Resisting Manipulative Bots in Meme Coin Copy Trading: A Multi-Agent Approach with Chain-of-Thought Reasoning

AI summary

Overview

Research area: Artificial intelligence (LLM-powered multi-agent systems, chain-of-thought reasoning) applied to blockchain economics and decentralized finance, specifically meme coin copy trading on Solana.

Technical level: Advanced. The paper combines large language model agent design, chain-of-thought prompting, on-chain bot detection algorithms, machine-learning prediction baselines, and formal bonding-curve return modeling with proofs.

Scope (one sentence): The paper builds and evaluates a GPT-4o-powered, chain-of-thought multi-agent system that decides which trader wallets and meme coins a copier should follow, using on-chain and social data from 6,000 meme coin projects, while formally characterizing four types of manipulative bots and the mechanical cost of imitating another trader.

What This Paper Is About

Copy trading lets users automatically replicate the trades of wallets that appear to be "smart money," and it has become the dominant way people enter meme coin markets. Because meme coins are extremely illiquid and volatile, and because many copiers are inexperienced, adversaries can deploy automated bots to front-run, hide their real positions, and fabricate hype, then sell into the copy traders they attracted. The paper's goal is to detect and quantify these bots and to build a copy-trading decision system that resists them.

Key Contributions

1. A multi-agent meme coin copy-trading system. The authors propose an LLM-based multi-agent framework with chain-of-thought reasoning, which they state is the first application of a multi-agent system to copy trading. It decomposes copy trading into three subtasks — coin evaluation, wallet selection, and timing assessment — handled by a wallet agent, a coin agent, and a timing agent on GPT-4o.
2. A systematic analysis of manipulative bots in copy trading. The authors formalize manipulation heuristics, develop detection algorithms, and empirically quantify bot activity. They cover position manipulation bots (naïve bundle, bundle bot, gradual bundle, sniper bot) and attention manipulation bots (bump bot, comment bot).
3. A structured feature space and the predictability of trader profitability. The features capture bot presence plus traders' historical performance (across short-, medium-, and long-term horizons) and experience, and the authors show trader profitability is predictable, giving a foundation for copy trading.
4. Bot-resilient evaluation. The system is tested on historical wallet-level data from 6,000 meme coins and compared against zero-shot and statistic-driven baselines on both prediction accuracy and economic return.

Main Findings

• Bot categories and detection rules: Bundle bots are detected by non-creator buys in the meme coin creation block (Algorithm 1); sniper bots by non-creator buys within the first K blocks after launch, with K defaulting to 5 (Algorithm 2); bump bots by a "flip" score, defined as the number of flips divided by net position change with a small constant default of 1, classified as bump bots when the score is at least 50 (Algorithm 3); comment bots by LLM classification of hype-style, contextless comments, with a coin labeled as exhibiting comment bot activity if more than one comment is classified as bot-generated.
• Position manipulation bots mostly redistribute rather than change project outcomes. Performance distributions with and without bundle bots or sniper bots largely overlap. Bundle bots appear in roughly one quarter of projects and are weakly associated with lower returns and shorter dump durations; sniper bots are widespread but show minimal observable performance effect. The authors conclude these bots mainly shift gains from slower retail traders to early or better-positioned adversaries.
• Attention manipulation bots are associated with better performance. Bump bots and comment bots are associated with improved performance metrics, though bump bots occur less frequently, which the authors attribute to their higher operational cost.
• Gradual bundles are the hardest case. Because creation and purchases are decoupled over time and may involve multi-layer funding networks, temporal signals and attribution are obscured; the authors note that experienced traders may rely on candlestick patterns, since gradual bundles show mechanically regular patterns versus more organic non-manipulative ones.
• Prediction accuracy: The multi-agent system with few-shot chain-of-thought outperforms the zero-shot multi-agent system baseline across nearly all thresholds on Precision and F1. XGBoost achieves the highest AUC overall, followed closely by the multi-agent system with chain-of-thought; the neural network and LASSO have lower AUC, and the zero-shot multi-agent system performs worst. The authors report the framework outperforms zero-shot and most statistic-driven baselines in prediction accuracy.
• Economic performance: The smart-money wallets identified by the framework achieve an average return of 14%, while the corresponding estimated copier return is 3% per meme coin investment under realistic market frictions. The authors state their approach outperforms all baselines in economic performance.
• The imitation penalty is a structural effect, not an implementation flaw. Theorem 2 shows that under the pump.fun bonding-curve execution model, a copier strictly overpays on every replicated buy, so the imitation penalty (smart money return minus copier return) is strictly positive.
• Manipulation is modeled as near zero-sum. Adversary profits come directly from the losses of copy traders who follow manipulated on-chain and off-chain signals.

Methodology in Plain English

The authors start with data: account-level transaction and transfer records from Flipside for the Solana ledger, plus off-chain comment data from pump.fun, covering 6,000 meme coin projects with their creators, traders, and complete trading histories. They then write down, in algorithmic form, the behavioral fingerprints of each bot type — same-block coordinated buys, ultra-early buys within a few blocks of launch, repeated buy-sell flips with near-zero net position change, and generic hype comments. These rules become four "bot" indicator features.

