Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

StagePilot: A Deep Reinforcement Learning Agent for Stage-Controlled Cybergrooming Simulation

Heajun An, Qi Zhang, Minqian Liu, Xinyi Zhang, Sang Won Lee, Lifu Huang, Pamela J. Wisniewski, Jin-Hee Cho

TL;DR

StagePilot introduces an offline deep reinforcement learning agent that controls stage transitions in cybergrooming simulations, treating dialogue as a long-horizon planning problem over six grooming stages. By combining adjacent-stage transition constraints with a sentiment-distance reward, StagePilot achieves coherent, emotionally engaging, and strategically progressive conversations, validated in LLM-driven predator–victim simulations. IQL+AWAC emerge as the strongest offline-RL baseline, delivering the highest Stage 6 reach and efficient dialogue length while maintaining sentiment alignment, and outperforming BC, CQL, and prompting baselines. The work provides a rigorous evaluation framework, expert validation, and demonstrates the potential of stage-level planning for safety-critical educational tools, with careful attention to ethical safeguards and applicability to broader counseling or negotiation simulations.

Abstract

Cybergrooming is an evolving threat to youth, necessitating proactive educational interventions. We propose StagePilot, an offline RL-based dialogue agent that simulates the stage-wise progression of grooming behaviors for prevention training. StagePilot selects conversational stages using a composite reward that balances user sentiment and goal proximity, with transitions constrained to adjacent stages for realism and interpretability. We evaluate StagePilot through LLM-based simulations, measuring stage completion, dialogue efficiency, and emotional engagement. Results show that StagePilot generates realistic and coherent conversations aligned with grooming dynamics. Among tested methods, the IQL+AWAC agent achieves the best balance between strategic planning and emotional coherence, reaching the final stage up to 43% more frequently than baselines while maintaining over 70% sentiment alignment.

StagePilot: A Deep Reinforcement Learning Agent for Stage-Controlled Cybergrooming Simulation

TL;DR

StagePilot introduces an offline deep reinforcement learning agent that controls stage transitions in cybergrooming simulations, treating dialogue as a long-horizon planning problem over six grooming stages. By combining adjacent-stage transition constraints with a sentiment-distance reward, StagePilot achieves coherent, emotionally engaging, and strategically progressive conversations, validated in LLM-driven predator–victim simulations. IQL+AWAC emerge as the strongest offline-RL baseline, delivering the highest Stage 6 reach and efficient dialogue length while maintaining sentiment alignment, and outperforming BC, CQL, and prompting baselines. The work provides a rigorous evaluation framework, expert validation, and demonstrates the potential of stage-level planning for safety-critical educational tools, with careful attention to ethical safeguards and applicability to broader counseling or negotiation simulations.

Abstract

Cybergrooming is an evolving threat to youth, necessitating proactive educational interventions. We propose StagePilot, an offline RL-based dialogue agent that simulates the stage-wise progression of grooming behaviors for prevention training. StagePilot selects conversational stages using a composite reward that balances user sentiment and goal proximity, with transitions constrained to adjacent stages for realism and interpretability. We evaluate StagePilot through LLM-based simulations, measuring stage completion, dialogue efficiency, and emotional engagement. Results show that StagePilot generates realistic and coherent conversations aligned with grooming dynamics. Among tested methods, the IQL+AWAC agent achieves the best balance between strategic planning and emotional coherence, reaching the final stage up to 43% more frequently than baselines while maintaining over 70% sentiment alignment.
Paper Structure (70 sections, 8 equations, 1 figure, 31 tables)

This paper contains 70 sections, 8 equations, 1 figure, 31 tables.

Table of Contents

  1. Introduction
  2. Why Create a Predator Chatbot?
  3. Unresolved Challenges.
  4. Addressing Gaps.
  5. Related Work
  6. Reinforcement Learning (RL) for Controlled Dialogue Generation.
  7. Chatbot for Cyber Safety.
  8. Proposed Approach: StagePilot
  9. DRL-Based Cybergrooming Simulator
  10. State Representation.
  11. Action Space.
  12. Reward Function.
  13. Training Procedure.
  14. Adaptation to StagePilot
  15. Experimental Setup
  16. ...and 55 more sections

Figures (1)

  • Figure 1: StagePilot overview. The RL policy selects stage transitions, while LLMs generate dialogue as environment dynamics.