miStudio + miLLM: Direction of the Mechanistic AI Suite

Where the Suite Is Going

miStudio and miLLM are evolving together as an integrated environment for understanding and working with the internal representations and mechanisms of learned systems.

The suite began with a strong practical emphasis on sparse autoencoders and mechanistic experimentation around language models. That remains important, but the direction is deliberately broader.

The objective is to support a research progression from:

observing internal state
        ↓
discovering representations
        ↓
understanding their structure
        ↓
manipulating them experimentally
        ↓
discovering how they interact
        ↓
identifying candidate mechanisms
        ↓
testing those mechanisms causally
        ↓
validating and reproducing the findings
        ↓
applying validated knowledge during inference

The underlying question is increasingly not simply:

What feature activates here?

It is:

What internal variables has this learned system developed, how are those variables represented, how do they interact, and which of those interactions actually contribute to the resulting behavior?

That shift drives much of the roadmap.

miStudio is the environment in which those questions are investigated experimentally.

miLLM is the instrumented inference environment in which sufficiently understood representations and interventions can be observed and ultimately applied during model execution.

The Suite Today

The roadmap starts from a platform that already has substantial mechanistic experimentation capability.

miStudio

miStudio is a rich interactive mechanistic interpretability workbench.

Its sparse-autoencoder capabilities already go well beyond feature extraction and visualization. In particular, miStudio has an extensive SAE steering environment designed for comparative experimentation.

A researcher can hold a prompt constant and generate a series of responses while progressively changing the steering strength applied to a feature:

Same Prompt

Feature X @ -2.0  → response
Feature X @ -1.0  → response
Feature X @  0.0  → baseline
Feature X @ +1.0  → response
Feature X @ +2.0  → response
Feature X @ +3.0  → response

This makes it possible to see whether an internal representation exhibits a meaningful intervention-response relationship rather than relying on a single steered generation.

miStudio can also perform equivalent experiments sequentially across multiple features, allowing their effects to be compared against the same prompt.

It can additionally steer several SAE features simultaneously.

That makes experiments possible in which representations are treated not only individually but as potentially interacting components:

Feature A   +1.4
Feature B   -0.7
Feature C   +2.0
        ↓
     inference
        ↓
      response

This existing capability is important to the direction of the product.

The roadmap is not primarily about adding steering.

It is about extending the experimental discipline already present in SAE steering to more general forms of model representation and intervention.

Jacobian Lens and J-space

miStudio is also expanding beyond an SAE-centered view of internal representations.

Jacobian Lens fitting has already been added and is being actively developed further, along with J-space analysis and characterization.

J-space swapping is currently under development.

That work is significant because it creates a second major representation framework within the product and forces an important architectural question:

Can the same experimental environment work regardless of whether the representation was discovered through an SAE, J-space, a probe, a direction, a subspace, or some future method?

The roadmap assumes the answer should be yes.

miLLM

miLLM provides the complementary runtime environment.

It is not intended to be merely a backend inference service. It has its own rich GUI and is designed to make mechanistically instrumented inference directly usable and inspectable.

As the suite develops, miLLM becomes the environment in which validated mechanistic knowledge can be applied efficiently while inference is taking place.

The division is therefore straightforward:

miStudio is optimized for discovery, experimentation, visualization, comparison, and validation.

miLLM is optimized for instrumented inference, runtime observation, and execution of validated interventions.

Both remain full interactive products.

Human Interfaces and Programmatic Interfaces

A central design principle for the suite is that automation should not come at the expense of rich interactive research.

Mechanistic interpretability is inherently exploratory.

A researcher often learns something precisely because an unexpected activation pattern, response transition, geometric relationship, or interaction becomes visually obvious.

Both miStudio and miLLM will therefore continue to receive rich purpose-built GUIs as functionality grows.

At the same time, the same capabilities should be accessible through programmatic interfaces where appropriate.

The suite currently includes a shared MCP server, and MCP controls are expanded as capabilities are added to miStudio and miLLM.

The intended interaction model is therefore:

                    Human Researchers
                    /               \
                   ▼                 ▼
            miStudio GUI         miLLM GUI
                 │                   │
                 ▼                   ▼
             APIs / Services / Core Capabilities
                         ▲
                         │
                  Shared MCP Server
                         ▲
                         │
               Agents / Research Harnesses

These interfaces serve different purposes.

