Scale-Dependent ESR

The observed pattern that ESR appears predominantly in the largest model tested, suggesting scale-dependence

Neighborhood — ranked by edge-count

concept

Endogenous Steering Resistance
extends
The central phenomenon introduced by this paper: inference-time recovery from irrelevant activation steering in LLMs

cosine ≥ 0.65 · no typed edge

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

Does ESR reflect model scale, architecture, or training procedures?question0.778
Central unresolved question about the mechanism behind ESR's apparent size-dependence
We cannot isolate whether ESR reflects scale, architecture, or training procedures in Llama-3.3-70Bclaim0.751
Epistemic limitation claim acknowledging confounds in the cross-model comparison
ESR Rate (metric)concept0.751
Primary metric: percentage of responses containing multiple attempts that successfully improve on the first attempt
Implicit ESRconcept0.742
Form of ESR occurring without explicit verbal self-interruption markers, not captured by current metrics
Scale Inseparabilityconcept0.736
Central tenet: biological computation is fundamentally non-modular; molecular, cellular, and population scales bidirectionally co-determine each other without privileged level.
Explicit ESRconcept0.732
The form of ESR focused on in this paper, measured by verbal self-interruption phrases as segment boundaries
Scale Resilienceconcept0.720
Meta-problem: alignment techniques may collapse under rapid self-improvement or extreme complexity
ESR Testing Pipelinemethod0.714
Three-step protocol: (1) object-level prompting, (2) SAE-latent steering, (3) judge model scoring of attempts