Conch: Multifaceted Research Constructs

• Conch is a polysemous term that defines various constructs including a visual-language foundation model in computational pathology, an interactive debate analysis system, a graph neural network, and notions in algebra and astronomy.
• In computational pathology, CONCH leverages a ViT-Base image encoder and autoregressive Transformer with contrastive pretraining, achieving notable zero-shot classification and segmentation performance across multiple benchmarks.
• Additional applications include semi-supervised classification in heterogeneous networks, formulation of conch maximal subrings in algebra, and the CONCH-SHELL catalog for exoplanet host-star identification.

Conch is a polysemous research term used for several unrelated constructs. In computational pathology, CONCH denotes “CONtrastive learning from Captions for Histopathology,” a task-agnostic visual-language foundation model built on the CoCa framework and pretrained on 1.17 million human H&E image-caption pairs (Lu et al., 2023). In other arXiv usage, Conch denotes an interactive visualization system for competitive debate analysis (Chen et al., 19 Jul 2025), a graph neural network for classification in heterogeneous information networks (Li et al., 2020), a notion of maximal subring in commutative algebra (Azarang, 2020), and the CONCH-SHELL catalog of nearby cool host stars for habitable-exoplanet searches (Gaidos et al., 2014).

1. CONCH as a computational-pathology foundation model

Lu et al. introduced CONCH as a visual-language foundation model specifically designed for computational pathology. Its core architecture comprises three modules: an image encoder $f(\cdot;\theta)$, a text encoder $g(\cdot;\phi)$, and a multimodal decoder $h(\cdot;\psi)$. The image encoder is a ViT-Base with 12 Transformer layers, 12 heads, embedding dimension $d=768$, and MLP hidden size 3072. Input images are tokenized into $16\times16$ patches with learned absolute positional embeddings. The image side uses a contrastive pooler with a single learned query to produce a 512-dimensional global image embedding, and a captioning pooler with 256 learned queries to produce 256 local image tokens for fine-grained decoding. The text side is a 12-layer autoregressive Transformer, while the multimodal decoder is a 12-layer GPT-style Transformer with cross-attention to the image tokens (Lu et al., 2023).

The pretraining objective combines symmetric InfoNCE alignment and caption generation. In compact form, several summaries write this as

$$L_{\mathrm{total}}=\lambda L_{\mathrm{contrastive}}+(1-\lambda)L_{\mathrm{caption}},$$

or equivalently as a weighted sum of an image-text contrastive term and an autoregressive captioning loss. Gustafsson and Rantalainen et al. describe the pretraining as two-stage: Stage 1 uses self-supervised iBOT on $\sim 16$ million tissue patches from $>21{,}000$ whole-slide images, and Stage 2 uses vision-language CoCa on 1.1 million automatically extracted pathology image-caption pairs from PubMed figures; in downstream grading experiments, the frozen image backbone is a ViT-Base with 12 transformer blocks, 12 attention heads per block, hidden dimension $d=768$, MLP-hidden size $\approx 3072$, and a final projection into a 512-dimensional patch embedding space (Gustafsson et al., 2024).

The pretraining corpus is described across papers as roughly 1.1-1.17 million histopathology image-caption pairs drawn from educational pathology sources, PubMed Central open-access figures, PubMed and related sources, and associated textual materials. In the original CONCH paper, Lu et al. further describe a three-stage curation pipeline involving YOLOv5 sub-image extraction, a GPT-style caption splitter, and a CLIP-based matcher, with final filtering for human tissues and H&E stains (Lu et al., 2023). This established CONCH as a domain-specific multimodal model rather than a direct adaptation of a general-purpose vision-LLM.

2. Transfer performance, feature extraction, and benchmarked use

In its original evaluation, CONCH was transferred without modification or with light fine-tuning across 13 benchmarks spanning classification, cross-modal retrieval, segmentation, and captioning. Reported zero-shot classification results include balanced accuracy $g(\cdot;\phi)$0 on TCGA BRCA, accuracy $g(\cdot;\phi)$1 on TCGA NSCLC, accuracy $g(\cdot;\phi)$2 on TCGA RCC, accuracy $g(\cdot;\phi)$3 on CRC100k, and quadratic $g(\cdot;\phi)$4 on SICAP Gleason. For zero-shot segmentation, Lu et al. report Dice $g(\cdot;\phi)$5 on SICAP and Dice $g(\cdot;\phi)$6 on DigestPath. For fine-tuned captioning on Source A, CONCH reaches METEOR $g(\cdot;\phi)$7 and ROUGE-1 $g(\cdot;\phi)$8 (Lu et al., 2023).

