method

active

method:layer-wise-anchoring-score-s-l-computation

layer-wise anchoring score S(ℓ) computation

Compute per-layer S(ℓ) = ρ̃d(ℓ) - d̃r(ℓ) - log k after whitening and standardization.

Neighborhood — ranked by edge-count

Frameworks (1)

framework

Unified Contextual Control Theory (UCCT)
implements
A theory that pretrained latent patterns are bound to task targets via external semantic anchors; formalized by anchoring strength S.

Concepts (1)

concept

anchoring strength S
implements
Composite score S = ρd − dr − log k predicting anchoring success.

Artifacts (1)

artifact

Semantic Anchoring in LLMs: Thresholds, Transfer, and Geometric Correlates
introduces
Main paper presenting UCCT and semantic anchoring framework.

Conceptual bridges

2-hop · via this method's ideas

Where ideas in this method connect to the rest of the corpus — the same concept, an analogy, or a restatement elsewhere.

anchoring strength S
~AIC/BIC Model Selection Criteria

Related by similarity (8)

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.

Peak Layer-wise Anchoring Score (S_max)concept0.832
Maximum of S(ℓ) across layers; geometry summary used to predict θ50
The anchoring score S is a predictive correlate of when anchoring succeeds and why small prompt changes yield threshold-like shifts.claim0.817
A central claim about the operational value of S.
(ii) does the anchoring score S = ρd − dr − log k consistently correlate with performance across anchoring methods?question0.787
Second research question in E2
Correlation between layer-wise S scores and task accuracy: ρ = -0.73, p < 0.001finding0.774
Shows S predicts anchoring effectiveness.
Math and code tasks show strongest mid-layer anchoring on LLaMA (S ≈ −1.65 at layers 8-12)finding0.774
Task-specific E3 finding showing compositional reasoning requires deeper processing
Layer-wise anchoring trajectoryconcept0.767
Plot of per-layer anchoring score S(ℓ) across model depth, revealing early dip, mid-layer peak, late standardization.
Layer-wise geometry summaries (Sbmax, AUSN) predict internal few-shot thresholds θ50claim0.762
Claim that geometry-to-behavior correlates exist
Math/code tasks S ≈ -1.65 at layers 8–12finding0.758
Task-specific peak anchoring score for structured reasoning domains.