CoinCLIP: A Multimodal Framework for Assessing Viability in Web3 Memecoins

Hongyang Li; Hou-Wan Long; Wei Cai

REVIEW 3 major objections 4 minor 10 references

CoinCLIP: A Multimodal Framework for Assessing Viability in Web3 Memecoins

T0 review · 3 major / 4 minor · reviewed 2026-08-11 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read The paper claims that memecoin viability can be predicted at 84.7% accuracy by combining a frozen CLIP model with community comment data, and tests this on a new 6,231-token dataset.

desk verdict Useful new memecoin dataset, but the evaluation leaks the label into the features and the 92% AUC does not measure predictive viability as claimed. read the letter →

arxiv 2412.07591 v2 pith:I67QWQIN submitted 2024-12-10 cs.CE

Hou-Wan Long , Hongyang Li , Wei Cai This is my paper · ORCID

classification cs.CE

keywords memecoinWeb3multimodalclassificationCLIPcommunitydataPump.funRaydiumlistingviabilityprediction

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that the viability of a newly launched memecoin can be predicted from how it presents itself and how its community reacts, before any substantial trading history exists. It introduces CoinVibe, a dataset of 6,231 Pump.fun tokens with logos, textual descriptions, user comments, timestamps, and likes, and labels each token viable if it was later listed on the Raydium exchange. On top of this it builds CoinCLIP, which keeps a frozen CLIP model, adds per-modality projection layers and residual feature adapters, and injects comment data weighted by likes. On a held-out test split CoinCLIP reaches 84.72% accuracy, 92.07% AUC, and 83.74% F1, outperforming CLIP itself, CLIP-Adapter, and unimodal text or image classifiers. The point of caring is that most Pump.fun tokens are bot-generated or short-lived, and a cheap multimodal screen could filter them before investors commit.

What carries the argument

The mechanism is a frozen CLIP vision-language backbone plus three lightweight add-ons. Linear projection layers map each modality's CLIP embedding into a task-aligned space; Feature Adapters with residual connections let the model adjust to memecoin data without drifting from pre-trained features; and a community module encodes user comments through the CLIP text encoder, embeds normalized timestamps and likes, and aggregates comments by a like-weighted sum. Image and text representations are merged by Hadamard product, the community vector is concatenated, and a small MLP makes the final viable/non-viable decision. CLIP is the named central object: a contrastively pre-trained model that puts images and text in one embedding space.

What would settle it

Take the same CoinVibe features but split the dataset by creation date rather than randomly, or relabel viability by whether a token still has liquidity and a non-zero price 90 days after launch; if CoinCLIP's accuracy drops substantially, then its current performance depends on leakage from post-listing engagement or on a label proxy that does not measure long-term survival. A simpler check is to collect tokens that were scraped as non-viable but listed on Raydium days later and see whether the model flags them as viable.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that community interaction data—comments, timestamps, and likes—carries signal for memecoin viability that neither image nor text alone contains, and that this signal can be fused into a frozen CLIP backbone with lightweight adapters rather than expensive fine-tuning. The authors claim that separating image and text streams with independent projection layers and feature adapters, then combining them through a Hadamard product and concatenating an aggregated community representation, yields the best classifier for the CoinVibe task. They support this with a full benchmark: CoinCLIP beats CLIP by 3.4 accuracy points, CLIP-Adapter by 2.5, and the best unimodal model (CLIP image-only) by 5.6, with similar margins in AUC and F1. The ablation study attributes the gains to the projection layers, the dual feature adapters, and the community data integration in turn.

Load-bearing premise

The load-bearing premise is that a token being listed on Raydium is a faithful label for 'viable,' and that the scraped comments, timestamps, and likes reflect community interest before the listing rather than being inflated by the listing event itself.