For prediction, they frame the question as: given a wallet's first position in a meme coin, will that position end up profitable (TRUE) or not (FALSE)? They order each trader's first transaction chronologically to avoid information leakage and split data 70%–15%–15% into training, validation, and test sets, with the label being the sign of realized profit.

Two model families are compared. The statistic-driven models are a LASSO linear model, XGBoost, and a neural network. The multi-agent system uses three specialized GPT-4o agents with few-shot chain-of-thought prompts. Knowledge is injected by first defining conditions for each feature — boolean constraints (e.g., bundle bot must be False), visual or language-based conditions (candlestick must not show a mechanical uptrend; comments must be organically positive), statistical conditions (e.g., t-statistic above 1.645), and continuous conditions tied to percentiles of the training set (e.g., purchase amount above the 25th percentile of the training set). Representative examples satisfying these conditions are annotated by hand into agent-specific chains of thought and embedded in the prompts.

Each agent outputs a TRUE/FALSE decision plus token logits; the linear probability of the TRUE token is read as a confidence score. The three scores form a probability vector, initialized with uniform weights, and the weights are tuned on the validation set to maximize the AUC of the aggregated prediction. The final prediction is a weighted combination of agent confidences evaluated on the test set. A zero-shot multi-agent version with everything else held constant serves as a chain-of-thought ablation.

Finally, to reason about economics rather than accuracy, the authors derive closed-form returns for smart money and for a one-to-one immediate copier under the pump.fun bonding curve, where each trade moves the price along a strictly increasing, convex path.

Why This Matters

Impact on research: The paper is positioned as the first multi-agent system for copy trading and one of the few defensive rather than descriptive treatments of meme coin manipulation. Prior work on crypto manipulation largely measured the phenomenon (pump-and-dump, rug pulls, wash trading, metric distortion); this work builds a defense, and it argues that trader profitability is predictable despite adversarial conditions.

Real-world applications:

• Copy-trading platforms such as GMGN could use bot filters and profitability predictions to flag or down-rank manipulative wallets before users replicate them.
• Retail traders could use the bot indicators (bundle, sniper, bump, comment) and candlestick/comment checks as pre-trade screens on new meme coins.
• Regulators and market-integrity researchers could use the detection algorithms and quantified bot prevalence as evidence when assessing manipulation on launch platforms like pump.fun.
• Wallet analytics and risk-scoring products could turn the feature set into real-time wallet reputation scores.

Industry relevance: Because Solana lacks a public mempool and has roughly 400 ms block times, the ability to front-run is concentrated among infrastructure-level participants, so ordinary copiers cannot out-speed bots and must instead out-reason them. The finding that position manipulation bots shift value between participants rather than destroy it, and that attention bots correlate with higher returns, gives platforms a concrete trade-off to manage: suppressing some bots protects users, while suppressing others removes signals.

Future Directions

• Detecting gradual bundles. The authors explicitly note that gradual bundles are much harder to detect accurately and that current temporal signals fail; a stronger detector is the obvious next step.
• Improving on the accuracy ceiling. XGBoost still achieved the highest AUC overall, ahead of the chain-of-thought multi-agent system, so the authors' own framing leaves room to close that gap.
• Narrowing the imitation penalty. The 14% smart-money return versus 3% copier return gap is proven to be strictly positive under the bonding curve, so a natural question is whether execution timing or partial replication can reduce it (the paper's economic-performance figure is truncated in the material provided, so the full breakdown is not reported there).
• Robustness to adaptation. The detection rules rest on current bot heuristics; as adversaries evolve beyond bundle, sniper, bump, and comment behaviors, both the detectors and the agents' chain-of-thought conditions would need updating.

Target Audience

Blockchain and DeFi security researchers, AI researchers working on multi-agent LLM systems and chain-of-thought reasoning, financial market microstructure researchers studying manipulation and copy trading, and practitioners building copy-trading, wallet-analytics, or meme coin risk products. Readers need comfort with machine learning evaluation and basic on-chain market mechanics; the formal return modeling and agent pipeline will be most accessible to advanced readers.

Note on completeness: the provided content is truncated at Figure 8 ("Test set average n..."), so the detailed economic-performance figure and the appendices referenced in the text (A.1–A.6, including the bonding-curve derivation and prompt details) could not be summarized.

Authors’ abstract

Copy trading has become the dominant entry strategy in meme coin markets. However, due to the market's extremely illiquid and volatile nature, the strategy exposes an exploitable attack surface: adversaries deploy manipulative bots to front-run trades, conceal positions, and fabricate sentiment, systematically extracting value from naïve copiers at scale. Despite its prevalence, bot-driven manipulation remains largely unexplored, and no robust defensive framework exists. We propose a manipulation-resistant copy-trading system based on a multi-agent architecture powered by a multi-modal large language model (LLM) and chain-of-thought (CoT) reasoning. Our approach outperforms zero-shot and most statistic-driven baselines in prediction accuracy as well as all baselines in economic performance, achieving an average copier return of 3% per meme coin investment under realistic market frictions. Overall, our results demonstrate the effectiveness of agent-based defenses and predictability of trader profitability in adversarial meme coin markets, providing a practical foundation for robust copy trading.