Game of Phones: Harnessing Game Theory and LLMs in 'Bot Wars' to Counteract Phone Scams

Nardine Basta, Robin Carpentier, Benjamin Zi Hao Zhao, Dali KaafararXiv 2026 · DOI · arXiv

This paper introduces a game-theoretic framework that transforms Large Language Models (LLMs) into adaptive "scam-baiters" designed to waste scammers' time and resources. By modeling interactions as a two-player game, the authors demonstrate that LLMs can discover and refine optimal counter-strategies through in-context experience alone, achieving significantly better performance than traditional static, template-based systems.

TL;DR

Researchers have developed a game-theoretic "Bot War" framework where LLMs learn to outmaneuver phone scammers by maximizing interaction time while minimizing data disclosure. Using in-context learning, these agents achieve up to 79.7% closer alignment to a theoretical Nash equilibrium than non-adaptive baselines, offering a scalable, low-cost way to impose real resource costs on scam operations.

What Problem Does This Paper Address?

Phone scams are a multi-billion dollar global industry. Current defenses—like call blocking or blacklisting—are purely reactive and impose zero cost on the perpetrator; a blocked scammer simply moves to the next number.

"Scam-baiting" (engaging scammers to waste their time) is a proactive alternative, but it’s difficult to scale. Human baiters are limited, and previous automated attempts relied on static scripts that were easily detected by adapting scammers. The challenge lies in creating an agent that can dynamically adapt its strategy to maintain credibility while dragging out the conversation as long as possible.

Methods

The authors treat the scammer-baiter interaction as a two-player non-cooperative game.

  1. Strategy Sets: Both players choose from 10 empirically grounded strategies (e.g., Scammers use Urgency or Authority; Baiters use Delay or Counter Questioning).
  2. Utility Function: A zero-sum payoff structure was designed where the scammer gains utility by extracting Financial Personally Identifiable Information (FPII) and loses utility as time passes. The baiter’s goal is the exact inverse.
  3. In-Context Learning (ICL): Instead of expensive fine-tuning, the agents use ICL. The "Baiter Strategy Selector" receives the current conversation history plus performance data from the 20 most recent games to sample its next move.
  4. Nash Equilibrium (NE) Benchmarking: The researchers computed the NE from empirical payoff matrices to serve as a formal "gold standard" for optimal play.
Fig. 1: System architecture and information flow during a turn. Data from the current game and 20 most recent games are processed to select a strategy. Mes- sage Generators translate selected strategies into contextually appropriate di- alogue. The Dialogue Analyzer extracts PII, computes utilities, and stores the data. Dashed arrows indicate feedback loops enabling learning across games.

Datasets

The study produced and released a new annotated dataset:

  • 300 Simulated Dialogues: 100 games each for GPT-4, Mixtral-8x22B, and DeepSeek-Chat.
  • 10,247 Annotated Turns: Each turn includes strategy labels, extracted PII states, and cumulative utility scores.
  • Validation: Manual validation confirmed a 94% accuracy rate for PII extraction across the models.

Key Insights and Ideas

  • Inductive Bias Trumps Parameter Count: Surprisingly, the 16B-parameter DeepSeek-Chat outperformed the much larger 104B-parameter Mixtral. This suggests that a model's internal "inductive bias" for strategic reasoning is more important for this task than raw scale.
  • Universal vs. Architecture-Specific Tactics: "Delay" and "Counter Questioning" emerged as universally effective strategies across all models. However, the "optimal mix" of these strategies varied significantly between LLM architectures.
  • The "Reassurance" Trap: Error analysis showed that scammers using the Reassurance strategy (S7) were most successful at "breaking" the baiters, leading to faster credential disclosure.
  • Emergent Strategy Discovery: LLMs were able to identify high-priority strategies and calibrate their usage frequencies over time without any parameter updates, proving that game-theoretic feedback can activate latent strategic reasoning in transformers.
Table 1: Overview of strategies. Full descriptions are provided in App. B

Key Results

  • Equilibrium Alignment: Adaptive strategy selection achieved 28.9–79.7% closer alignment to the Nash equilibrium compared to random or non-adaptive baselines.
  • Performance: DeepSeek-Chat was the "MVP," achieving a 41% scammer win rate (the lowest among tested models) and the longest average interaction duration at 42 turns.
  • Economic Viability: At current API pricing, a DeepSeek-based baiter costs roughly $0.10–$0.30 per interaction. Given that the average loss per successful phone scam is orders of magnitude higher, the "cost asymmetry" heavily favors the defenders.
  • Duration Boost: The adaptive framework increased average interaction duration by up to 83% compared to previous static prompting methods.

Limitations and Future Work

  • Zero-Sum Simplification: The model assumes a zero-sum game, whereas real-world scenarios where scammers manage multiple targets simultaneously might require a more complex "general-sum" formulation.
  • Simulated Environment: The study used LLM-vs-LLM simulations. While this allows for controlled learning analysis, the authors acknowledge that validation against real human scammers is the necessary next step for "ecological validity."
  • Utility Function Sensitivity: The Nash equilibrium predictions are tied to the specific weights in the hand-crafted utility function; different parameterizations might yield different "optimal" behaviors.
  • Future Directions: The authors suggest integrating these bots into live carrier-level detection pipelines and moving toward "extensive-form" game representations for even more granular strategic analysis.

Summary updated

Abstract

Phone scams impose substantial and growing costs on individuals and financial systems worldwide, yet existing countermeasures impose no resource cost on perpetrators. Strategic scam-baiting, in which automated agents engage scammers to deplete their operational capacity, offers a proactive alternative but requires dynamic strategy adaptation that static systems cannot provide. We introduce a game-theoretic framework enabling LLMs to learn effective counter-strategies through in-context experience without fine-tuning. The framework models scammer-baiter interactions as a two-player game with empirically grounded strategies and utility functions, using Nash equilibrium as a formal benchmark for evaluating learned behaviour. Across 300 simulated dialogues and three LLM architectures, adaptive strategy selection achieves 28.9-79.7% closer alignment to Nash equilibrium than non-adaptive baselines. The code and an annotated dataset of 300 dialogues with turn-level strategy labels and utility scores are released at https://github.com/nardinebasta/GameofPhones_/ .
BibTeX
@misc{basta2026game,
  author = {Nardine Basta and Robin Carpentier and Benjamin Zi Hao Zhao and Dali Kaafar},
  title = {{Game of Phones: Harnessing Game Theory and LLMs in 'Bot Wars' to Counteract Phone Scams}},
  year = {2026},
  eprint = {2610.26781},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2610.26781},
}
Figure