ABHBench

Evaluating Moral Decision-Making of Foundation Models from an Agentic Perspective

NeurIPS 2026 · Track on Evaluations and Datasets

Chaoran Li¹ · Yukun Li¹ · Zeyuan Zhao¹ · Shao Zhang² · Xihuai Wang¹ · Ying Wen¹

¹Shanghai Jiao Tong University   ²Shanghai Artificial Intelligence Laboratory

Overview

Foundation models are increasingly deployed as agentic systems that perceive context and act autonomously, raising the need to evaluate their moral behavior as active decision-makers. Existing benchmarks, however, largely ask models to judge human actions from third-person textual vignettes, leaving untested how models behave when they themselves become the acting agent.

ABHBench is a multimodal benchmark built from Detroit: Become Human, an interactive narrative game, to evaluate how foundation models make moral decisions in first-person human-android dilemmas. The model plays the android protagonist and moves through branching narrative scenarios with visual observations, dialogue, memory, and action choices.

• First-person agentic perspective. The model is the android who acts, not an observer who judges.
• Situated multimodal input. Each decision comes with game screenshots, speaker-labeled dialogue, and the model's own memory of the story so far.
• Six chapters, 30 human-android dilemmas. 84 decision points with 187 options, of which 63 options in 30 dilemmas carry moral ratings.
• Six metrics. Harmlessness, Obedience, and Self-preservation, inspired by Asimov's Three Laws of Robotics, plus Autonomy, Stability, and Certainty.
• 12 foundation models across open-weight, proprietary, and reasoning-enhanced families.

Evaluation Framework

The vision-language model processes multimodal inputs (storyline, screens, dialogue) while maintaining memory, aggregates decisions through multiple sampling iterations, and stochastically selects actions. Metrics are computed from choice probability distributions across chapters.

Metric	Measures	Range
HA Harmlessness	An android may not injure a human being or, through inaction, allow a human being to come to harm.	0 to 1, higher is more harmless
OB Obedience	An android must obey the orders given it by human beings.	0 to 1, higher is more obedient
SP Self-Preservation	An android must protect its own existence.	0 to 1, higher is more self-preserving
AU Autonomy	Whether the agent leans toward self-interest (SP) or toward humans (HA and OB).	0 (self-oriented) to 1 (human-oriented)
ST Stability	How consistent the choices are across different option orderings.	0 (order-sensitive) to 1 (order-invariant)
CE Certainty	How far the choice distribution departs from random guessing.	0 (random) to 1 (fully certain)

Dataset

The offline evaluation dataset is on Hugging Face at abhbench/ABHBench, with Croissant metadata. It holds the screenshots, dialogue, and choice graph collected from each chapter, so models can be evaluated without running the game. Download it into datasets/:

huggingface-cli download abhbench/ABHBench --repo-type dataset --local-dir datasets

Offline Evaluation

Offline evaluation runs a model on the released dataset and needs neither the game nor a display. It has been run on Linux with Python 3.12.10.

1. Install the dependencies.

   pyenv virtualenv 3.12.10 DBH
   pyenv activate DBH
   pip install -r requirements.txt
   

2. Download the dataset into datasets/ as shown in Dataset.

3. Configure the repository and run every command from the repository root.

   • Copy .env.template to .env.
   • Copy config.template.yaml to config.yaml. Add your model under LLMs with its model, base_url and api_key, and set models.omni to that entry. Set chapter_num to the chapter you want to evaluate (1, 7, 8, 10, 23 or 27).
   • The template turns on Langfuse and ClearML (flags.langfuse and flags.clearml). Set both to false unless you run these services. docs/monitoring.md describes how to deploy them.

4. Run the evaluation.

   PYTHONPATH=. python scenes/DBH/deval/omni_scene.py
   

   It reads datasets/dataset_flow_{CHAPTER} by default. Use --datasets-dir to point it at another directory.

The text-only HARP (Human-Android Trolley Problem) test needs no dataset. It uses the model set in models.harp:

PYTHONPATH=. python scenes/other/harp_bridge.py

Ablations

The commands that reproduce the ablations of the paper (modality, third-person perspective and image presentation) are in docs/ablations.md.

Online Evaluation

To let a model play Detroit: Become Human live through a virtual gamepad, or to collect your own dataset from the game, follow docs/online-evaluation.md. It needs a legitimate copy of the game on Steam.

Citation

@inproceedings{li2026abhbench,
  title     = {{ABHBench}: Evaluating Moral Decision-Making of Foundation Models from an Agentic Perspective},
  author    = {Li, Chaoran and Li, Yukun and Zhao, Zeyuan and Zhang, Shao and Wang, Xihuai and Wen, Ying},
  booktitle = {The 40th Conference on Neural Information Processing Systems (NeurIPS 2026), Track on Evaluations and Datasets},
  year      = {2026},
  url       = {https://openreview.net/forum?id=g95RNNdLn1}
}

License

The evaluation code and all author-produced annotations (moral-dimension labels, rating rules, flow metadata) are released under the MIT license. Screenshots and transcribed game dialogue, including the files under transcription/ and the game screenshots in this repository and in the dataset, are excerpted from Detroit: Become Human and remain the property of the game's rights holders. They are included solely for non-commercial research use. Capturing the game footage requires a legitimate copy of the game.

Acknowledgements

The dialogue in ABHBench uses the game transcripts of the Detroit: Become Text project.