External benchmarking in weakly supervised computational pathology refines that picture. In “Benchmarking foundation models as feature extractors for weakly-supervised computational pathology,” CONCH yielded the highest performance when compared to vision-only foundation models, with Virchow2 as close second. The reported mean AUROC values are $g(\cdot;\phi)$9 for morphology, $h(\cdot;\psi)$0 for biomarkers, $h(\cdot;\psi)$1 for prognosis, and $h(\cdot;\psi)$2 across all 31 tasks. CONCH outperformed Virchow2 with statistical significance in 9/29 binary tasks, and a score-average ensemble of CONCH and Virchow2 reached mean AUROC $h(\cdot;\psi)$3, versus CONCH $h(\cdot;\psi)$4 and Virchow2 $h(\cdot;\psi)$5 (Neidlinger et al., 2024).

CONCH is also used as a generic feature extractor outside histopathology-specific benchmarks. In content-based medical image retrieval, the model is applied to 2D images and 2D slices of 3D volumes by resizing to $h(\cdot;\psi)$6, extracting the final $h(\cdot;\psi)$7 token, projecting to a 512-dimensional vector, and ranking database items by cosine similarity. On the two 3D MedMNIST datasets, CONCH achieves averaged mAP@5 $h(\cdot;\psi)$8, mMV@5 $h(\cdot;\psi)$9, ACC@1 $d=768$0, ACC@3 $d=768$1, and ACC@5 $d=768$2, placing it at the top of the 3D retrieval leaderboard in that study (Mahbod et al., 2024).

These results establish an important distinction. CONCH is not only a multimodal pretraining system; it is also routinely deployed as a frozen image encoder inside slide-level aggregation pipelines, kNN retrieval systems, MIL architectures, and zero-shot prompt-based inference procedures.

3. Representational structure, dimensionality, and robustness under shift

Mishra and Lotter analyzed six computational-pathology foundation models using representational similarity analysis on H&E image patches from TCGA. For each model $d=768$3, they construct a representational dissimilarity matrix

$d=768$4

with $d=768$5 patches per batch, and define inter-model similarity by the Spearman correlation of upper-triangular RDM entries: $d=768$6 In this analysis, CONCH is one of three vision-language contrastive-learning models. Its mean RSA to the other five models is $d=768$7, placing it in the middle of the representational-similarity spectrum; Prov-GigaPath has the highest average similarity at $d=768$8, while UNI2 and Virchow2 are lowest at $d=768$9 and $16\times16$0. The study explicitly reports that having the same training paradigm did not guarantee higher representational similarity (Mishra et al., 18 Sep 2025).

The same study quantifies slide-dependence and disease-dependence with Cliff’s Delta. CONCH’s slide-specificity is $16\times16$1 with range $16\times16$2, while disease-specificity is $16\times16$3 with range $16\times16$4. Under the stated convention, $16\times16$5 is “large,” $16\times16$6 is “medium,” and $16\times16$7 is “small,” so CONCH exhibits a moderately large slide-dependence and a small disease-dependence. Among the six models studied, CONCH has the lowest slide-specificity. Macenko normalization reduces CONCH’s slide-specificity from $16\times16$8 to $16\times16$9, a relative decrease of about $$L_{\mathrm{total}}=\lambda L_{\mathrm{contrastive}}+(1-\lambda)L_{\mathrm{caption}},$$0, and reduces disease-specificity from $$L_{\mathrm{total}}=\lambda L_{\mathrm{contrastive}}+(1-\lambda)L_{\mathrm{caption}},$$1 to $$L_{\mathrm{total}}=\lambda L_{\mathrm{contrastive}}+(1-\lambda)L_{\mathrm{caption}},$$2, a $$L_{\mathrm{total}}=\lambda L_{\mathrm{contrastive}}+(1-\lambda)L_{\mathrm{caption}},$$3 relative decrease (Mishra et al., 18 Sep 2025).

