Research map / Multi-agent systems and games / LLM agents
Social Agents: Collective Intelligence Improves LLM Predictions
Aanisha Bhattacharyya, Abhilekh Borah, Yaman Kumar Singla, Rajiv Ratn Shah, Changyou Chen, Balaji Krishnamurthy
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Social Agents, a multi-agent LLM framework that mimics collective human decision making to improve predictions, based only on the abstract's motivation.
Problem
LLM predictions may be improved by harnessing the benefits of collective intelligence (diverse, independent, distributed judgments) that have historically outperformed individual experts.
Approach
Uses a society of LLM agents with diverse perspectives and independent judgments whose aggregated outputs are expected to outperform individual predictions, in the spirit of wisdom-of-crowds averaging.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| aggregate persona-conditioned agents instead of a single no-persona expert | GPT-4o · Tweet engagement | accuracy (%): 70.27 -> 86.90 | Improved |
| aggregate personas instead of single prompt | GPT-4o · Webpage likability | Pearson r: 0.28 -> 0.74 | Improved |
| aggregate personas | GPT-4o · Ad CTR (Creative) | MAPE: 72.45 -> 47.60 | Improved |
| aggregate personas | Qwen3 32B · Ad CTR (Real Estate) | 10-way accuracy (%): 26.10 -> 23.40 | Hurt |
| aggregate personas | LLaMA 3.2 90B Vision · Webpage likability | accuracy (%): 57.73 -> 56.70 | Hurt |
| aggregate personas | LLaMA 3.3 70B · ROAS (Creative) | PE@20: 0.31 -> 0.24 | Hurt |
| aggregate personas | GPT-4o · Video memorability | Spearman rho: 0.33 -> 0.41 | Improved |
| aggregate personas | Qwen3 32B · Video memorability | Spearman rho: 0.33 -> 0.33 | No clear effect |
| aggregate personas | GPT-4o · Tweet generation | BLEU-4: 23.77 -> 31.71 | Improved |
| aggregate personas | GPT-4o · Behavioral attribute Persuasion | accuracy (%): 61.53 -> 61.81 | No clear effect |
| aggregate personas vs repeated no-persona calls or multi-LLM ensemble | Social Agents · Ad CTR, webpage likability | MAPE / alignment | Improved |
| increase number of personas beyond 10-20 | Social Agents · Ad CTR | MAPE | No clear effect |
Closest papers by meaning
- TwinMarket: A Scalable Behavioral and Social Simulation for Financial Markets (NeurIPS 2025 · similarity 0.88)
- MF-LLM: Simulating Population Decision Dynamics via a Mean-Field Large Language Model Framework (NeurIPS 2025 · similarity 0.87)
- Steering the Herd: A Framework for LLM-based Control of Social Learning (ICLR 2026 · similarity 0.87)
- Exploring Prosocial Irrationality for LLM Agents: A Social Cognition View (ICLR 2025 · similarity 0.87)
- AI-Mediated Communication Can Steer Collective Opinion (NeurIPS 2026 · similarity 0.87)
- LLM Agents Can Be Choice-Supportive Biased Evaluators: An Empirical Study (AAAI 2025 · similarity 0.87)
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.
