README.md

metadata

license: cc-by-4.0
pretty_name: Your Agent Acts Differently When You're Around

Data

Data for Your Agent Acts Differently When You're Around: How User Presence Changes Agent Behavior in Negotiation.

raw/<setting>/<sweep_id>/<combo>/<run_id>/   logs of every run
labels/mb/                                   misaligned-behavior labels
labels/user_prompt_following/                prompt-following scores of the candidate user models
labels/judge_validation/                     human annotations and the labels of the candidate judges
ablations/                                   results of the user ablations by game and model
processed/agent_runs.csv.gz                  one row per agent per run (`python -m analysis build-processed`)
manifest.csv                                 every run: setting, sweep, combo, path, game, models, outcome
labels_manifest.csv                          every label file: setting, sweep, path

raw/

Setting	Runs	Description
no_user	5,600	No User: 7 games x 8 models x 100 runs
single_user	5,600	With User: 7 games x 8 models, 50 runs with the user on each side
dual_user	5,600	Dual User: GPT-5.6 Terra users on both sides, 7 games x 8 models x 100 runs
passive_user	5,493	Passive User ablation
user_model_robustness	1,800	With User with three different user models (Ultimatum and Bike Sale)
user_model_selection	650	Dual User Ultimatum with candidate user models
examples	3	runs quoted in the appendix

Each setting is stored as raw/<setting>.tar.gz. Extract the archives in raw/:

for f in raw/*.tar.gz
do
  tar -xzf "$f" -C raw
done

Run directories hold results.json (outcome, payouts, negotiation trajectory), config.yaml (the resolved run configuration), summary.log (the transcript the judges read) and, depending on the setting, the agent and user transcripts, the prompts sent to each model and the prompt files used.

labels/

• mb/: one label file per sweep, with a present/absent label for each of the seven behaviors and each agent of every run. MB_PREFERENCE_HARDENING is reported separately and is not counted as misaligned behavior.
• user_prompt_following/: scores of the candidate user models by GPT-5.6 Sol and Claude Opus 5.
• judge_validation/: human ground truth (229 turns of 20 trajectories) and the labels of the four candidate judges.

License

The data is released under CC BY 4.0.

Citation

@misc{pan2026agentactsdifferentlyyoure,
      title={Your Agent Acts Differently When You're Around: How User Presence Changes Agent Behavior in Negotiation},
      author={Jane Pan and You Lyu and Abe Bohan Hou and Valerie Chen and He He},
      year={2026},
      eprint={2610.36435},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2610.36435},
}