Model Steering — Proving Causation

Steering is the definitive proof of your research. By manipulating specific features during generation, you demonstrate that a feature causally influences model behavior — not just correlates with it.

Steering Modes

miStudio provides three distinct steering modes:

Mode	What It Does	Best For
Individual	One feature at multiple strengths	Understanding a single feature's dose-response curve
Comparison	Multiple features at the same strength, side-by-side	Comparing related features (e.g., "French" vs "German")
Combined	All selected features applied simultaneously	Discovering synergistic effects between features

Strength Values

Residual-stream SAEs only

Steering adds each feature's direction at its layer's residual output, so it refuses MLP, attention, resid_pre and resid_mid SAEs. See Hook Types and What Accepts an SAE.

Steering strengths are raw coefficients added to the model's residual stream, compatible with Neuronpedia's scale:

Range	Effect	Example
0	No intervention (baseline)	Unsteered output
0.07 – 5	Subtle influence	Slight shift in topic or tone
5 – 50	Moderate effect	Clear behavioral change
50 – 100	Strong effect	Dominant feature influence
100 – 200	Very strong	Feature overwhelms other signals
200 – 300	Extreme	Often causes repetition or collapse
Negative	Suppression	Inhibits the feature's concept

Strength Calibration

The effective range depends on the SAE and layer. Start with strengths around 5–20 and increase gradually. Values above ±100 frequently cause the model to "collapse" into repetitive or incoherent output.

Multi-Strength Testing

Each feature supports up to 3 additional strengths tested simultaneously. Set a primary strength and additional values to see the dose-response curve in one generation pass. For example: primary=10, additional=[5, 20, 50].

Generation Parameters

Parameter	Default	Range	Effect
Max Tokens	100	1–2,048	Length of generated text
Temperature	0.7	0–2.0	Randomness. 0 = deterministic, 1.0 = creative
Top-P	0.9	0–1.0	Nucleus sampling threshold
Top-K	50	0–500	Vocabulary restriction per token
Repetition Penalty	1.15	0.5–2.0	Penalizes repeated tokens. Increase if output loops.
Seed	—	Optional	Set for reproducible results across runs

The Matrix Testing Workflow

Unlike tools with a single slider, miStudio uses a grid approach:

1. Select Features: Add up to 20 features (e.g., "Honesty" + "French Language")
2. Add Prompts: Multiple test prompts processed in batch
3. Configure Strengths: Set primary and additional strengths per feature
4. Include Baseline: Toggle "Include Unsteered" to compare against the natural output
5. Execute: miStudio runs all combinations, presenting results in a structured comparison view

Combined Mode Synergies

Steering multiple features simultaneously is NOT the same as running them separately. "Scientific Tone" + "Excitement" combined may produce different text than either alone. Combined mode reveals circuit behavior where features interact.

Viewing Results

After execution, results are presented in a structured comparison view showing baseline and steered outputs side-by-side with perplexity metrics:

Cluster strength budget

A feature added on its own gets a starting strength from its activation frequency. A whole cluster sent from the Clusters panel, or a circuit loaded into Steering, gets a computed allocation instead: one budget for the cluster, shared out between its members. The calculation runs on the server; the panel only redistributes the budget when you edit a strength.

A single feature's starting strength

S = clamp(2.9 − 2.6 · f, 1.0, 3.0)      rounded to 0.1

f is the feature's activation frequency, the fraction of tokens it fires on. Denser features start weaker. The tile shows an auto badge. A feature with no recorded frequency starts at 10 with a default badge. The line was fitted on one SAE (LFM2.5-1.2B layer 12, frequencies 0.037–0.484), so treat it as a starting point on any other SAE.

How a cluster's budget is computed

Each member brings its similarity to the cluster (from the Clusters panel), its activation frequency, and a sign: +1 to boost, −1 to suppress.

Step	Quantity	Rule
1	Weights wᵢ	wᵢ = sᵢ / Σ sⱼ over similarities. A missing similarity takes the mean of the known ones, and a negative one counts as 0.
2	Effective frequency f_eff	Σ wᵢ fᵢ / Σ wᵢ, over the members whose frequency is known
3	Direction budget B_dir	clamp(2.9 − 2.6 · f_eff, 1.0, 3.0): the single-feature line, applied to the cluster. With no known frequency it is the midpoint, 2.0.
4	Cohesion gain G	‖Σ σᵢ wᵢ dᵢ‖, the length of the direction actually injected, where dᵢ is the member's decoder column exactly as the steering hook adds it. Identical members give G = 1; members that point apart give less.
5	Total budget B	min(B_dir / max(G, 0.05)^γ, Σᵢ clamp(2.9 − 2.6 · fᵢ, 1.0, 3.0)). The exponent γ defaults to 0, so B = B_dir, capped at the sum of the members' single-feature strengths (2.0 for a member without a frequency).
6	Strengths	sᵢ = σᵢ · B · wᵢ, rounded to 0.1. The rounding remainder is handed out 0.1 at a time in weight order, never flipping a member's sign, so `Σ

Why γ = 0. Dividing by G keeps the injected vector at the single-feature magnitude, which is what the model was designed around. Validation on real clusters (2026-07-16) found that it overdrove by about 2×, so the fitted default ignores G for the size of the budget. G is still measured, shown, and used for the cancellation check.

A cluster of one member with no known frequency starts at 10, as a single feature does.

