ziyjiang/WorldAuditBench-trajectories

WorldAuditBench Trajectories

Paper · Code · Project · Dataset

Simplified experiment records: screenshots, actions, model responses, final reports, and judge results. Each experiment covers 213 tasks: 126 Unreal tasks and 87 Three.js tasks.

Downloads

Experiment	Task records	Completed runs	Screenshots	File
GPT-6 Astra VLM	213	213	8,733	gpt-6-astra-vlm.tar.gz
Qwen3.8 Flash VLM	213	132	6,542	qwen3.8-flash-vlm.tar.gz
Qwen3.8 Flash VLA	213	206	25,560	qwen3.8-flash-vla.tar.gz

Qwen records preserve the selected final attempt for each task, including 73 failed and 8 interrupted VLM runs, and 7 failed VLA audits. Partial runs retain their available actions and messages. Missing reports and judge results are null; missing results are not assigned zero scores. The VLA exploration recordings remain available even when the subsequent audit failed.

GPT-6 Astra VLA responses are also available as text-only JSONL: 213 tasks, 224 nonempty auditor responses, submitted reports, and judge results. Its screenshot archive is not currently published here.

Folder structure

Each archive contains one experiment folder:

qwen3.8-flash-vlm/
├── A01/
│   ├── trajectory.json
│   └── screenshots/
│       ├── 000.jpg
│       ├── 001.jpg
│       └── ...
├── A02/
└── ...

Image paths in trajectory.json are relative to the task folder.

trajectory.json

Field	Contents
task	Task ID.
status	Original run status, included in the Qwen exports.
input	Original instruction and scene description.
agent_final_report	Complete final response and submitted summary and bugs. null when no report was submitted.
judge_output	Original score and explanation translated into English. 1 means the target anomaly was identified; 0 means it was not. null if judging did not run.
initial_observation	Initial screenshot, or null if none was recorded.
steps	Original actions or controls, with corresponding observation paths.
responses	Original emitted text messages.
reasoning	Qwen's original logged thinking text.
recording	VLA explorer model, tick duration, and screenshot interval.

VLM: actions and observations

For VLM runs, 000.jpg is the initial observation; 001.jpg is the observation after action 1. Actions include move, turn, look, wait, and interact.

{
  "step": 1,
  "action": {"kind": "turn", "params": {"degrees": 90.0}},
  "observation": "screenshots/001.jpg"
}

GPT-6 Astra has 40 actions per task and 9,220 response messages overall. Qwen VLM includes partial runs and contains 6,729 responses and 8,427 reasoning blocks. For both models, after_step gives the most recent completed action (0 before the first action). Multiple messages can occur between two actions.

Qwen: responses and reasoning

Qwen VLM and VLA logs contain both text responses and thinking blocks. They are exported separately without rewriting their content. sequence preserves their joint chronological order:

messages = [
    (m["sequence"], kind, m["text"])
    for kind in ("responses", "reasoning")
    for m in trajectory[kind]
]
for sequence, kind, text in sorted(messages):
    print(sequence, kind, text)

The GPT-6 logs do not contain reasoning text, so no reasoning trace is added to those exports.

VLA: exploration followed by auditing

Open-P2P first explores the environment; the language model then audits that recording. Each VLA recording contains 1,200 control ticks and 120 screenshots, sampled every 10 ticks. steps[].action.params stores the original keys, dx, and dy. observation is null on ticks without a saved screenshot. The first saved screenshot follows tick 0 and is also referenced by initial_observation.

Qwen VLA contains 853 auditor responses and 1,631 reasoning blocks. These messages describe the later audit, so they do not have after_step or represent the explorer's per-tick decisions.

Screenshots and evidence

Unreal screenshots were rendered again along the recorded actions without the minimap mask; previously verified renders were recovered from backup when available. Lighting, reflections, and repaired surfaces can differ from the historical observations. Three.js screenshots are the original observations. Original actions, emitted text, reports, and judge scores are retained. Judge explanations are translated into English. Historical instructions may still mention the original HUD mask. Judging was not rerun on the new screenshots.

VLM report references such as a23 correspond to screenshots/023.jpg. In VLA reports, a0 refers to screenshots/000.jpg, and a0.fN refers to screenshot N + 1 (for example, a0.f29 is screenshots/030.jpg). Crop references such as a28#c1 and judge figure numbers refer to the original judging evidence; separate crops and intermediate video frames are not included.

Usage

hf download ziyjiang/WorldAuditBench-trajectories qwen3.8-flash-vlm.tar.gz --repo-type dataset --local-dir .
tar -xzf qwen3.8-flash-vlm.tar.gz

import json
from pathlib import Path

folder = Path("qwen3.8-flash-vlm/A01")
trajectory = json.loads((folder / "trajectory.json").read_text())
print(trajectory["agent_final_report"])
print(trajectory["judge_output"])
for step in trajectory["steps"]:
    image = folder / step["observation"] if step["observation"] else None
    print(step["step"], step["action"], image)