The GUI is optimized for exploration, visualization, comparison, and human judgment.

The APIs provide deterministic application integration and automation.

MCP provides a semantic interface through which agents and research harnesses can participate in the same workflows.

MCP does not replace the applications, and agents do not replace visual research.

The intended model is complementary.

A research agent might execute a broad experiment campaign and identify an unusual feature interaction. A researcher could then open those experiment results in miStudio and investigate the phenomenon visually.

Likewise, something discovered interactively in the GUI could become the basis for a much larger automated campaign.

A Common Scientific Model

The suite is increasingly organized around a simple research model:

Stimulus / Observation
        ↓
Internal Representation
        ↓
Representation Interaction
        ↓
Mechanism
        ↓
Prediction / Action / Output

For each stage, miStudio should progressively help answer a stronger sequence of questions.

Can we observe it?

Is there an identifiable internal state associated with the phenomenon?

Can we decode or interpret it?

What information appears to be represented?

How is it represented?

Is it:

• a sparse feature,
• several related features,
• a J-space representation,
• a direction,
• a subspace,
• a distributed population code,
• a spatial state,
• a temporal state?

Where and when does it appear?

Does the representation emerge at a particular layer, training checkpoint, modality transition, or temporal stage?

What does it interact with?

Which other internal variables appear upstream, downstream, or jointly involved?

Can we manipulate it?

Does changing the representation alter other internal states or model behavior?

Does the effect behave predictably?

Do increasing intervention strengths produce corresponding changes?

Do combinations of representations interact as expected?

Does the mechanism survive context changes?

Is the relationship stable across prompts, datasets, models, and modalities?

Can the result be reproduced?

Does the result survive replication and controlled variation?

The roadmap increasingly treats these questions as the organizing structure of mechanistic research.

From Features to Representation Systems

An SAE feature is one possible representation object, not the universal unit of interpretability.

miStudio should therefore develop a common Representation framework capable of describing multiple views of internal model state:

Model Internal State
        │
        ├── Native Activation Space
        ├── SAE Feature Space
        ├── Jacobian Lens / J-space
        ├── Probe Directions
        ├── Steering Directions
        ├── Learned Subspaces
        ├── Population Codes
        └── Future Representation Methods

This matters scientifically because different phenomena may naturally appear in different forms.

One concept might correspond closely to an individual SAE feature.

Another might require several interacting SAE features.

Another might occupy a multidimensional subspace.

Another might be most clearly characterized in J-space.

A temporal model may represent an internal variable through a trajectory rather than a static vector.

The system should follow the representation rather than requiring the representation to conform to the tool.

This also makes comparison across methods possible.

If an SAE, J-space analysis, and linear probe independently expose related internal information, miStudio should be able to relate those observations and use agreement or disagreement among them as additional evidence.

From Steering to General Intervention Experiments

The existing SAE steering environment provides a useful blueprint for broader causal experimentation.

Today, a typical experiment can be described as:

Prompt
  ×
SAE Feature
  ×
Steering Strength

The generalized form becomes:

Stimulus
  ×
Representation System
  ×
Representation Object(s)
  ×
Intervention
  ×
Intervention Magnitude
  ×
Context

Over time, that framework should support interventions including:

• SAE steering,
• J-space swapping,
• future J-space steering,
• activation patching,
• ablation,
• clamping,
• representation replacement,
• subspace projection,
• subspace removal,
• circuit intervention.

The fundamental experiment remains familiar:

Controlled Stimulus
        ↓
Baseline
        ↓
Intervention
        ↓
Counterfactual Inference
        ↓
Observe Internal Changes
        ↓
Observe Behavioral Changes
        ↓
Compare With Prediction
        ↓
Replicate

The current SAE steering GUI already provides strong qualitative comparison.

The roadmap adds increasingly quantitative measurements alongside it, allowing intervention effects to be evaluated through both human inspection and structured metrics.

Evidence, Not Just Interpretations

A useful distinction for the suite is between an interpretation and an evidenced mechanistic claim.

A representation that has been automatically labeled is not epistemically equivalent to a representation whose role has survived controlled interventions across multiple contexts.

