Direct answer
An epigenetic steering vector is a compact control vector that changes model behavior without replacing the full somatic substrate. In ModelBreeder terms, it is best treated as an experimental adapter-like genome: small, inheritable, measurable, cheap to store, and always evaluated against the full runtime behavior it induces.
The source directive describes a low-dimensional vector as an “epigenetic steering wheel.” The site promotes that as a research pattern, not as a claim that a 20-dimensional vector universally captures arbitrary intelligence.
Interpretation
| Biological metaphor | Engineering interpretation |
|---|---|
| Somatic substrate | Frozen base model or local runtime package. |
| Epigenetic vector | Small learned or searched behavior-control vector. |
| Phenotype | Observable output behavior under defined prompts and tasks. |
| Hybrid vigor | Child vector improves a defined metric without excess cost. |
| Speciation threshold | Compatibility and diversity constraint that prevents redundant merges. |
Pseudocode: bounded steering experiment
PROCEDURE evaluate_steering_vector(base_model, tau_vector, eval_suite, resource_budget)
candidate <- ATTACH_STEERING_VECTOR(base_model, tau_vector)
behavior <- RUN_EVAL_SUITE(candidate, eval_suite)
resources <- PROFILE_RUNTIME(candidate)
score <- WEIGHTED_SUM(
utility = behavior.delta_utility,
calibration = behavior.delta_calibration,
diversity = behavior.delta_behavioral_distance,
cost = -resources.delta_cost
)
IF resources.exceeds(resource_budget) THEN
RETURN NO_OP("resource closure failed")
END IF
RETURN EVIDENCE_CARD(score, behavior, resources)
END PROCEDURECompatibility rule
A steering vector is not portable unless the base model family, tokenizer, insertion point, runtime implementation, and eval contract are named. Treat unscoped vectors as research artifacts, not reusable packages.
Good first experiment
Start with a mock model or a small local model. Create five steering vectors for five narrow styles or tasks. Evaluate each against utility, calibration, diversity, latency, and output length. Keep the champion, one complementary specialist, and one diverse challenger; retire the rest with evidence.
Source reports used for this guide
These reports are preserved verbatim in the site archive. The guide above is an editorial synthesis and may narrow, qualify, or reorganize claims from the source material.