CuriousTogether: Implications of Curiosity and Cultural Grounding in Multi-Agent LLM Interactions
Organizations: University of Michigan – Ann Arbor, USA · Jinesis Lab, University of Toronto & Vector Institute · MPI for Intelligent Systems, Tübingen, Germany · EuroSafeAI
Abstract
Large Language Models (LLMs) are increasingly deployed in multi-agent settings, exhibiting emergent collective behaviors, yet the role of curiosity as a driver of such dynamics remains understudied. We introduce CuriousTogether, a framework evaluating curiosity-driven multi-agent LLM interactions across three complementary dimensions: novelty-seeking, epistemic and social curiosity. Across eight tasks spanning reasoning, creative generation, and social bias mitigation, curiosity-guided multi-agent settings yield improvements over non-curious ones on seven out of eight tasks, with gains of up to 20.6% in story generation (semantic diversity) and a 14.9 pp reduction in stereotype rates. Furthermore, country-conditioned cultural grounding show benefits for story generation and social tasks, but can hurt precision-oriented reasoning. Finally, we show that inference-time multi-agent curiosity-based approach offers favorable quality-token trade-offs, compared with single-agent baselines. Together, these findings support curiosity-guided interaction as a practical mechanism for improving multi-agent collaboration, with downstream benefits for reasoning, creative, and societal tasks.
Figures & tables
| Task | Input | Curious agent questions (examples) | Output |
| Novelty-Seeking | |||
| Story gen. | Title: “The Last Lighthouse Keeper” | “What twist could we introduce that surprises the reader?” | Novel story |
| Hypo. gen. | LLMs hallucinate due to training noise. | “Do you know if the problem lies in retrieval or generation?” | Creative hypothesis |
| Codenames | Clue word: bank . Choose from: robin, river, mud, … | “Do we anchoring on the same meaning of this word?” | {robin, river, piggy, blood} ✓ |
| Epistemic | |||
| CSQA | Where do people go to borrow a book to read at home? Choices: library, bookstore.. | “Where would find a book to also take home?” | library ✓ |
| Novelty-Seeking | Epistemic | Social | |||||||||||
| Story Gen. | Hypothesis Gen. | Codenames | CSQA | SQA | Nov.-Conc. | 1-Ster. | Ethics | ||||||
| Nov. | SD | LR | PPL | IA | ID | BLEU | RL | Acc | Acc | Acc | Acc | Acc | |
| Curiosity | .647 0 | .620 0 | .787 0 | 185.1 11.8 | .595 0 | .840 0 | .062 0 | .261 0 | 73.77 2.7 | 69.56 2.8 | 34.33 2.9 | 75.36 3.8 | 49.39 3.1 |
| No-Curiosity | .600 0 | .514 0 | .729 0 | 174.2 12.6 | .542 0 | .790 0 | .046 0 | .177 0 | 72.81 2.8 | 70.33 2.8 | 30.21 2.9 | 60.46 4.3 | 48.97 3.1 |
| Culture | .647 0 | .602 0 | .772 0 | 175.4 11.9 | .578 0 | .830 0 | .060 0 | .253 0 | 73.41 2.7 | 68.37 2.9 | 32.12 2.9 | 76.78 3.8 | 49.39 3.1 |
| No-Culture | .588 0 | .579 0 | .774 0 | 164.9 11.2 | .564 0 | .832 0 | .060 0 | .243 0 | 73.77 2.7 | 70.33 2.8 | 34.33 2.9 | 74.68 3.8 | 48.97 3.1 |
Appendix figures & tables20 assets
Supplementary material from the paper’s appendix.
