Research map / Multi-agent systems and games / Games and equilibria
CSP4SDG: Constraint and Information-Theory Based Role Identification in Social Deduction Games with LLM-Enhanced Inference
Kaijie Xu, Fandi Meng, Clark Verbrugge, Simon Mark Lucas
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 |
|---|---|---|---|
| add assertions as high-weight soft constraints to Strict (ablation) | CSP4SDG · Avalon-NLU objective | marginal accuracy: 0.28 -> 0.33 | Improved |
| add IG-weighted hypotheses on top of assertions (ablation) | CSP4SDG · Avalon-NLU objective | marginal accuracy: 0.33 -> 0.34 | Improved |
| IG hypothesis weights instead of manual weights (ablation) | CSP4SDG · Avalon-NLU objective | marginal accuracy: 0.34 vs 0.33 | No clear effect |
| add soft constraints (hypotheses) to Strict (ablation) | CSP4SDG · Mafia-UCB objective | marginal accuracy: marginal 0.70 -> 0.73; MAP 0.90 -> 0.76 | Mixed |
| scaling all hypothesis weights by 0.25 to 4 (ablation) | CSP4SDG · Avalon-NLU / Mafia | MAP accuracy: MAP change <= 1.5 pp | No clear effect |
| add CSP posteriors to the LLM prompt (ablation) | LLM (GPT-4o-mini) · Avalon-NLU objective | marginal accuracy: 0.14 -> 0.19 | Improved |
| add CSP posteriors to the LLM prompt (ablation) | LLM (GPT-4o-mini) · Mafia-UCB objective | marginal accuracy: GChat 0.56 -> 0.69; TChat 0.66 -> 0.69 | Mixed |
| use turn-based chat instead of global chat | LLM · Mafia-UCB objective | marginal accuracy: 0.56 -> 0.66 | Improved |
| use a stronger LLM backbone (GPT-4o-mini to GPT-4.1 / DeepSeek-v3) (ablation) | LLM-only · Avalon-NLU objective | marginal accuracy: 0.14 -> 0.15 / 0.16 | No clear effect |
| use CSP+HypIG instead of LLM-only on AvalonLogs 9-player | CSP vs LLM · AvalonLogs 9-player objective | marginal accuracy: 0.280 -> 0.360 | Improved |
| feed CSP posteriors to LLM on AvalonLogs (ablation) | LLM+CSP · AvalonLogs 6-player objective | marginal accuracy: 0.289 vs CSP Strict 0.290 | No clear effect |
Closest papers by meaning
- BALROG: Benchmarking Agentic LLM and VLM Reasoning On Games (ICLR 2025 · similarity 0.87)
- Modeling Boundedly Rational Agents with Latent Inference Budgets (ICLR 2024 · similarity 0.86)
- Theory of Mind and Persuasion Beyond Conversation: Assessing the Capacity of LLMs to Induce Belief States via Planning and Action (NeurIPS 2026 · similarity 0.86)
- LLM Strategic Reasoning: Agentic Study through Behavioral Game Theory (NeurIPS 2025 · similarity 0.85)
- LieCraft: A Multi-Agent Framework for Evaluating Deceptive Capabilities in Language Models (AAAI 2026 · similarity 0.85)
- Sparks of Cooperative Reasoning: LLMs as Strategic Hanabi Agents (ICML 2026 · 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.
