Research map / Large language models / AI and society
Evaluating Online Moderation via LLM-Powered Counterfactual Simulations
Giacomo Fidone, Lucia Passaro, Riccardo Guidotti
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use personalized neutral moderation message (PMI-N) instead of one-size-fits-all message (ablation) | COSMOS ex ante moderation · COSMOS full-population simulations | mass divergence of toxicity: OSFA -0.06, +0.04, -0.05, -0.06, +0.05 vs PMI-N -0.09, +0.00, -0.08, -0.11, -0.11 | Improved |
| make personalized moderation prescriptive (PMI-P) instead of neutral (PMI-N) (ablation) | COSMOS PMI · COSMOS full-population simulation 5 | mass divergence of toxicity: -0.11 -> -0.02 | Hurt |
| make personalized moderation empathizing (PMI-E) instead of neutral (PMI-N) (ablation) | COSMOS PMI · COSMOS full-population simulation 4 | mass divergence of toxicity: -0.11 -> -0.02 | Hurt |
| lower ban tolerance from e=8 to e=1 (ablation) | COSMOS BAN · COSMOS full-population simulation 1 | mass divergence of toxicity: -0.06 -> -0.54 | Improved |
| lower ban tolerance from e=8 to e=1 (ablation) | COSMOS BAN · COSMOS full-population simulations | content loss ratio: CLR 0.04 -> 0.45 | Hurt |
| lower ban tolerance to e=1 (healthy content lost) (ablation) | COSMOS BAN · COSMOS full-population simulations | fraction of healthy content lost: 0.27 ± 0.03 of content below THR lost; macro-average recall 0.55 at e=1, 0.60 at e=2, 0.58 at e=4 | Hurt |
| use cal_tox prompt template instead of no_tox (ablation) | COSMOS user prompt · PANDORA toxicity distribution | KL divergence: 1.37 -> 0.07 | Improved |
| use cal_tox prompt template instead of yes_tox (ablation) | COSMOS user prompt · PANDORA toxicity distribution | KL divergence: 0.57 -> 0.07 | Improved |
| use cal_tox instead of no_tox template (ablation) | COSMOS user prompt · candidate template evaluation | percent toxic outputs: 0.03 -> 8.12 | Improved |
| use cal_tox instead of no_tox template (ablation) | COSMOS user prompt · candidate template evaluation | percent correctly formatted outputs: 84.95 -> 71.66 | Hurt |
| use SOLAR-10B instead of Llama-2-13B (ablation) | COSMOS LLM · PANDORA sample | perplexity: 53.63 -> 22.50 | Improved |
| restrict simulation to sub-population of 5 most toxic agents plus least toxic agent | COSMOS · sub-population simulation | agent median toxicity: parent-child toxicity correlation rho=+0.39 | Hurt |
Closest papers by meaning
- Position: LLM Social Simulations Are a Promising Research Method (ICML 2025 · similarity 0.86)
- OdysSim: Building Foundation Models for Human Behavior Simulation (NeurIPS 2026 · similarity 0.86)
- Measuring And Improving Persuasiveness Of Large Language Models (ICLR 2025 · similarity 0.86)
- Towards Real-world Human Behavior Simulation: Benchmarking Large Language Models on Long-horizon, Cross-scenario, Heterogeneous Behavior Traces (NeurIPS 2026 · similarity 0.86)
- DEBATE: A Large-Scale Benchmark for Evaluating Opinion Dynamics in Role-Playing LLM Agents (NeurIPS 2026 · similarity 0.86)
- MF-LLM: Simulating Population Decision Dynamics via a Mean-Field Large Language Model Framework (NeurIPS 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