Appendix
| Dimension | Dataset | # Examples |
| Epistemic | CommonSenseQA (CSQA) | 1221 |
| StrategyQA (SQA) | 1603 | |
| Novelty Concepts (BigBench) | 32 | |
| Novelty-Seeking | Story Generation (1K Stories) | 1000 |
| Hypothesis Generation (HypoGen) | 5500 | |
| Codenames (BigBench) | 90 |
| Task | Effect | Holm |
| Hypothesis | ||
| Codenames | ||
| Stereotypes | pp | |
| Story | ||
| Novel Concepts | pp | |
| CSQA | pp |
| Qwen-2.5-14B | LLaMA-3-8B | Mistral-7B | DeepSeek-7B-Chat | Gemma-3-12b | OLMo-7B | |||||||||||||
| Story Generation Nov. / Sem. Diversity / Lexical Richness | ||||||||||||||||||
| Nov. | SD | LR | Nov. | SD | LR | Nov. | SD | LR | Nov. | SD | LR | Nov. | SD | LR | Nov. | SD | LR | |
| Curiosity | .600 | .730 | .830 | .717 | .668 | .778 | .655 | .605 | .767 | .622 | .482 | .753 | .635 | .683 | .839 | .654 | .550 | .755 |
| No-Curiosity | .580 | .520 | .760 | .591 | .610 | .730 | .620 | .531 | .644 | .605 | .429 | .733 | .601 | .588 | .804 | .606 | .405 | .703 |
| Culture | .600 | .730 | .820 | .717 | .668 | .730 | .655 | .602 | .761 | .622 | .432 | .749 | .635 | .683 | .830 | .654 | .499 | .745 |
| No-Culture | .550 | .570 | .830 | .591 | .610 | .730 | .563 | .605 | .738 | .598 | .482 | .753 | .615 | .656 | .839 | .613 | .550 | .755 |
| Llama-3-8B | Qwen-2.5-14B | Mistral-7B | DeepSeek-7B | Gemma-3-12B | OLMo-7B | |
| CSQA — Accuracy (%) | ||||||
| Curiosity | 78.21 | 83.70 | 72.89 | 67.57 | 79.52 | 66.99 |
| No-Curiosity | 76.99 | 84.28 | 71.01 | 65.93 | 78.87 | 60.70 |
| Culture | 76.41 | 83.13 | 71.58 | 64.19 | 79.03 | 66.99 |
| No-Culture | 78.21 | 84.28 | 72.89 | 67.57 | 79.52 | 61.39 |
| Single-Agent | 74.86 | 83.05 | 67.98 | 58.89 | 79.93 | 59.51 |
| Qwen-2.5-14B | Llama-3-8B | Mistral-7B | DeepSeek-7B | Gemma-3-12B | OLMo-7B | |
| Stereotype Rate | ||||||
| Curiosity | 24.52 | 31.33 | 29.56 | 12.75 | 35.76 | 5.42 |
| No-Curiosity | 40.62 | 62.03 | 40.84 | 17.82 | 65.78 | 9.69 |
| Culture | 24.52 | 31.33 | 29.56 | 12.75 | 43.05 | 5.42 |
| No-Culture | 29.80 | 34.82 | 29.60 | 12.75 | 35.76 | 10.87 |
| Single-Agent | 65.77 | 55.45 | 52.25 | 40.00 | 55.86 | 61.29 |
| Setting | Llama-3 8B | Qwen-2.5 14B |
| Single-agent | 55.45 | 65.77 |
| Self-refine | 40.33 | 36.04 |
| Non-interacting ensemble | 49.55 | 32.38 |
| Interacting multi-agent | 33.08 | 27.16 |
| Task | Condition | Quality | Coherent |
| Story | Curiosity | 3.35 | 85.0% |
| Story | No curiosity | 2.93 | 87.5% |
| Hypothesis | Curiosity | 3.60 | 90.0% |
| Hypothesis | No curiosity | 2.93 | 70.0% |
| Pooled | Curiosity | 3.48 | 87.5% |
| Pooled | No curiosity | 2.93 | 78.75% |
| Task | Setting | Outputs | Relevance | Coherence |
| Hypothesis | Multi-agent, all-curious | 710 | 3.74 | 3.86 |
| Hypothesis | Multi-agent, no-curiosity | 710 | 2.84 | 2.29 |
| Story | Multi-agent, all-curious | 300 | 3.98 | 2.98 |
| Story | Multi-agent, no-curiosity | 300 | 2.04 | 1.42 |
| Metric | Pearson | 95% CI | Spearman | |
| Relevance/quality | .755 | [.642, .836] | .760 | |
| Coherence | .726 | [.602, .816] | .738 |
| Task | Metric | Qwen-2.5-14B | Qwen-2.5-72B | LLaMA-3-8B | LLaMA-3.3-70B |
| Story Generation | Novelty | 0.583 | 0.564 | 0.654 | 0.564 |
| Semantic Diversity | 0.638 | 0.548 | 0.639 | 0.479 | |
| Lexical Richness | 0.810 | 0.806 | 0.742 | 0.797 | |
| Hypothesis Generation | Perplexity | 133.5 | 50.5 | 208.5 | 52.4 |
