steering-coefficient-tuning

name

steering-coefficient-tuning

description

How to set the strength of any additive intervention on internal representations — steering vector, CAA, DAS dose-response, representation engineering, SAE feature scaling, ROME-style edits. Use whenever the plan sets a steering strength (`α`, `β`, `dose`, `magnitude`, `scale`, `coefficient`, `k`) to a fixed value or small range, especially when copied from a paper. Covers a coarse `layer × β` sweep (β in σ_proj units, scored on a target metric plus a fluency / general-ability metric), why the best coefficient is layer-dependent, the mid- vs late-layer behavior (late layers break into repetition / format-spam), and the rule to use the smallest sufficient β. Prevents two symmetric failures: TOO SMALL → effect drowned in noise → false "no causal effect"; TOO LARGE → off-distribution collapse → fluency breaks / random direction matches it → false "specificity fails". Triggers include `dose ∈ {-3..3}`, `α = 3`, "random direction beat my steering vector", "steering had no effect", "model output garbage after steering".

Steering Coefficient Tuning

The point

A steering coefficient that is too small does nothing; one that is too large damages the model's general ability and breaks fluent generation. You want the moderate range in between, and you find it by sweeping.

Three things to keep in mind:

1. The best coefficient depends on where you intervene — different layers (and different sites) need different coefficients. A value tuned at one layer does not transfer to another.

2. Mid layers usually work best, because of how "semantic maturity" varies with depth (task-dependent):

   • Early layers — closer to tokens / local patterns.
   • Mid layers — more often carry high-level control signals: behavior, intent, style, refusal.
   • Late layers — closer to logits and surface token choice. So even a slightly large coefficient at late layers tends to produce repeated tokens, format-symbol spam, or broken semantics.

3. Tuning tricks (condensed):

   • Express the coefficient in units of the projection std (β · σ_proj, σ_proj = std(hᵀu)) so values are comparable across layers/directions/models — a paper's "α = 3" almost always means 3σ.
   • Always sweep a β = 0 baseline, and log a fluency / general-ability metric alongside the target metric so you can see collapse, not just effect.
   • Prefer the smallest coefficient that already meets the target — don't push past it.
   • For long generations, add decay / gating so the push doesn't accumulate into collapse.

Recipe

1. For each candidate layer l, extract v_l = mean(h_pos − h_neg).
2. Build three forms: raw v_l, unit u_l, and mean-centered v_l.
3. Estimate the projection std per layer: σ_l = std(h_lᵀ u_l).
4. Coarse sweep:
   • layer ∈ [20%, 30%, 40%, 50%, 60%, 70%, 80%] of depth
   • β ∈ [0, ±0.25, ±0.5, ±1, ±2, ±4]
5. Score each point on target metric + fluency/general metric; keep the Pareto candidates.
6. Fine sweep β around the candidate layers.
7. If the target is already met, use the smallest β.
8. For long generation, add decay / gating.
9. Before shipping, test out-of-domain / general ability to confirm no broad damage.

Composition

Lock the site set via ../steering-block-selection/ first (the usable coefficient range is site-dependent), then tune β here on that locked site set. Re-tune whenever the site, direction, or model changes.