NormProbe for less-aligned layers
A normalized projection probe reweights and maps earlier candidate layers into the retrieval space.
Motivation
Late-interaction retrievers preserve fine-grained signals but require many vectors per page. Dense retrievers store one vector, but typically rely on the final layer alone. MINER studies whether internal representations can improve that single-vector embedding without changing its dimensionality.
Layer-wise analysis
Normalized CKA identifies layers with retrieval-relevant structure. The alignment ratio then distinguishes earlier, less-aligned representations from the final aligned regime.
Which layers preserve retrieval structure?
Normalized CKA measures whether a layer organizes pages and queries similarly to the final retrieval space.
Layers 15–36
22 of 36 layers selected at CKA = 0.60
Markers show the measured fraction of Jina’s 2,048 neurons retained at the default cutoff.
At the default cutoff, Jina selects layers 15–36. Earlier candidates are less aligned; the last three selected layers are already aligned with the final retrieval space.
Method: retrieval-aligned layer probing and adaptive sparse fusion
MINER aligns layer-wise representations to cross-modal final-layer anchors, then combines the selected representations into a single vector with the same dimension as the original dense embedding.
BaseProbe for aligned layers
Element-wise feature weights preserve the geometry of the last three selected layers, which are already directly aligned.
Adaptive sparse multi-layer fusion
Validation utility determines each layer’s feature retention. A learned weighted sum and global bias produce one D-dimensional embedding.
Experiments and results
We evaluate MINER with Jina-Embeddings-v4, Eager-Embed-v1, and MoCa-3B on the ViDoRe V1, V2, and V3 visual document retrieval suites.
8 / 9
backbone × suite averages improved
+4.5%
largest relative average gain
42.4×
smaller than Jina late interaction · ViDoRe V2
5.3×
faster than Jina late interaction · ViDoRe V2
Jina
Dense53.3
MINER55.2
Eager
Dense56.0
MINER58.5
MoCa
Dense58.3
MINER59.1
MINER improves all three backbone averages on ViDoRe V2 and eight of nine averages across the complete evaluation. The gains come from better use of the corresponding retriever’s internal representations, while the backbone remains frozen and the output remains a single dense vector.
View backbone averages
| Benchmark | Metric | Backbone | Dense baseline | MINER | Relative change |
|---|---|---|---|---|---|
| ViDoRe V1 | nDCG@5 | Jina | 84.3 | 84.4 | +0.1% |
| ViDoRe V1 | nDCG@5 | Eager | 82.5 | 84.8 | +2.8% |
| ViDoRe V1 | nDCG@5 | MoCa | 85.7 | 85.6 | −0.1% |
| ViDoRe V2 | nDCG@5 | Jina | 53.3 | 55.2 | +3.6% |
| ViDoRe V2 | nDCG@5 | Eager | 56.0 | 58.5 | +4.5% |
| ViDoRe V2 | nDCG@5 | MoCa | 58.3 | 59.1 | +1.4% |
| ViDoRe V3 | nDCG@10 | Jina | 49.0 | 49.7 | +1.4% |
| ViDoRe V3 | nDCG@10 | Eager | 47.6 | 48.8 | +2.5% |
| ViDoRe V3 | nDCG@10 | MoCa | 49.1 | 49.5 | +0.8% |
Retrieval quality and serving cost
Because MINER still emits one D-dimensional vector, it uses the same dense index and dot-product search procedure as its backbone.
Showing retrieval quality against index size.
Dense Jina
MINER
Late interaction
Full-corpus Qdrant measurements; higher nDCG and QPS are better, lower storage is better.
On Jina and ViDoRe V2, MINER matches the dense baseline’s 22.24 MB index and retains 93% of its search throughput while improving average nDCG@5 by 1.9 points. Relative to Jina late interaction, it uses 42.4× less index storage and searches 5.3× faster.
View efficiency measurements
| Method | Vectors per page | QPS | Index storage | Average nDCG@5 |
|---|---|---|---|---|
| Dense Jina | 1 | 15.41 | 22.24 MB | 53.3 |
| MINER-Jina | 1 | 14.36 | 22.24 MB | 55.2 |
| Jina late interaction | Many | 2.75 | 943.38 MB | 57.6 |