The roadmap therefore introduces an evidence progression such as:

Observed
   ↓
Correlated
   ↓
Decoded
   ↓
Localized
   ↓
Attributed
   ↓
Intervened
   ↓
Dose / Response Established
   ↓
Causally Supported
   ↓
Context Validated
   ↓
Replicated

The existing steering-strength sweeps make the dose/response stage particularly natural.

If progressively changing an internal representation systematically changes model behavior, that provides stronger evidence than a single isolated steering result.

Likewise, simultaneous multi-feature steering provides the basis for testing interaction effects.

miStudio should increasingly preserve this evidence as part of the representation or mechanism itself.

The long-term result is not merely a collection of feature labels.

It is a body of mechanistic claims whose provenance and evidentiary support can be inspected.

Dataset Engineering as Experimental Design

Data increasingly becomes part of the experiment rather than merely an input to it.

miStudio already supports two-way Hugging Face dataset exchange.

The roadmap extends that into a broader Dataset Studio capable of importing, constructing, transforming, composing, and versioning controlled experimental datasets.

Sources should progressively include:

• Hugging Face,
• local and uploaded files,
• object storage,
• HTTP sources,
• Git-hosted data,
• synthetic/generated datasets,
• derived miStudio datasets.

Researchers should be able to compose experimental corpora deliberately.

For example:

General text        50%
Source code         20%
Scientific text     15%
Synthetic examples  10%
Rare examples        5%

The same research campaign could then be repeated against another mixture.

That allows questions such as:

Does changing training or stimulus distribution alter where a representation emerges?

Does the same concept have the same internal geometry across domains?

Is an apparent mechanism robust to a different data distribution?

Dataset lineage, sampling, transformations, seeds, and revisions should therefore become part of experiment provenance.

Workflow and Research Campaigns

As the number of experimental stages grows, individual actions need to become reproducible research procedures.

Workflow Studio should represent processes such as:

Dataset
  ↓
Activation Extraction
  ↓
SAE Training
  ↓
Feature Discovery
  ↓
Interpretation
  ↓
Steering Experiment
  ↓
Validation

or:

Dataset
  ↓
Activation Capture
  ↓
Jacobian Collection
  ↓
J-Lens Fit
  ↓
J-space Analysis
  ↓
Swap Experiment
  ↓
Validation

The intent is not to turn miStudio into a generic workflow product.

It is to make mechanistic research procedures repeatable, resumable, inspectable, and shareable.

A Research Campaign sits one level above a workflow.

A workflow executes an experiment.

A campaign executes a family of experiments.

For example:

3 datasets
×
4 layers
×
2 representation methods
×
5 intervention strengths
×
3 random seeds

miStudio should be able to expand that experiment space, schedule the work, preserve provenance, aggregate results, and provide a rich visual environment for understanding the outcome.

This is a natural extension of the experiment-sweep methodology already present in SAE steering.

Representation Geometry and Emergence

Once representation types become method-independent, geometry becomes increasingly important.

The question may no longer be:

Which feature represents X?

but instead:

What is the dimensional structure of X inside the model?

miStudio should therefore support analyses such as:

• PCA,
• SVD,
• CCA,
• intrinsic dimensionality,
• subspace relationships,
• representational similarity,
• manifold analysis,
• population-code analysis,
• cross-layer comparison,
• cross-model comparison.

A concept may turn out to require a 20-dimensional subspace rather than one interpretable direction.

That is not a failure of interpretability.

It is information about the actual representation used by the model.

The suite should also support representation emergence analysis.

Rather than selecting an interesting layer manually, researchers should be able to analyze representations across layers, checkpoints, architectures, model scales, and dataset conditions.

Controlled Stimulus
        ↓
Capture Multiple Layers
        ↓
Probe / SAE / J-space Analysis
        ↓
Measure Representation Accessibility
        ↓
Identify Transition Region
        ↓
Characterize Representation
        ↓
Perform Intervention

This moves the product from studying what representations exist in finished models toward studying how representations form.

From Representation Discovery to Mechanism Discovery

Finding an internal representation does not establish how it participates in computation.

Mechanism discovery therefore becomes a broader concept than traditional circuit visualization.