Editorial extensions

If this is right

On coins like those in CoinVibe, multimodal screening reaches about 84.7% accuracy, which beats unimodal methods and plain CLIP by several points.
Separate adapters per modality outperform a single shared adapter, suggesting images and text need independent adjustment for this domain.
Community engagement features (comment volume, timing, likes) add measurable signal beyond the token's own image and description.
Image-only classification using CLIP's visual encoder already outperforms text-only methods, so visual branding matters more than narrative text for memecoin survival.
The frozen-backbone design means the method can be applied without expensive fine-tuning of large vision-language models.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If the Raydium-listing label is replaced by a longer-horizon outcome such as 90-day survival or trading volume persistence, the reported 84.7% accuracy is likely to decrease; that gap would measure how much of the model's signal is about reaching the listing threshold rather than about lasting value.
Because the article's features include comment likes and timestamps gathered at scrape time, a temporal split (train on earlier coins, test on later coins) is the natural stress test; the paper's random split may overstate generalizability.
The same architecture should transfer to other bonding-curve launchpads on other blockchains, and its community module could be extended to social-platform signals such as X/Twitter engagement without changing the fusion design.
At 83.7% macro F1, the false-negative rate for viable coins is still high enough that a practical deployment should treat the model as a filter rather than a validator.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Quick take: the CoinVibe dataset is the real contribution here; the CoinCLIP framework is a modest CLIP-Adapter variant, and the evaluation has a temporal leakage problem that makes the headline accuracy numbers untrustworthy as evidence of predictive viability.

What's actually new and good: 6,231 Pump.fun memecoins with text, logos, comments, timestamps, likes, labeled by Raydium listing. That's a new public dataset, and it's non-trivial to collect. They run multiple baselines, report mean/std over three seeds, and publicly release code and data. They also honestly state that CLIP-Adapter is conceptually similar, which is more than many papers do.

The soft spots are real. The biggest issue is temporal leakage. The label is whether the coin was listed on Raydium. The community features—comments, timestamps, likes—are scraped without any stated cutoff before the listing event. Listing drives attention and comments, so the model can learn 'this coin is attracting engagement after hitting the 69k threshold,' which is not the same as predicting viability from pre-listing information. The paper never frames this as a prediction task with a temporal boundary, but the abstract and intro clearly imply filtering before success. As it stands, the 92% AUC probably overstates what the model would do on genuinely new coins.

A second issue: they drop entries with missing descriptions or comments. Non-viable coins are more likely to have no comments, so this filtering biases the sample and inflates apparent accuracies. That's a design choice, but it should be discussed.

Third, the ablation in Table 2 is on a 'hate detection task' from CoinVibe, which is confusing—CoinVibe is a memecoin viability dataset. The accuracy values (71-76%) are incompatible with Table 1 (81-84%), suggesting it's a different task entirely. That needs clarification or correction.

On method novelty: separate per-modality adapters plus community-data concatenation is a small step from CLIP-Adapter. Not a fatal flaw, but it is what it is.

Who's this for? Researchers working on blockchain or crypto analytics, and anyone who wants an example of how easy it is to leak the label into the features in a retrospective dataset. A serious referee could help the authors fix the leakage by re-annotating with features as of a fixed time before the Raydium threshold, or re-framing the task as post-hoc classification.

Recommendation: send it to review. The dataset is worth publishing with the evaluation fixed; the current central claim doesn't stand, but it's fixable.

Referee Report

3 major / 4 minor

Summary. The paper introduces CoinVibe, a multimodal dataset of 6,231 Pump.fun memecoins (text descriptions, logos, user comments, timestamps, likes), labels coins as viable if they were listed on Raydium, and proposes CoinCLIP, a frozen-CLIP architecture with per-modality projection layers, feature adapters with residual connections, and community-data integration. On an 80/10/10 split of CoinVibe, CoinCLIP is reported to outperform unimodal and multimodal baselines, reaching 84.72% accuracy, 92.07% AUROC, and 83.74% macro-F1 (Table 1). An ablation study is presented in Table 2, and the code and dataset are made publicly available.