Intrinsic dimensionality is analyzed by mean-centering and concatenating embeddings into $$L_{\mathrm{total}}=\lambda L_{\mathrm{contrastive}}+(1-\lambda)L_{\mathrm{caption}},$$4, computing the singular values $$L_{\mathrm{total}}=\lambda L_{\mathrm{contrastive}}+(1-\lambda)L_{\mathrm{caption}},$$5, normalizing $$L_{\mathrm{total}}=\lambda L_{\mathrm{contrastive}}+(1-\lambda)L_{\mathrm{caption}},$$6, and examining the cumulative explained variance $$L_{\mathrm{total}}=\lambda L_{\mathrm{contrastive}}+(1-\lambda)L_{\mathrm{caption}},$$7. CONCH shows the sharpest early rise among all models, indicating the most compact embedding space; it reaches $$L_{\mathrm{total}}=\lambda L_{\mathrm{contrastive}}+(1-\lambda)L_{\mathrm{caption}},$$8 by $$L_{\mathrm{total}}=\lambda L_{\mathrm{contrastive}}+(1-\lambda)L_{\mathrm{caption}},$$9, whereas vision-only models such as UNI2 and Prov-GigaPath require $\sim 16$0 for the same coverage. The study further reports that vision-LLMs demonstrate relatively compact representations compared to the more distributed representations of vision-only models (Mishra et al., 18 Sep 2025).

Robustness under real distribution shift is substantially weaker than in-domain performance might suggest. Gustafsson and Rantalainen et al. evaluated frozen CONCH features in prostate cancer grading with ABMIL and reported in-domain performance of $\sim 16$1 on full PANDA cross-validation, but only $\sim 16$2 under Radboud$\sim 16$3Karolinska covariate/OOD shift and $\sim 16$4 under Radboud-U$\sim 16$5Karolinska-U. The paper states that the fact that foundation models have been trained on large and varied datasets does not guarantee that downstream models always will be robust to common distribution shifts (Gustafsson et al., 2024). A common misconception is therefore directly addressed in the literature: scale and multimodal pretraining do not by themselves ensure stain- or scanner-invariant deployment behavior.

4. Specialization without labels and zero-shot prompt design

Qiu et al. study annotation-free adaptation of CONCH by continued pretraining on domain- and task-relevant image-caption pairs mined from the cleaned Quilt1M corpus. Their workflow has three stages. First, image-caption pairs are retrieved by string matching to organ/site keywords for domain-specific pairs (“DAPT”) and then further filtered by class keywords for task-specific pairs (“TAPT”). Second, candidate pairs are ranked by cosine similarity under the frozen pretrained CONCH encoders,

$\sim 16$6

Third, both the image and text encoders are unfrozen and optimized with the same symmetric contrastive loss used originally by CoCa/CLIP, without captioning loss and without manual class labels. On SICAP, baseline zero-shot performance of $\sim 16$7 rises to $\sim 16$8 under TAPT with only $\sim 16$9 pairs; on BACH and MHIST, TAPT improves balanced accuracy by about 10–20 percentage points over baseline. The paper also reports that TAPT alone, with 16 shots of unlabeled data, matches CoOp at 8 labeled shots, and with 100 unlabeled shots, TAPT matches CoOp at 16 labeled shots (Qiu et al., 11 Aug 2025).

A complementary line of work studies prompt sensitivity rather than continued pretraining. In “Investigating Zero-Shot Diagnostic Pathology in Vision-LLMs with Efficient Prompt Design,” CONCH is evaluated on 3,507 digestive whole-slide images across seven tissue types with prompts that vary along four axes: detail specificity, anatomical precision, instructional framing, and output constraints. For invasiveness classification, the best prompt is $>21{,}000$0: “Given an image of a tissue sample stained with hematoxylin and eosin from the $>21{,}000$1, identify whether the sample is cancerous or not. Provide only a single word label.” With this prompt, CONCH reaches slide-level AUC $>21{,}000$2, versus $>21{,}000$3 for Quilt-LLAVA and $>21{,}000$4 for Quilt-Net; its overall average AUC across all nine prompts is approximately $>21{,}000$5. For dysplasia classification, CONCH achieves AUC $>21{,}000$6 with “Dysplasia,” $>21{,}000$7 with “Atypia,” and $>21{,}000$8 with “Precancerous” (Sharma et al., 30 Apr 2025).

These results also delimit several objective controversies. The paper reports that reduction of anatomical precision consistently degrades performance across all models, that “expert” framing did not improve and sometimes worsened AUC, and that model complexity alone does not guarantee superior performance: CONCH ($>21{,}000$9 M parameters) outperforms Quilt-LLAVA ($d=768$0 B) when prompts and pre-training are properly aligned to histopathology (Sharma et al., 30 Apr 2025). This suggests that downstream specialization can proceed either through representation-level continued pretraining on retrieved image-caption pairs or through carefully constrained prompt design at inference time.

