Skip to main content
. 2026 Jul 29;14:e99639. doi: 10.2196/99639

Table 2.

Layer-level trajectory characteristics per model.a

Model Peakb (layer, MRRc) Troughd (layer, MRR) Final MRR Recovery PR expansione Cosine dropf
BGEg-base L1, 0.400 L4, 0.106 0.872 2.60 2.57× +0.236
GTEh-base L1, 0.408 L4, 0.110 0.881 2.59 2.63× +0.172
BioLORD L0, 0.602 L7, 0.414 0.798 2.04 1.90× +0.439
Nomic L1, 0.545 L6, 0.225 0.863 2.00 2.66× +0.313
Nomic-nopfx L1, 0.535 L6, 0.174 0.863 1.91 2.28× +0.357
MedCPTi L2, 0.033 L6, 0.025 0.816 Degeneratei 4.01× +0.242
BERT-base L2, 0.240 L7, 0.094 0.250 1.07 1.15× +0.009
BioBERT L1, 0.350 L7, 0.170 0.398 1.27 1.35× –0.030
ClinicalBERT L1, 0.167 L3, 0.102 0.230 1.97 1.38× +0.025
BioMistral L0, 0.440 L11, 0.035 0.424 0.96 1.45× +0.144
Phi-3-mini L0, 0.551 L12, 0.031 0.522 0.94 2.68× –0.236
E5-Mistral L0, 0.379 L11, 0.044 0.318 0.82 0.97× +0.036
E5-Mistral-ablation L0, 0.430 L11, 0.036 0.663 1.59 2.24× +0.149

aAll values are averaged across 6 corpus × query-format conditions. Final MRR values match the intervention-pipeline final-layer baseline; layer-wise trajectory extraction (streams A and G) produces final-layer MRR values differing by <1% from the intervention-pipeline baseline. Peak, Trough, PR expansion, and Cosine drop are computed from the layer-wise trajectory extraction. Values for new matched controls (BERT-base-uncased, BioMistral-7B) are highlighted.

bEarly peak=highest MRR@10 in the first quarter of layers (excluding layer 0 for MedCPT).

cMRR: mean reciprocal rank.

dTrough=lowest MRR@10 in the middle layers (35%-55% relative depth).

ePR expansion=participation ratio at the final layer divided by participation ratio at the trough.

fCosine drop=average cosine at the peak layer minus average cosine at the final layer (negative values indicate increased anisotropy from peak to final).

gBGE: Beijing Academy of Artificial Intelligence General Embedding.

hGTE: General Text Embeddings.

iMedCPT: layer-0 MRR@10=1.000 (excluded; reflects shared vocabulary embeddings in the dual-encoder architecture). Layers 1-6 produce near-zero MRR (range: 0.025-0.033) before a sharp recovery in the later layers; the recovery ratio, as defined in Methods, has a near-zero denominator (peak – trough ≈ 0.008) and is not meaningfully computable for this model. The PR expansion factor (4.01×) remains interpretable and is the largest in the panel.