Flags

Flag	Meaning	What the panel does
low_cohesion	The cluster's cohesion is below the gate (0.5)	Keeps each member's single-feature strength, shows no budget bar, and says why
cancellation	Among the N boosted members, the weighted average of their unit-length decoder directions is shorter than 1/√N, so they partly cancel. cancellation_pair names the two most opposed members. Suppressed members are left out of this check	Shows the warning
cap_bound	The budget was capped at the sum of the members' single-feature strengths	Shows the warning
default_budget	No member has a known frequency	Shows the warning
approximate	The decoder could not be read, so G is taken as 1	Shows the warning
nonunit_decoder	A decoder column's length differs from 1 by more than 0.1; the frequency line was fitted on unit-length columns	Shows the warning
uniform_weights	No usable similarity; members share equally	Shows the warning
inactive_member	A member's weight is 0 (similarity 0), so it gets strength 0	Shows a warning. Its current wording, "rarely activates", describes frequency, but the flag is about similarity
grain_limited	The 0.1 strength grain kept `Σ	sᵢ

Editing strengths, pins and the intensity dial λ

While a budget governs the selection, a budget bar shows Σ |strength| against B, with G beside it. It turns amber when the strengths exceed the budget.

• Editing a strength pins that member. The rest of the budget, B − Σ |pinned|, is shared out again between the unpinned members by their weights, and each keeps its sign. When the pinned strengths alone exceed B, the unpinned members drop to 0. A pinned value is never rescaled.
• Unpin a member from its tile. It takes a share again at the next edit.
• B and G do not change when you edit strengths.
• λ (0–2, default 1) multiplies every strength when the request is sent, rounded to 0.1. Tiles keep showing the unscaled values, and λ = 0 previews the unsteered output.
• A saved cluster profile restores its strengths, budget and λ exactly. The budget is not recomputed.

Constants and the API

POST /api/v1/steering/cluster-allocation computes an allocation without loading the model. The body is sae_id, 1–20 members (feature_idx, layer, and optionally similarity, activation_frequency, sign, and a per-member sae_id), plus optional group_cohesion and circuit_id. The answer carries B, B_dir, G, f_eff, weights, strengths, flags, cancellation_pair, constants_used and formula_id (freq-budget/sim-alloc@1). It answers 404 for an unknown SAE, 400 for an SAE that is not ready or an index out of range, and 422 for an SAE recorded at a non-residual hook. A member on a layer other than its SAE's is 400 in a single-layer request. In a multi-layer request it is 422, listing every such member, and so is a layer whose members name two SAEs or a request naming more than 8 SAEs. A multi-layer request is answered per layer: see Per-layer budgets, one dial. Agents call the same computation through the MCP tool compute_cluster_allocation.

The constants a (2.9), b (2.6), m (1.0), M (3.0), cohesion_gate (0.5) and gamma (0) can be overridden, globally or per SAE, with the backend setting steering_cluster_constants_json: {"default": {...}, "per_sae": {"sae_…": {...}}}. A value that is not valid JSON is logged and ignored.

Multi-layer circuits — steering across SAEs

Everything above steers features from one SAE at one layer. A discovered circuit, though, spans layers — its members live on the layers each was found on, and each layer has its own SAE. Feature 015 makes steering follow the circuit: every member steers through the SAE trained on its own layer, in a single generation.

The own-layer rule

Each feature carries the SAE it belongs to. When you load a multi-layer circuit into the steering panel, the members arrive with their per-layer SAEs already set — an L13 member steers through the L13 SAE, an L14 member through the L14 SAE, in the same Blended run. You cannot steer an L14-trained feature at L10: that's rejected with a clear message naming the offending member, never quietly served through the wrong decoder (which would inject a direction from the wrong layer's basis — the bug this feature exists to fix).

Per-layer budgets, one dial

The budget model runs independently per layer — each layer gets its own budget bar (total budget B, cohesion gain G, similarity-weighted allocation) computed against that layer's SAE. One global intensity dial (λ) scales the whole circuit at once, so you tune the entire cross-layer behaviour with a single control while each layer keeps its principled starting strengths. The applied-features summary groups members by layer so you can see, and verify, that each one steered through its own SAE.

Hazard warnings — compounding and cancellation

Steering an upstream feature that drives a downstream feature you're also steering makes their influences compound (or, with opposite signs, cancel). miStudio surfaces this before you generate — an amber banner listing the pairs — but never silently corrects it; you decide.

The warning is as strong as the evidence:

• If the circuit has a validated edge between the pair (a rung-2 edge that survived causal validation), the warning is quantified from the measured effect size — "validated edge, ES=0.8 — combined influence ≈ higher than the naive sum."
• If there's no validated edge, a weight-prior heuristic (how aligned the upstream feature's output direction is with the downstream feature's input direction) still warns — but every such warning is explicitly labeled heuristic, never presented as causal. It's a hint to check, not a proven mechanism.

This is the payoff of the whole circuits arc: the same validated evidence that earns a circuit its rung-2 badge is what makes the steering hazard warning trustworthy instead of a guess.

VRAM

Each distinct SAE the circuit references loads onto the GPU (~130 MB for an 8k-feature SAE). A typical two- or three-layer circuit adds well under a gigabyte; only the SAEs the circuit actually uses are loaded, and exiting steering mode frees them all.