5. Conch in analytic and graph-based artificial intelligence

Outside pathology, Conch names an interactive system for competitive debate analysis. The system proposed in “Conch: Competitive Debate Analysis via Visualizing Clash Points and Hierarchical Strategies” uses a three-stage pipeline: data ingestion and structuring, multi-level NLP extraction, and interactive visual encoding. Debate video is first transcribed with Whisper and then corrected with GPT-4. The corrected text is segmented into sessions, turns, and atomic claims by a few-shot GPT-4 prompt, with the resulting claim segmentation achieving over 93% precision and Fleiss’ $d=768$1 against expert annotation. Hierarchical information extraction then identifies clash points, disagreements, viewpoints, and refutation strategies; strategy recognition achieves 95% precision with Fleiss’ $d=768$2. The central Process View is a parallel spiral visualization defined by

$d=768$3

augmented by a chord diagram for cross-claim interactions. In a user study with 27 experienced debaters and coaches, the Conch group averaged 51.1 minutes, saving 32.8% over the video group and 45.9% over the text group, with significantly lower extraneous load and higher germane load (Chen et al., 19 Jul 2025).

ConCH is also the name of a graph neural network for semi-supervised classification in heterogeneous information networks. The model formulates learning as a multi-task objective,

$d=768$4

combining cross-entropy on labeled nodes with a self-supervised mutual-information loss on unlabeled nodes. Its key design choices are meta-path-guided neighborhoods, top-$d=768$5 PathSim filtering, explicit meta-path “context” embeddings, shallow graph convolutions on node-context bipartite graphs, and attention-based fusion across meta-paths. In the reported experiments on DBLP, Yelp-Restaurant, and Freebase-Movie, ConCH wins all 24 settings across training fractions $d=768$6. The paper further reports that it converges in around 100–200 sec, whereas MAGNN or HGT take 40–50× longer for the same or lower accuracy (Li et al., 2020).

These two usages share only the name. One is a human-centered visual analytics system driven by large-language-model prompting; the other is a multi-task GNN for scarce-label HIN classification.

6. Conch in commutative algebra and astronomy

In commutative algebra, Azarang defines a conch maximal subring of a commutative ring $d=768$7 through the intermediate notion of an $d=768$8-conch subring. A subring $d=768$9 is an $\approx 3072$0-conch subring if there exists a unit $\approx 3072$1 such that $\approx 3072$2, $\approx 3072$3, and $\approx 3072$4 is maximal among subrings satisfying those two conditions. Equivalently,

$\approx 3072$5

The paper proves an existence theorem: if $\approx 3072$6 with $\approx 3072$7 an integral domain and $\approx 3072$8 a prime element satisfying $\approx 3072$9 and $g(\cdot;\phi)$00, then $g(\cdot;\phi)$01 has a conch maximal subring. It also proves a dichotomy: either $g(\cdot;\phi)$02 has a conch maximal subring, or for every subring $g(\cdot;\phi)$03 one has $g(\cdot;\phi)$04. Further results address lifting conductors through integral extensions, affine $g(\cdot;\phi)$05-algebras with primes of height at least $g(\cdot;\phi)$06, and the characterization that a normal affine integral domain $g(\cdot;\phi)$07 is an integrally closed maximal subring of some larger ring $g(\cdot;\phi)$08 if and only if $g(\cdot;\phi)$09 (Azarang, 2020).

In astronomy, CONCH-SHELL abbreviates the “Catalog Of Nearby Cool Host-Stars for Habitable ExopLanets and Life.” Gaidos et al. present an all-sky catalog of 2,970 nearby, bright M- or late K-type dwarf stars selected from the SUPERBLINK proper motion catalog. The stars satisfy $g(\cdot;\phi)$10, are mostly within $g(\cdot;\phi)$11 pc, and 86% have been confirmed by spectroscopy. The catalog uses color, reduced-proper-motion, and absolute-magnitude cuts; spectroscopic vetting removes about 0.2% as hotter or evolved interlopers. Reported population statistics include 13% with H$g(\cdot;\phi)$12 emission, mean metallicity $g(\cdot;\phi)$13 with standard deviation $g(\cdot;\phi)$14 dex, and a resolved multiplicity fraction of $g(\cdot;\phi)$15 among stars observed with SNIFS. The catalog is then used to predict survey yields, including about 17 TESS detections at SNR $g(\cdot;\phi)$16, rising to about 27 at SNR $g(\cdot;\phi)$17, and infrared radial-velocity yields ranging from about 7 detections under $g(\cdot;\phi)$18 to about 32 under $g(\cdot;\phi)$19 (Gaidos et al., 2014).

Across these domains, “Conch” functions as a recurring label rather than a unified concept. The computational-pathology model is the most extensively developed technical usage in current AI literature, but the term also has established and mathematically precise meanings in algebra, visualization, graph representation learning, and exoplanet target selection.