kulits/BrickBench-Runs

BrickBench Runs

Peter Kulits    Yiqing Xu    R. Kenny Jones    Cordelia Schmid    Jiajun Wu

[Project Page] | [Gallery] | [Code]

This repo contains the outputs for each of the BrickBench evaluations.

samples.parquet                     one row per build: model, split, task, verdict, parts, agent, files
<model>/<split>/<task>/
    build.py                        the build program if applicable; build() returns the assembly
    *.py                            additional modules build.py imports, if any
    model.mpd                       the model
    trace.jsonl.gz                  the agent's session log
    thumb-512.avif
    hero-1024.avif
    view-0.avif . . . view-7.avif   full 2k renders for each view

<split> is model, set, or alt; <task> is the task id (model-001, . . .).

Folder	Model
astra	GPT-6 Astra
sol61	GPT-6.1 Sol
opus55	Claude Opus 5.5
opus5	Claude Opus 5
qwen	Qwen 3.8 Flash
sol56	GPT-5.6 Sol
gemini	Gemini 3.8 Flash
deepseek	DeepSeek V4.1 Flash
luna	GPT-5.6 Luna
muse	Muse Spark 1.3
glm	GLM 5.3 Flash
bricknet	BrickNet-14B
brickgpt	BrickGPT
brickwright	BrickWright (GPT-6 Astra)

Citation

@misc{kulits2026brickbenchevaluatingagenticbrick,
      title={BrickBench: Evaluating Agentic Brick Design}, 
      author={Peter Kulits and Yiqing Xu and R. Kenny Jones and Cordelia Schmid and Jiajun Wu},
      year={2026},
      eprint={2610.12452},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2610.12452}, 
}