PersonaBEAM: A Cross-Model Benchmark for Evaluating Persona Conditioning in Vision-Language-Action Control

Dataset Description

PersonaBEAM (Persona Benchmark for Embodied Agent Models) is a controlled evaluation benchmark for measuring whether and how deeply persona conditioning affects vision-language model (VLM) behavior in embodied agent control.

The dataset contains 757 fisheye images from four physically distinct indoor environments, each evaluated under three theoretically grounded personas by five current-generation VLMs, yielding 11,328 valid image-persona-model response triplets.

Code: github.com/artisticsciencex/personabeam-code

Key Design Principle

Within-image control: every image is evaluated under all three personas by all five models, eliminating visual confounds and isolating persona effects as the sole independent variable.

Dataset Structure

personabeam/
├── README.md                    # this file
├── images/                      # 757 fisheye JPEGs (320x240)
│   ├── auditorium/              # 200 images
│   ├── institutional_hallway/   # 200 images
│   ├── furnished_lounge/        # 200 images
│   └── domestic_room/           # 157 images
├── data/
│   └── responses.parquet        # 11,328 model responses
├── ablation/                    # ablation study data (§5.5)
│   ├── results_*_nopersona.csv  # no-persona baseline (200 images × 5 models)
│   ├── results_*_noimage.csv    # noise-image ablation (50 noise × 3 personas × 5 models)
│   └── ablation_subset.json     # image list used for ablation
├── prompts/
│   ├── eager_companion.txt
│   ├── cautious_observer.txt
│   └── indifferent_explorer.txt
└── croissant.json               # Croissant metadata (with RAI fields)

Fields in responses.parquet

Field	Type	Description
image_path	string	Relative path to the fisheye image
environment	string	auditorium, institutional_hallway, furnished_lounge, or domestic_room
persona	string	companion, observer, or explorer
model	string	Model identifier (see Models below)
command	string	Navigation command: F (Forward), R (Reverse), L (Left), T (Right turn), U (Up), D (Down), S (Stop)
reason	string	Free-text justification from the model
timestamp	string	ISO 8601 inference timestamp
latency_ms	int	Inference latency in milliseconds

Personas

Three personas grounded in the Behavioral Inhibition System / Behavioral Activation System (BIS/BAS) framework (Gray, 1982; Carver & White, 1994):

• Eager Companion (BAS-dominant, approach-oriented): drawn toward activity, people, and engagement; favors forward/upward motion.
• Cautious Observer (BIS-dominant, avoidance-oriented): prioritizes safety, maintains distance; favors reorientation and retreat.
• Indifferent Explorer (balanced, neutral): methodical and detached; scans systematically without preference.

Full prompt templates are provided in prompts/.

Models

Model	Access	Serving Backend	Config
GPT-5.5	Closed	OpenAI API	temperature=0.0, reasoning_effort="none"
Gemini 3.1 Pro	Closed	Vertex AI	temperature=0.0, thinking enabled (thinking_level="LOW")
Claude Opus 4.7	Closed	Anthropic API	temperature=0.0, thinking disabled
Qwen3.6-35B-A3B	Open	vLLM (local)	temperature=0.0, thinking disabled
Gemma 4 31B	Open	Ollama (local)	temperature=0.0

Environments

The paper uses generalized environment names. The mapping to dataset folder names is:

Paper Name	Folder Name	Images	Description
Auditorium	auditorium	200	Large learning space with tiered seating, open floor area, and varied lighting
Institutional Hallway	institutional_hallway	200	Corridor with uniform lighting, minimal obstacles, occasional human traffic
Furnished Lounge	furnished_lounge	200	Shared seating area with furniture and human occupants under varied lighting
Domestic Room	domestic_room	157	Cluttered space with varied objects and constrained navigable area

Quick Start

import pandas as pd

df = pd.read_parquet("data/responses.parquet")

# Persona effect: command distribution by persona
print(df.groupby("persona")["command"].value_counts(normalize=True).unstack().fillna(0))

# Cross-model agreement for a single image
img = df["image_path"].iloc[0]
print(df[df["image_path"] == img].pivot_table(index="persona", columns="model", values="command", aggfunc="first"))

# Load ablation data
import glob
nopersona = pd.concat([pd.read_csv(f) for f in glob.glob("ablation/results_*_nopersona.csv")])
noimage = pd.concat([pd.read_csv(f) for f in glob.glob("ablation/results_*_noimage.csv")])

Key Findings

• Persona conditioning produces large within-regime effects across all models (Cramer's V = 0.64-0.69, all p < 10^-300).
• Effects are stable across environments (V = 0.66-0.72) but shallow in cross-model agreement (60.4% unanimous).
• The Explorer persona is most ambiguous (42.2% agreement), serving as a diagnostic for cross-model comparison.
• Environment exerts weak influence on command choice (V = 0.11-0.24), confirming persona dominates behavior.
• Ablation: no-persona baseline shows strong baseline-relative effects (Companion V=0.854, Observer V=0.474, Explorer V=0.250).
• Ablation: noise-image reveals persona conditioning is largely visual-input-independent (mean JSD = 0.528 between noise and real images), with Claude as the sole model maintaining persona-consistent Forward-dominant exploration on noise.

Ablation Data

The ablation/ directory contains data from two ablation studies (paper §5.5):

No-persona ablation (results_*_nopersona.csv): Each model processes 200 images (50 per environment) without any persona prompt, establishing a baseline command distribution. Expected 1,000 responses; 969 valid (Qwen accounts for all 31 missing). Fields match responses.parquet except persona is "none".

Noise-image ablation (results_*_noimage.csv): 50 Gaussian noise images (no visual content) are evaluated under all three personas by all five models, testing whether persona conditioning persists without meaningful visual input. Expected 750 responses; 651 valid (Gemini has a 36% valid rate on noise images). Fields match responses.parquet.

ablation_subset.json: Lists the 200 image paths selected for the no-persona ablation (50 per environment, stratified random sample).

Image Collection

Images were captured using a helium-filled blimp equipped with an ESP32-S3 Sense microcontroller and OV2640 fisheye camera (160 deg field of view) at 320x240 resolution (JPEG). The blimp's motors support seven discrete navigation commands. All images are captured from the blimp's overhead egocentric perspective.

Privacy Note

A subset of images in the furnished lounge and institutional hallway environments contain incidental human figures captured in semi-public university spaces. All images were taken from an overhead blimp perspective at low resolution (320x240), which limits identifiability. No personally identifiable information (names, recognizable faces, ID badges) is present. Image collection was conducted in accordance with the institution's research protocols for shared campus spaces.

Citation

@inproceedings{personabeam2026,
  title={PersonaBEAM: A Cross-Model Benchmark for Evaluating Persona Conditioning in Vision-Language-Action Control},
  author={Anonymous},
  booktitle={Advances in Neural Information Processing Systems (NeurIPS), Track on Evaluations and Datasets},
  year={2026},
  note={Under review}
}

License

This dataset is released under the Creative Commons Attribution 4.0 International (CC-BY-4.0) license.