| IA Score | 0.791 | 0.630 | 0.380 | 0.606 | |
| Idea Distinctness | 0.832 | 0.672 | 0.850 | 0.683 |
| Task | Model | Curious | Control | |
| Bias | DeepSeek-7B | 60.9 | 59.8 | +1.1 |
| Bias | Mistral-7B | 59.3 | 53.3 | +6.0 |
| CSQA | DeepSeek-7B | 26.7 | 13.3 | +13.3 |
| CSQA | Mistral-7B | 66.7 | 66.0 | +6.7 |
| SQA | DeepSeek-7B | 43.3 | 43.3 | 0.0 |
| SQA | Mistral-7B | 80.0 | 70.0 | +10.0 |
| Continent | Countries |
| North America | United States, Canada, Mexico |
| South America | Brazil, Argentina, Chile |
| Europe | Germany, France, United Kingdom, Italy, Spain |
| Asia | India, China, Japan, South Korea |
| Africa | Nigeria, Kenya, South Africa, Egypt |
| Oceania | Australia, New Zealand |
| Task | Specific | Generic | NN Sim. | High-Sim |
| Story | 0.868 | 0.132 | 0.217 | 0.002 |
| Hypothesis | 0.280 | 0.720 | 0.791 | 0.719 |
| Codenames | 0.888 | 0.112 | 0.461 | 0.049 |
| CSQA | 0.791 | 0.209 | 0.498 | 0.177 |
| SQA | 0.835 | 0.165 | 0.352 | 0.032 |
| Novel Concepts | 0.781 | 0.219 | 0.468 | 0.099 |
| Task | Main curiosity prompt | Ablated curiosity prompt |
| Story Generation | You are a curious storyteller (optionally from a cultural background). The story so far: [story] Based on the story so far, please: (1) Add exactly one new sentence that continues the narrative. Keep it consistent with the flow. (2) Pose one thoughtful question to guide the next storyteller. In culture mode, weave in a subtle cultural element without overemphasizing it. Format: Sentence: [..] Question: [..] | You are an exploratory storyteller (optionally from a cultural background). The story so far: [story] Based on the story so far, extend the narrative by introducing an interesting uncertainty, possibility, or consequence. (1) Add exactly one new sentence that develops this exploratory direction. (2) Ask one forward-looking question that encourages the next storyteller to expand this possibility or uncertainty. Keep the flow natural and coherent. Format: Sentence: [..] Question: [..] |
| Hypothesis Generation | You are a research collaborator whose lens is curiosity-driven, integrative reasoning. Given: Title, BIT, FLIP. Task: (1) Propose one precise, testable hypothesis that reconciles or adjudicates BIT vs FLIP. (2) Provide a 1–2 sentence rationale referencing mechanisms or conditions. (3) Propose a minimal evaluation plan (data, method, controls). (4) List 2–4 quality metrics or statistical tests. (5) Ask exactly one thoughtful follow-up question that moves the collaboration forward. In culture mode, use cultural background only to broaden perspectives when helpful. Output strict JSON. | You are an exploratory research collaborator. Given: Title, BIT, FLIP. Task: (1) Propose one precise, testable hypothesis that identifies the key condition, mechanism, or moderator explaining when BIT versus FLIP holds. (2) Provide a 1–2 sentence rationale focusing on when and why each applies. (3) Propose a minimal evaluation plan. (4) List 2–4 quality metrics or tests. (5) Ask exactly one follow-up question aimed at identifying the most informative missing condition, assumption, or test. Output strict JSON. |
| Codenames | You are a curious collaborator. Propose guesses or clues while actively soliciting teammates’ interpretations and reasoning. Ask questions to better understand associations before deciding. | You are an exploratory collaborator. Probe alternative associations, ambiguities, and overlooked connections between clues and words. Actively explore multiple possible interpretations before deciding. |