A candidate mechanism may combine different types of internal objects:

SAE Feature Group
        ↓
J-space Representation
        ↓
Subspace
        ↓
Downstream Representation
        ↓
Output

The roadmap includes increasingly systematic support for:

• cross-layer dependencies,
• attribution,
• gradient attribution,
• mediation,
• feature interactions,
• candidate circuit discovery,
• context comparison,
• mechanism persistence,
• causal intervention.

The mature SAE steering environment already provides useful tools for testing some of these relationships.

For example:

If Feature A is suppressed, does the effect associated with Feature B disappear?

Do two representations only influence behavior when manipulated together?

Does intervening on an upstream representation change the predicted downstream internal state?

This is the transition from finding components to explaining computation.

Multimodal Representations and Internal Variables

The same research model should progressively extend beyond language.

The roadmap includes:

Language
Vision
Video
Audio
Multimodal Models
Learned Dynamic / World Models
Scientific Foundation Models

This expansion requires representations to be associated with more than token positions.

The platform therefore needs a generalized Activation Context:

Activation Context
   ├── token span
   ├── image patch
   ├── image region
   ├── audio interval
   ├── time/frequency region
   ├── video frame range
   ├── spatial-temporal region
   ├── latent state
   └── multimodal alignment

Vision

For a vision-language model, miStudio should eventually be able to follow an internal variable across:

Image
  ↓
Vision Encoder
  ↓
Projection / Adapter
  ↓
Language Model
  ↓
Response

A rich visual interface can show image regions, patch activations, representation changes across layers, and intervention effects.

The research question becomes not merely whether an object was recognized, but how that visual information is internally represented and transformed before contributing to language behavior.

Video and Temporal Models

Video is particularly important because it introduces persistent internal state.

Relevant internal variables may include:

• position,
• motion,
• velocity,
• trajectory,
• object persistence,
• relationships among objects,
• temporal order,
• state transitions,
• expected future state.

The research problem becomes:

Observation
   ↓
Internal State
   ↓
State Transition
   ↓
Prediction / Action

The same intervention framework can then ask whether manipulating an internal state variable causes the predicted counterfactual change.

Audio

Audio adds another type of structured representation involving time and frequency.

Candidate internal variables may include:

• phonetic information,
• speaker information,
• prosody,
• pitch,
• acoustic events,
• linguistic meaning.

Again, the objective is to trace how those variables are represented, transformed, and used.

Research-to-Runtime Integration

miStudio and miLLM become most powerful when the boundary between research and inference is explicit.

A successful miStudio experiment may eventually produce a validated Intervention Artifact describing:

Model / Revision
Representation System
Required SAE or Other Artifact
Target Representation(s)
Intervention Method
Parameters
Applicable Context
Validation Evidence
Expected Effects
Known Side Effects

The lifecycle becomes:

miStudio Experiment
       ↓
Intervention Tests
       ↓
Validation
       ↓
Intervention Artifact
       ↓
miLLM
       ↓
Instrumented Runtime Execution

Both products already integrate directly with Hugging Face for artifact acquisition.

miStudio supports:

• model download,
• SAE download/upload,
• dataset download/upload.

miLLM supports:

• model download,
• SAE download,
• dataset download.

miLLM therefore does not need miStudio to physically transfer every model or SAE used by a validated intervention.

A runtime artifact can identify exact Hugging Face dependencies, and miLLM can resolve them independently.

miStudio also currently publishes SAEs and feature activation labels to Neuronpedia.

The intended ecosystem model is deliberately open: the suite should orchestrate and interoperate with research repositories rather than attempting to replace them.

White-Box Evaluation

Most model evaluation operates entirely from externally visible behavior:

Input → Output → Score

Mechanistic instrumentation creates the possibility of another dimension:

Input
  ↓
Internal State
  ↓
Mechanism
  ↓
Output

miStudio should eventually allow researchers to define and validate assertions such as:

• an expected representation should be present,
• a prohibited representation should not be present,
• an expected mechanism should be active,
• a known failure signature should not appear,
• an internal state should be consistent with the resulting output.

These should not immediately become runtime rules.

They should first be treated as hypotheses that can be tested experimentally.

Once sufficiently validated, miLLM can execute them as white-box evaluations during inference.

This preserves the product division:

miStudio establishes whether an internal signal deserves to be trusted.

miLLM applies validated signals during runtime.

Conditional Mechanistic Control

The mature form of intervention is not necessarily constant steering.