Significance. If the reported result measured true predictive viability, the paper would provide a useful benchmark and a reasonable strong baseline for a practically relevant Web3 problem. The stated strengths are the public release of code and data, a three-seed evaluation protocol with mean and standard deviation, and a broad comparison against unimodal and multimodal baselines. However, the central claim of assessing viability before success is not supported as reported, because the community features are collected without any demonstrated temporal cutoff relative to the Raydium-listing label. The evaluation therefore cannot distinguish predictive signal from leakage of post-listing engagement, and the ablation study is described as being on a different task, further weakening the empirical case.

major comments (3)

[§3.1, §3.3, §4, Table 1] The main load-bearing issue is temporal leakage between features and label. Section 3.3 defines viability by whether the coin was listed on Raydium, and Section 2 states that a coin must reach a $69,000 market capitalization to become eligible for that listing. Section 3.1 describes scraping comments, timestamps, and likes, but it does not state that these were collected before the listing event; the natural reading is that all community data were scraped after coin creation, which for viable coins is after the Raydium listing. Section 4 then uses these same comments, timestamps, and likes as features, aggregating comment representations by a like-weighted sum. Because post-listing engagement is a consequence of listing, the model can learn to detect 'this coin already attracted attention after crossing the listing threshold,' which is not predictive viability. The reported 92.07% AUROC in Table 1 therefore does not establish the paper's claim of assessing memecoin viability before success. The authors must provide a temporal split or otherwise demonstrate that community features were observed before the label was determined; without that, the central empirical claim is not supported.
[§3.2, §3.3] The data-cleansing step excludes tokens with missing descriptions or comments. This systematically removes non-viable coins that attracted no community engagement, and it changes the composition of the dataset relative to the population of all Pump.fun launches. The paper reports that 44.27% of the cleaned dataset is viable, but no comparison with the overall listing rate is provided. This filtering can inflate apparent separability because the remaining classes are conditioned on having at least some content and community activity. The paper should report the number of tokens removed at each cleansing step and show that the results are robust to this exclusion.
[Table 2, §5.3] The ablation study does not support the conclusion drawn from it. Table 2's caption states that the ablation was conducted on the 'hate detection task from the CoinVibe dataset,' while the main experiments and the paper's stated contribution concern memecoin viability classification. The accuracy values in Table 2 (71.23–76.44) are also incompatible with the viability accuracies in Table 1 (70.17–84.72 for the corresponding configurations). The progressive integration of projection layers, feature adapters, and community data is therefore not shown to contribute to the viability result. The paper should either report the ablation on the same task and split as Table 1 or explicitly state that Table 2 concerns a separate task and discuss why that evidence bears on CoinCLIP's viability performance.

minor comments (4)

[§6] The conclusion mentions 'semantic-aware initialization' as a contributing component, but this component is not introduced in Section 4 and does not appear in the ablation table. The manuscript should either describe and evaluate it or remove the reference.
[§3.2] The text says that corrupted or missing images are 'replaced with zero vectors,' but a zero vector is not a standard valid input to the CLIP image encoder. Please clarify whether the replacement happens before encoding, whether the encoder sees a zero image, or whether a zero embedding is used instead.
[§4] The term 'disentangle' is used for the linear projection layers, but there is no training objective or evaluation that enforces or measures disentanglement. Consider using a more neutral description such as 'separate linear projections for each modality.'
[Abstract and §3.3] The abstract frames the goal as filtering out 'low-quality or bot-driven projects,' but the operational label is Raydium listing, which is a market-capitalization threshold and not a direct measure of quality or bot activity. The gap between the stated motivation and the actual label should be acknowledged explicitly.

Circularity Check

1 steps flagged · score 6.0 of 10

CoinCLIP's community features can encode the label's own causal effects because viability is defined by Raydium listing and no temporal cutoff is imposed.

self definitional [Section 3.3 (Data Annotation); Section 2 (Pump.fun narrative); Section 4 (Community Data Integration)]
"Viability was determined by whether the memecoin was successfully listed on Raydium, a decentralized exchange (DEX) on the Solana blockchain. Tokens listed on Raydium were labeled as 'Viable', while those not listed were labeled 'Non-Viable'. This classification is based on the premise that being listed on Raydium indicates market validation, strong community engagement, and financial health, making it a reliable indicator of long-term potential. // We enhance the model by incorporating community data, including user comments, timestamps, and likes."

