout accuracy on SpinBench.

./scripts/run_experiment.sh --experiment context-scaling

This project is licensed under the Apache License 2.0.

If you find S-Space useful, please cite:

@misc{mirros2026sspace,
    title  = {S-Space: Exploring Spatial Workspace in Multimodal Models},
    author = {{MirroS Team}},
    year   = {2026},
    month  = {September},
    url    = {https://mirros.ai/blog/s-space},
    note   = {Blog post}
}