finding

active

finding:gpt-4-1-verbalized-probabilities-for-us-states-show-pearson-r-0-741-spearman-0-896-correlation-with-corpus-based-reference-frequencies

GPT-4.1 verbalized probabilities for US states show Pearson r=0.741, Spearman ρ=0.896 correlation with corpus-based reference frequencies

Indicates verbalized probabilities contain meaningful distributional information for constrained answer spaces

Source paper

extracted_from

Verbalized Sampling: How to Mitigate Mode Collapse and Unlock LLM Diversity

(2025) · Jiayi Zhang · Simon C.H. Yu · Derek Chong · Anthony Sicilia +3

Neighborhood — ranked by edge-count

Papers (1)

paper

Verbalized Sampling: How to Mitigate Mode Collapse and Unlock LLM Diversity
mentions

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.

Verbalized probabilities for programming languages show only Pearson r=0.182 (GPT-4.1) correlation with corpus frequencies, indicating weaker calibrationfinding0.874
Limits of verbalized probability calibration when corpus frequency and perceived popularity diverge
Spearman's rank correlation among different alignment metrics (CKA, SVCCA, Mutual k-NN, CKNNA) over 78 vision models is high across variants, with all p-values below 2.24×10^-105finding0.765
Validates robustness of alignment metric choice
Full-vector regression over 32 traits reaches leave-one-out Spearman correlations ~0.58-0.61 depending on layer windowfinding0.765
Vector geometry features provide moderate secondary signal beyond elicitation-only screen
Average Spearman correlation of Elo trait rankings between all three models increases to 0.87 after loving constitution character trainingfinding0.758
Post-training inter-model convergence in trait preferences demonstrating persona convergence
Pearson correlation of 0.74 between A/1/3450 activation and Arabic script proxy over 40M tokensfinding0.757
Joint measure of sensitivity and specificity for the Arabic script feature
SP ASR vs AS ASR correlation on Llama-3.1-8B is Pearson r=-0.492 (p=0.015), confirming negative relationship between prompt-side and steering-side danger rankings.finding0.751
Statistical evidence for SP/AS ranking inversion
82% of features in 1M SAE had maximum Pearson correlation ≤0.3 with any MLP neuron, and manual inspection showed no semantic resemblance.finding0.747
SAE features are not simply mirroring individual neurons.
Interest concept: Spearman ρ=0.76, isotonic R²=0.54 between logit self-report and probe score in LLaMA-3.2-3B (n=400)finding0.746
Strongest pooled introspective coupling across the four emotive concepts in the primary model