The outcome is defined by the Raydium-listing event, and Section 2 says that event grants the coin 'broader exposure within the Solana DeFi ecosystem.' The model's input then includes exactly the community engagement that listing both signals and amplifies: comments, timestamps, and likes, aggregated by like-weighted sum. Because the paper never specifies that comments were collected before the listing date, a coin labeled 'viable' can contribute its own post-listing engagement as features. A high AUC can therefore be achieved by recognizing consequences of the label rather than independent predictors of long-term viability; the claimed 'prediction' partly restates the label's effects.

full rationale

There are no load-bearing self-citations or imported uniqueness theorems here; the architecture is a standard CLIP-Adapter-style design with frozen CLIP encoders, linear projections, feature adapters, and an MLP classifier. The circularity lies in the construction of the task. Section 3.3 operationalizes viability as Raydium listing and explicitly justifies that choice by saying listing indicates 'strong community engagement'; Section 4 then feeds user comments, timestamps, and likes into the classifier. Since Section 2 states that Raydium listing brings 'broader exposure,' community engagement is partly caused by the label event, and no temporal cutoff is described. The random 80/10/10 split in Section 5.1 cannot remove that leakage, so the reported 84.72% accuracy and 92.07% AUC are not evidence that CoinCLIP assesses viability before the outcome is known. Additionally, the ablation in Table 2 is internally inconsistent: its caption refers to a 'hate detection task' and its complete-model accuracy of 76.44% does not match Table 1's 84.72%, so the component-wise support is unreliable. These are correctness and validation concerns rather than self-citation problems; they still make the central empirical claim partially circular by construction of the label/feature pairing.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the validity of the Raydium-listing label, the timeliness of scraped community features, and the representativeness of the comment-bearing sample. The learned model parameters are numerous and undisclosed, making the reported gains hard to audit.

free parameters (3)

Projection layer weights = Not reported
Learned on CoinVibe; dimensions and initialization are not given.
Feature adapter weights = Not reported
Learned on CoinVibe; depth and width of adapters are not specified.
MLP classifier weights = Not reported
Learned on CoinVibe; layer count, hidden size, learning rate, and number of epochs are not specified.

assumptions (4)

domain assumption CLIP ViT-L/14 frozen encoders provide useful representations for memecoin logos and descriptions.
Invoked in Section 4 under Zero-shot CLIP; no evaluation of domain shift for memecoin imagery is provided.
ad hoc to paper Raydium listing is a reliable indicator of viability.
Defined in Section 3.3 as the labeling rule; it is a $69,000 market-cap threshold, not a measure of long-term success.
ad hoc to paper Community comments and likes are collected independently of the listing outcome.
Sections 3.1 and 4 describe scraping and use of likes and timestamps without any temporal cutoff before Raydium listing; this assumption is likely violated.
ad hoc to paper Tokens without comments or descriptions can be excluded without biasing the dataset.
Section 3.2 excludes missing entries; fewer than 1% of Pump.fun launches are represented, so the sample is biased toward engaged tokens.

how reviews work

0 comments

Cite this review

Pith. "Pith review of CoinCLIP: A Multimodal Framework for Assessing Viability in Web3 Memecoins." pith.science (2026). https://pith.science/paper/I67QWQIN

@misc{pith2026241207591,
  author       = {Pith},
  title        = {Pith review of: CoinCLIP: A Multimodal Framework for Assessing Viability in Web3 Memecoins},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/I67QWQIN}},
  note         = {Machine review of arXiv:2412.07591}
}

read the original abstract

The rapid growth of memecoins within the Web3 ecosystem, driven by platforms like Pump.fun, has made it easier for anyone to create tokens. However, this democratization has also led to an explosion of low-quality or bot-generated projects, often motivated by short-term financial gain. This overwhelming influx of speculative tokens creates a challenge in distinguishing viable memecoins from those that are unlikely to succeed. To address this issue, we introduce CoinVibe, a comprehensive multimodal dataset designed to evaluate the viability of memecoins. CoinVibe integrates textual descriptions, visual content (logos), and community data (user comments, timestamps, and number of likes) to provide a holistic view of a memecoin's potential. In addition, we present CoinCLIP, a novel framework that leverages the Contrastive Language-Image Pre-Training (CLIP) model, augmented with lightweight modules and community data integration, to improve classification accuracy. By combining visual and textual representations with community insights, CoinCLIP provides a robust, data-driven approach to filter out low-quality or bot-driven projects. This research aims to help creators and investors identify high-potential memecoins, while also offering valuable insights into the factors that contribute to their long-term success. The code and dataset are publicly available at https://github.com/hwlongCUHK/CoinCLIP.git.