It is contextual control.

Instead of:

Always amplify Feature X

a validated policy might eventually state:

WHEN internal state X is present
AND context Y applies

THEN
  apply intervention Z

UNTIL condition Q

The intervention may itself contain several representations.

For example:

Feature A  +1.2
Feature B  -0.7
Feature C  +0.4

miStudio would provide the environment for creating, visually inspecting, simulating, and validating such policies.

miLLM would provide runtime execution, visibility, and control.

MCP would allow agents and external research harnesses to interact with the same validated capabilities programmatically.

Runtime Evidence Should Return to Research

Research-to-runtime integration should not be one-way.

Real inference may expose:

• contexts absent from the original research dataset,
• unexpected representation combinations,
• intervention side effects,
• representation drift,
• mechanism failures,
• previously unseen internal states.

miLLM should therefore be able to return structured mechanistic evidence to miStudio.

       miStudio
 Research / Validation
          │
          ▼
        miLLM
 Runtime / Inference
          │
          ▼
       miStudio
 Refine / Replicate / Retest

The resulting feedback loop allows mechanistic understanding to improve through both controlled experiments and observed runtime behavior.

Capability Horizons

The roadmap is better understood as a sequence of capability horizons than as a rigid series of product releases.

Foundation and Orchestration

Build on existing capabilities with:

• richer Dataset Studio,
• workflow orchestration,
• scheduling,
• research campaigns,
• provenance and reproducibility.

This makes large mechanistic experiments easier to construct, repeat, and share.

Representation Expansion

Generalize the internal architecture to support:

• SAE representations,
• J-space,
• probes,
• directions,
• feature groups,
• subspaces,
• population representations.

This removes unnecessary dependence on any one interpretability technique.

Causal Experimentation

Build outward from the mature SAE steering subsystem and active J-space work into:

• J-space swapping,
• generalized interventions,
• feature-interaction studies,
• patching,
• ablation,
• representation replacement,
• quantitative causal validation.

The goal is to move progressively from descriptive evidence toward causal evidence.

Geometry and Emergence

Add tools for understanding:

• representation dimensionality,
• geometry,
• transitions across layers,
• emergence across training,
• changes across models and datasets.

This makes the suite capable of studying the formation and organization of internal variables.

Multimodal and Dynamic Systems

Extend the same research model into:

• vision,
• VLMs,
• video,
• temporal models,
• audio,
• multimodal systems,
• learned dynamic models.

The key architectural shift is from token-specific representations toward generalized internal state.

Operationalization

Connect validated research findings to miLLM through:

• portable intervention definitions,
• runtime representation observation,
• white-box evaluations,
• conditional interventions,
• mechanistic policies,
• structured runtime telemetry.

This closes the loop between interpretability research and instrumented inference.

Long-Term Direction

The suite is ultimately moving toward seven connected functions.

Discover

Find internal variables and representations using multiple interpretability methods.

Experiment

Manipulate them individually and in combination.

Understand

Determine what they encode, how they are structured, where they emerge, and how they change.

Relate

Determine how internal variables interact and form mechanisms.

Validate

Use controlled counterfactual intervention and replication to separate correlation from causal contribution.

Operationalize

Package sufficiently validated mechanistic knowledge in forms that can be consumed during inference.

Observe and Control

Use miLLM to inspect and, where justified, alter internal model behavior while retaining a feedback path into continued research.

The intended end state is therefore broader than a feature-discovery workbench or a steered inference server.

It is an integrated environment for investigating increasingly deep questions about learned systems:

What internal variables does the system develop?

How are they represented?

How do they change across layers, time, data, models, and modalities?

How do those variables interact?

Which interactions actually contribute causally to behavior?

Can those mechanisms be manipulated predictably?

Can validated mechanistic knowledge remain useful during real inference?

miStudio provides the rich experimental laboratory in which those questions can be explored visually, programmatically, and through agent-assisted research.

miLLM provides the rich instrumented runtime in which validated mechanistic understanding can be observed and applied.

The architecture deliberately does not assume that the important internal variable will turn out to be an SAE feature, a J-space coordinate, a direction, a subspace, a population code, a temporal state, or any particular representation form.

The suite should remain capable of following the internal structure wherever the learned system actually puts it.