Scalable Evaluation of Online Facilitation Strategies via Synthetic Simulation of Discussions (2503.16505v2)

Abstract: Limited large-scale evaluations exist for facilitation strategies of online discussions due to significant costs associated with human involvement. An effective solution is synthetic discussion simulations using LLMs to create initial pilot experiments. We propose a simple, generalizable, LLM-driven methodology to prototype the development of LLM facilitators, and produce high-quality synthetic data without human involvement. We use our methodology to test whether current facilitation strategies can improve the performance of LLM facilitators. We find that, while LLM facilitators significantly improve synthetic discussions, there is no evidence that the application of more elaborate facilitation strategies proposed in modern Social Science research lead to further improvements in discussion quality, compared to more basic approaches. Additionally, we find that small LLMs (such as Mistral Nemo 12B) can perform comparably to larger models (such as LLaMa 70B), and that special instructions must be used for instruction-tuned models to induce toxicity in synthetic discussions. We confirm that each component of our methodology contributes substantially to high quality data via an ablation study. We release an open-source framework, "SynDisco" (pip install syndisco), which implements our methodology. We also release the "Virtual Moderation Dataset" (https://paperswithcode.com/dataset/vmd), a large, publicly available dataset containing LLM-generated and LLM-annotated discussions using multiple open-source LLMs.

• The paper introduces SynDisco, a framework utilizing synthetic discussions by LLMs to evaluate six online moderation strategies.
• The paper demonstrates that an RL-inspired 'Moderation Game' strategy improves discussion quality by reducing toxicity more effectively than traditional guidelines.
• The paper finds that using smaller LLMs produces more diverse discussions compared to larger, heavily aligned models, highlighting crucial trade-offs.

Scalable Evaluation of Online Facilitation Strategies via Synthetic Simulation of Discussions

Introduction

The research presents a methodology using synthetic simulations by LLMs for evaluating online moderation strategies, aiming to address the challenges of involving human participants in moderation experiments. The paper evaluates six different moderation configurations, offering a comprehensive evaluation of synthetic LLM-driven discussions as a viable alternative to human involvement in moderation research. The introduced approach includes the development of a framework named "SynDisco" for simulating discussions, accompanied by the release of the Virtual Moderation Dataset (VMD).

Methodology

The methodology relies on synthetic experiments conducted by LLMs to evaluate different moderation strategies. Synthetic discussions are defined mathematically, utilizing sets of users and comments to simulate discussions with or without moderation. These comments are generated recursively by LLM user-agents, employing specific instructions varied by role and strategy. Six moderation strategies, including both real-life inspired and experimental approaches, are tested to determine their effectiveness (Figure 1).

Figure 1: Effects of moderation strategy for toxicity and argument quality. Error bars represent the 95% confidence interval.

The evaluation employs a statistical framework to compare moderation strategies, fundamentally relying on comment-level measures of toxicity and argument quality. Measures are annotated by LLMs to assess the effectiveness of each strategy. SynDisco is designed to efficiently manage synthetic discussions, annotation, and randomization tasks, ensuring reliable experimentation over hundreds of synthetic discussions.

Results

The experiments reveal that the proposed RL-inspired "Moderation Game" strategy notably improves synthetic discussions, outperforming other moderation configurations, including those derived from real-life guidelines (Figure 2). Surprisingly, established moderation guidelines do not significantly outperform basic or out-of-the-box baseline strategies, particularly in terms of reducing toxicity.

Figure 2: Effects of LLM user-intent for toxicity and argument quality. Error bars represent the 95% confidence interval.

LLM moderators demonstrate a tendency toward over-intervention, which is unrepresentative of human moderators and may require addressing in future fine-tuning sessions. Results also highlight the negative impact of LLM user-agents with trolling intentions on the quality and toxicity of discussions.

In examining model diversity, smaller LLMs, such as Qwen-2.5, result in more varied discussions compared to larger LLMs like LLaMa-3.1, which tend toward stereotypically polite behavior and less diverse linguistic expressions (Figure 3). This suggests that intense alignment procedures in larger models may limit variability in synthetic discussions.

Figure 3: Histogram of the length of comments (number of words) produced by various LLMs.

Technical Implementation

The technical implementation utilizes three open-source LLMs with varying parameters and specifications for generating and annotating discussions. Different models are employed for user-agents and annotation, with the LLaMa-3.1 model being specifically used for synthetic annotation due to its reliable toxicity assessment.

Virtual Moderation Dataset

The resulting VMD dataset consists of a broad array of synthetic discussions and annotations, providing extensive data for further research. This dataset can be utilized to explore the behaviors of different LLMs under various moderation strategies and to further analyze the efficacy of synthetic simulations of online moderation.

Conclusion

The research successfully demonstrates the potential for using LLMs to simulate online discussions with different moderation strategies, significantly reducing the need for human involvement in moderation experiments. It highlights the promising performance of LLMs, particularly in moderation roles, while pointing out areas where further refinement and understanding are necessary.

Future work involves exploring correlations between synthetic simulations and real-world moderation outcomes to understand synthetically replicated discussion dynamics and their potential application in training human moderators. Additionally, more advanced moderator models, possibly including reinforcement learning steps or RAG-based models, could be developed and compared with baseline models to gain deeper quantitative insights.