Figures

Figures reproduced from arXiv: 2412.07591 by the authors.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

10 extracted references · 4 canonical work pages

[1]

Giovanni Burbi, Alberto Baldrati, Lorenzo Agnolucci, Marco Bertini, and Alberto Del Bimbo. 2023. Mapping memes to words for multimodal hateful meme classifi- cation. In Proceedings of the IEEE/CVF International Conference on Computer Vision . 2832–2836

work page 2023
[2]

Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. BERT: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)

arXiv 2018
[3]

Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xi- aohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. 2020. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)

arXiv 2020
[4]

Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. 2024. CLIP-Adapter: Better Vision-Language Models with Feature Adapters. International Journal of Computer Vision 132, 2 (2024), 581–595

work page 2024
[5]

Douwe Kiela, Hamed Firooz, Aravind Mohan, Vedanuj Goswami, Amanpreet Singh, Pratik Ringshia, and Davide Testuggine. 2020. The Hateful Memes Challenge: Detecting Hate Speech in Multimodal Memes. In Advances in Neural Information Processing Systems

work page 2020
[6]

Chen Ling, Ihab AbuHilal, Jeremy Blackburn, Emiliano De Cristofaro, Savvas Zan- nettou, and Gianluca Stringhini. 2021. Dissecting the meme magic: Understanding indicators of virality in image memes. Proceedings of the ACM on human-computer interaction 5, CSCW1 (2021), 1–24

work page 2021
[7]

Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sand- hini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al
[8]

Siddhant Bikram Shah, Shuvam Shiwakoti, Maheep Chaudhary, and Haohan Wang. 2024. MemeCLIP: Leveraging CLIP Representations for Multimodal Meme Classification. arXiv preprint arXiv:2409.14703 (2024)

arXiv 2024

Show all 10 references

[9]

Shardul Suryawanshi, Bharathi Raja Chakravarthi, Mihael Arcan, and Paul Buite- laar. 2020. Multimodal meme dataset (MultiOFF) for identifying offensive content in image and text. In Proceedings of the second workshop on trolling, aggression and cyberbullying. 32–41

2020
[2021]

In International Conference on Machine Learning

Learning transferable visual models from natural language supervision. In International Conference on Machine Learning . PMLR, 8748–8763

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.

[1] [1]

Giovanni Burbi, Alberto Baldrati, Lorenzo Agnolucci, Marco Bertini, and Alberto Del Bimbo. 2023. Mapping memes to words for multimodal hateful meme classifi- cation. In Proceedings of the IEEE/CVF International Conference on Computer Vision . 2832–2836

work page 2023

[2] [2]

Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. BERT: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)

arXiv 2018

[3] [3]

Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xi- aohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. 2020. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)

arXiv 2020

[4] [4]

Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. 2024. CLIP-Adapter: Better Vision-Language Models with Feature Adapters. International Journal of Computer Vision 132, 2 (2024), 581–595

work page 2024

[5] [5]

Douwe Kiela, Hamed Firooz, Aravind Mohan, Vedanuj Goswami, Amanpreet Singh, Pratik Ringshia, and Davide Testuggine. 2020. The Hateful Memes Challenge: Detecting Hate Speech in Multimodal Memes. In Advances in Neural Information Processing Systems

work page 2020

[6] [6]

Chen Ling, Ihab AbuHilal, Jeremy Blackburn, Emiliano De Cristofaro, Savvas Zan- nettou, and Gianluca Stringhini. 2021. Dissecting the meme magic: Understanding indicators of virality in image memes. Proceedings of the ACM on human-computer interaction 5, CSCW1 (2021), 1–24

work page 2021

[7] [7]

Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sand- hini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al

[8] [8]

Siddhant Bikram Shah, Shuvam Shiwakoti, Maheep Chaudhary, and Haohan Wang. 2024. MemeCLIP: Leveraging CLIP Representations for Multimodal Meme Classification. arXiv preprint arXiv:2409.14703 (2024)

arXiv 2024