ai4he/hidden-profile-llm-groups

Hidden Profile LLM Groups

Dataset accompanying the paper Hidden Profile Decision Making in Multi-Agent LLM Groups (Companion of the 2026 ACM Conference on Computer-Supported Cooperative Work and Social Computing (CSCW Companion '26)).

Session-level records for replications of the hidden-profile paradigm with three-agent LLM groups in a 2x2 design (synchronous vs asynchronous communication x shared vs hidden information). Includes the original run (240 sessions; votes of two models could not be parsed), three corrected runs with fixed vote elicitation (a re-vote for Gemini 2.5 Flash and GPT-OSS-120B, Kimi K2 and Qwen3-32B with a dedicated voting step, and a 180-session run with three Llama/Qwen models), the post-review length-only control re-run (90 sessions with full transcripts plus a 14-session anchor), the runner scripts containing the task materials, and the original-run figure.

Paper	Hidden Profile Decision Making in Multi-Agent LLM Groups
Venue	Companion of the 2026 ACM Conference on Computer-Supported Cooperative Work and Social Computing (CSCW Companion '26)
Authors	Carlos Toxtli, Manuel Delaflor
License	Creative Commons Attribution 4.0 International (cc-by-4.0); code: MIT License
Dataset DOI	10.5281/zenodo.24179025
Version	1.0.0
Maintainer	Human-AI Empowerment Lab, Clemson University

Paper abstract

The hidden profile paradigm is a foundational tool of CSCW research on computer-mediated group decision making: a group can identify the collectively optimal option only if its members surface their uniquely held information, and asynchronous communication is known to improve that pooling in human groups. As multi-agent LLM systems take on group decision-support roles, it is unclear whether the same channel-level interventions improve their performance. We replicate the hidden profile paradigm with groups of three LLM agents from four families (Kimi K2, GPT-OSS 120B, Gemini 2.5 Flash, Llama 4 Scout 17B) in a two-by-two design crossing communication mode (synchronous, asynchronous) with information distribution (shared, hidden). Across 219 valid sessions, only one model-by-mode cell exceeds the three-candidate chance baseline of 33 percent: Kimi K2 under asynchronous communication, at 87 percent correct (Fisher's exact, two-sided p equals 0.002), an early-stage result from 15 sessions that awaits larger-n replication; two other models share private information at rates above 75 percent but never aggregate it into the correct decision. In this task and these model families, information sharing and information integration appear to be dissociated capabilities, and we develop the implications for CSCW systems that scaffold AI-mediated distributed decision making.

Dataset structure

Path	Format	Size	Records
code/exp03_hidden_profile_cmc.py	py	20.4 KB
code/run_e3_cross_category_fixed.py	py	7.9 KB
code/run_e3_fixed_votes.py	py	19.7 KB
data/analysis/exp03_hidden_profile_cmc_figure.pdf	pdf	49.1 KB
data/corrected_runs/e3_cross_category_fixed.json	json	35.8 KB	120
data/corrected_runs/e3_fixed_votes.json	json	89.0 KB	180
data/corrected_runs/e3_revoted.json	json	31.5 KB	62
data/length_control_rerun/data_rerun.jsonl	jsonl	4.2 MB	90
data/length_control_rerun/data_rerun_groq_gptoss.jsonl	jsonl	222.7 KB	14
data/original_run/exp03_hidden_profile_cmc.json	json	93.5 KB	240

10 files, 4.8 MB in total (metadata files excluded). Records are rows for CSV/TSV/JSONL files and list length for JSON arrays.

Loading the data

from datasets import load_dataset

ds = load_dataset("ai4he/hidden-profile-llm-groups", "data_rerun")

Available subsets: data_rerun, data_rerun_groq_gptoss. All other files can be downloaded with huggingface_hub.snapshot_download or from the Zenodo record.

Known limitations

• This release contains the latest and corrected data, which do not match every number in the published poster: the original-run file uses Qwen3-32B where the poster reports Llama 4 Scout, several cell counts differ, and GPT-OSS/Gemini votes were unparsed in the original run (see data/corrected_runs/).
• The length-only control re-run matches the poster's control analysis exactly.
• No script for the poster's inferential statistics is included; the re-vote script is omitted because it embeds credentials.

Personal and sensitive information

None. No human participants.

License

The data and documentation are released under the Creative Commons Attribution 4.0 International. Source code files (.ipynb, .js, .m, .py, .sh) are released under the MIT License.

Citation

If you use this dataset, please cite the paper:

@inproceedings{Toxtli2026Hidden,
 title = {Hidden Profile Decision Making in Multi-Agent LLM Groups},
 author = {Carlos Toxtli and Manuel Delaflor},
 booktitle = {Companion of the Computer-Supported Cooperative Work and Social Computing (CSCW Companion ’26)},
 year = {2026},
 doi = {10.1145/3785651.3831525}
}

and the dataset:

@dataset{Toxtli2026HiddenProfileLlmGroupsData,
  author = {Toxtli, Carlos and Delaflor, Manuel},
  title = {{Hidden Profile LLM Groups}: Dataset for ``Hidden Profile Decision Making in Multi-Agent LLM Groups''},
  year = {2026},
  version = {1.0.0},
  publisher = {Zenodo},
  doi = {10.5281/zenodo.24179025},
  url = {https://doi.org/10.5281/zenodo.24179025}
}