Larger Sbmax associated with smaller θ50 in E3 sweep

Geometry-to-behavior correlate within E3.

Source paper

extracted_from

(2025) · Edward Yi Chang · Kaya, Zeyneb N. · Ethan Chang

claim

community

Few-shot anchoring & latent structure
members_of
How minimal examples disambiguate and recruit latent arithmetic/reasoning interpretations in LLMs
Layer-wise geometry predicting few-shot learning
members_of
Silhouette-based metrics (Sbmax, AUSN) across LLM layers predict task accuracy and few-shot thresholds.
Neural activation geometry and behavioral prediction
members_of
Quantifies relationships between layer-wise activation statistics (Sbmax, AUSN) and task performance metrics in LLMs, bridging internal representation geometry to behavioral outcomes.

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.

Larger S_max correlates with smaller θ50 across backbones in E3 (negative association consistent across pooling and metric choices)finding0.830
Key geometry-to-behavior bridge finding in E3; robust to pooling choice, cosine vs. L2, and frozen external encoder
AUS_N is a weaker correlate of θ50 than S_max across E3 backbonesfinding0.768
E3 finding distinguishing the two geometry summaries; breadth less predictive than peak height
Layer-wise geometry summaries (Sbmax, AUSN) predict internal few-shot thresholds θ50claim0.766
Claim that geometry-to-behavior correlates exist
Sbmax mean -1.896 ± 0.211finding0.749
Geometry summary peak anchoring score averaged over seeds.
Multi-attempt rate peaks at 2.7% around -0.3σ below threshold in boost sweep experimentfinding0.742
Quantitative characterization of ESR operating regime in boost level sweep
Uncalibrated sweep units and restricted coefficient ranges are the primary cause of prior reports showing P2 outperforming MD injectionsclaim0.731
Mechanistic explanation for discrepancy with Banayeeanzade et al.; addressed by centroid unit and unbounded sweep contributions
LLaMA-3.1-8B: Sbmax = -1.896 ± 0.211, AUSN = -2.119 ± 0.198, peak layer ℓ* = 10 (median)finding0.730
Seed-pooled geometry-only statistics (per-dev z units).
Boost Level Ablation Sweepmethod0.725
Systematic sweep of 10 boost levels from threshold-3σ to threshold+3σ to characterize ESR vs. steering strength