| Scenario / Stereotype Reduction | You are a curious agent. Explain your reasoning about your task assignments. Engage with others’ reasoning and respond to their assignments. Ask at least one explicit question to other agents about why they made specific assignments. Remain open to revising your views to reach a well-justified consensus. | You are an exploratory agent. Explain your reasoning about your assignments. Identify at least one uncertainty, tension, or overlooked assumption in the current assignments. Ask at least one pointed question that could reveal missing evidence or a better alternative assignment. Build on or challenge others’ reasoning where appropriate. |
| Task | Main curiosity prompt | Ablated curiosity prompt |
| CSQA / StrategyQA | You are a curious reasoning agent. Reason step by step to determine the correct answer. Ask one question other agents whether they agree with your choice or if they see a better option. Review group responses and finalize the answer based on the most typical real-world reasoning. | Now, be curious : identify the single most uncertain part of your reasoning (the step you are least confident about). Ask one question to the group that would reduce that uncertainty and help you decide. |
| Novel Concepts | You are a curious agent generating ideas. Generate diverse, novel ideas while maintaining coherence and relevance to the task. Encourage variation and exploration across outputs. | You are an exploratory agent generating ideas. Focus on uncovering unexpected or underexplored conceptual combinations. Explicitly explore alternative directions, edge cases, or unconventional possibilities. |
| Task | Generic-question instruction used |
| Story | Ask one clarification or verification question that checks the current story plan; do not explore alternatives, uncertainty, evidence gaps, or assumptions. |
| Hypothesis | Ask one clarification or verification question that checks the current hypothesis; prohibit new perspectives, speculation, alternatives, uncertainty, evidence gaps, and overlooked assumptions. |
| CSQA / StrategyQA | Ask one clarification or verification question that checks the current reasoning, then recommend the best answer. Exploratory questioning is explicitly prohibited. |
| Stereotype Rate | Ask one verification question about the current reasoning. Questions about social roles, new perspectives, alternatives, and overlooked assumptions are prohibited. |
| ETHICS | Ask one verification question about the current reasoning, while prohibiting perspective exploration, alternatives, uncertainty, and assumptions. |
| Task | Metric | Curiosity Question | Generic Question | Difference |
| Story | Novelty | 0.647 | 0.579 | -0.068 |
| Story | Semantic Diversity | 0.620 | 0.538 | -0.082 |
| Story | Lexical Richness | 0.787 | 0.787 | 0.000 |
| Hypothesis | Perplexity | 174.2 | 72.66 | -101.54 |
| Hypothesis | IA Score | 0.595 | 0.590 | -0.005 |
| Hypothesis | Idea Distinctness | 0.840 | 0.656 | -0.184 |
Multi-Agent LLM Systems
Multi-agent large language model systems coordinate multiple AI agents to solve tasks through communication, specialization, shared tools, and memory.
See all 1,114 papers