PumpSense: Real-Time Detection and Target Extraction of Crypto Pump-and-Dumps on Telegram

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "PumpSense: Real-Time Detection and Target Extraction of Crypto Pump-and-Dumps on Telegram".

Jane: Here's the gist of the paper. The gist The paper introduces PumpSense, a system designed for real-time detection and target extraction of cryptocurrency pump-and-dump schemes coordinated via Telegram, addressing the lack of publicly available message-level labeled data in this area.

Tom: Data Collection and Annotation The research introduced a corpus of over 280,000 Telegram posts from 39 pumporganizing groups, all manually reviewed to identify 2,246 pump announcements and their targeted cryptocurrency and exchange <ref:2605.09431#pg2>.

Jane: The dataset contains a total of 283,017 messages collected from 39 different Telegram groups known for coordinating pump-and-dump fraud <ref:2605.09431#pg5>.

Tom: A pump announcement was defined as the message that explicitly announced the beginning of a pump <ref:2605.09431#pg6>.

Title and authors: Tom: So we're looking at this paper called "PumpSense: Real-Time Detection and Target Extraction of Crypto Pump-and-Dumps on Telegram." It seems like they're tackling a problem where we just don't have good data for this kind of manipulation, especially message-level labels.

Jane: That sounds really frustrating for anyone trying to track these schemes because the lack of labeled data makes it super hard to build reliable tools. So what is the core idea behind this paper?

Tom: The main thing they introduce is a system that can detect these pump-and-dump schemes in real time and then figure out exactly which cryptocurrency and exchange they are targeting. It's about using a big dataset from thirty-nine coordinating groups to do that <ref:2605.09431#pg1>.

Lu: What I see is they’re setting up two distinct tasks, real-time detection, and then target extraction based on those detections. That separation is important for building a practical system later on.

Meng: From where I sit in production, the part that sounds most promising is if this actually works fast enough to keep up with the speed of Telegram messages.

Lalam: It’s about using over two hundred eighty thousand Telegram posts from thirty-nine groups that they manually reviewed to find two thousand two hundred forty-six pump announcements and their specific crypto and exchange targets <ref:2605.09431#pg1,over 280,000 Telegram posts from 39>.

Tom: Right, so it's a massive dataset for training and testing their detection models against known examples of these schemes. It sets a new standard for having labeled data in this niche.

The paper's summary: Jane: So, to summarize what we just heard, the paper introduces PumpSense as a system designed to handle crypto pump-and-dump schemes coordinated on Telegram by giving us real-time detection and target extraction capabilities. What does that mean for someone listening who isn't deep in the weeds?

Tom: It means they’ve built a pipeline where you can watch messages and have the system flag potential pump announcements almost instantly, and then it tells you what coin they are pushing and which exchange they want to use.

Lu: The core methodology involves using a dataset of over two hundred eighty thousand Telegram posts from thirty-nine groups that they manually labeled to define what a pump announcement actually is <ref:2605.09431#pg1,of over 280,000 Telegram posts from 39>. They even defined it as the message that explicitly announces the beginning of a pump, and if that wasn't there, they inspect surrounding messages for missing info.

Meng: The focus on manual annotation is crucial because it shows how hard this data is to get naturally. It confirms that message-level labeled data for this specific fraud type just doesn't exist publicly in a usable form right now.

Lalam: And for extraction, they compare different approaches—rule-based methods, long-context transformers, and large language models—to see what gets the best results when pulling out the crypto and exchange names.

Tom: They found that LLMs are pretty strong here, achieving up to zero point nine one joint accuracy across all three extraction benchmarks they tested <ref:2605.09431#pg1>. That's a solid number when you compare it to older dictionary-based heuristics which they said were unreliable.

The paper's improvements: Jane: Okay, so we know what the paper does, but what are the specific tweaks or better ways they suggest doing things? They mentioned several improvements in their approach to make the system more robust.

Tom: One big thing is that they compared a lightweight tree-based LightGBM classifier against a transformer-based BGE-M3 for detection, and they found that LightGBM had a lower latency of nine point four microseconds per sample, which made it better for being a first-line detector <ref:2605.09431#pg1,a lightweight tree-based LightGBM classifier>.

Lu: They also showed that their proposed approach can achieve near instant pump detection right at the level of individual Telegram message windows. They noted that LightGBM got an F1 score of zero point seven nine with that low latency, while BGE-M3 was slower at fifty milliseconds per sample but got a slightly higher F1 score of zero point eight three <ref:2605.09431#pg1>.

Meng: From a deployment standpoint, the low latency of LightGBM seems like what you need for real-time monitoring on live traffic before things get too costly to run complex models on every message.

Lalam: For extraction, they found that LLMs like DeepSeek and Gemini are performing very well compared to a baseline Longformer model which only got a joint accuracy of zero point three seven. DeepSeek achieved zero point nine two cryptocurrency accuracy and zero point eight seven exchange accuracy when you look at the extraction task alone <ref:2605.09431#pg1>.

Tom: They also showed that the performance of LLMs in extraction was substantially better than those traditional approaches, even when considering the context construction optimization they did for their input windows, where a symmetric size of eleven or 20k features gave them an F1 score of zero point eight two.

Conclusion: Jane: So wrapping up this discussion on "PumpSense: Real-Time Detection and Target Extraction of Crypto Pump-and-Dumps on Telegram," the researchers are pointing toward a practical pipeline for monitoring these schemes by combining fast detection with powerful extraction tools. What’s the final message they want us to have about this work?

Tom: The main implication is that we can move away from just reactive analysis and build a system that actively flags these schemes before they fully execute, using lightweight models for speed and LLMs only when needed for detailed extraction.

Lu: They are also cautious about the extraction accuracy, noting that the performance of those LLMs can be pretty volatile depending on which specific model or prompt you use. They also pointed out that they have one hundred forty manually annotated cancelled pump events in their dataset, which could help us understand coordination failures better.

Meng: What I see is that this system isn't perfect, and the authors acknowledge that the performance of LLMs can vary a lot depending on how you set them up, which means deployment will require some fine-tuning.

Lalam: And they stress that this whole setup is designed for real-time deployment to protect investors before harm occurs in the market. It’s about creating a surveillance pipeline that adapts to new manipulation language through human verification and retraining.

Tom: So, PumpSense gives us a concrete framework—a way to combine fast, low-latency detection with high-accuracy extraction—for dealing with these Telegram-coordinated schemes. That's the essence of it.

Jane: It’s a useful foundation for anyone looking to apply this kind of structured approach to monitoring social media manipulation in finance and ensuring investors get a heads up before they lose money.