| --- |
| pretty_name: AssemblyWorldBench Results |
| license: other |
| license_name: per-source-terms |
| license_link: LICENSE.md |
| tags: |
| - 3d |
| - assembly |
| - agents |
| - benchmark |
| - agent-trajectories |
| size_categories: |
| - 1K<n<10K |
| configs: |
| - config_name: default |
| data_files: |
| - split: all |
| path: index/samples.parquet |
| extra_gated_prompt: >- |
| These results embed part geometry, manual pages and reference images from PartNet (via |
| Manual-PA), IKEA-Manual, AssemblyBench (Fusion 360 Gallery) and Fantastic Breaks inside |
| the recorded episodes and transcripts. These assets keep their original rights. Access is |
| granted for non-commercial research only, and you must comply with the terms of every |
| source dataset. |
| extra_gated_fields: |
| Name: text |
| Affiliation: text |
| I will use this dataset for non-commercial research only: checkbox |
| I will comply with the terms of each source dataset listed in the card: checkbox |
| --- |
| |
| # AssemblyWorldBench Results |
|
|
| Complete records of every agent run behind the main result tables of *AssemblyWorld: |
| Rethinking 3D Assembly with General-Purpose Agents*: for each evaluated sample, the prompt, |
| the full agent transcript with timestamped tool calls, the agent's own report, the exported |
| final episode, and its geometric scores, plus the aggregates the paper reports. The benchmark |
| itself is [`AssemblyWorld/AssemblyWorldBench`](https://huggingface.co/datasets/AssemblyWorld/AssemblyWorldBench). |
|
|
| `index/samples.parquet` has one row per sample (4,624 rows): table, system, block, sample, |
| run status, the agent's self-reported outcome, SCD / PA / SR, wall time, tokens, cost and |
| WebMCP call count, with the path to the sample directory. It is what the dataset viewer shows. |
|
|
| ## Contents |
|
|
| | Directory | Paper | Agent system | Samples | |
| |---|---|---|---| |
| | `assemblyworldbench/<system>/` | Table 1, AssemblyWorldBench | 8 systems: Claude Fable 5.1, Claude Opus 5, Claude Sonnet 5 (Claude Code); GPT-6 Astra, GPT-5.6 Sol, GPT-5.6 Terra, DeepSeek V4.1 Flash, Qwen3.8 Max (Codex) | 100 each | |
| | `partnet/gpt-6-astra/` | Table 2, PartNet-Assembly | GPT-6 Astra + Codex | storage 148 × 3 references; chair 791 and table 533 without reference and with a final image; chair and table 100-object manual subsamples | |
| | `ikea-manual/gpt-6-astra/` | Table 3, IKEA-Manual | GPT-6 Astra + Codex | 102 | |
| | `assemblybench/gpt-6-astra/` | Table 4, AssemblyBench | GPT-6 Astra + Codex | 280 (279 reported) | |
| | `fantastic-breaks/gpt-6-astra/` | Table 5, Fantastic Breaks | GPT-6 Astra + Codex | 150 | |
|
|
| Every directory follows `<table>/<system>/<block>/` with `run.json`, `task.txt`, |
| `samples/<sample>/` (`input.json`, `prompt.txt`, `result.json`, `conversation.jsonl`, |
| `final.episode.zip`) and `evaluation/` (`metrics.jsonl`, `metrics_summary.json`, `meta.json`). |
| Each system has an `experiment.json` naming the paper table and row. `AGENTS.md` documents |
| every file and field, the per-harness usage conventions, and all edits made at export. |
|
|
| Supplementary studies in the paper (robustness to layout, distractor and missing-part |
| perturbations; GARF refinement) are not included. |
|
|
| ## Reproducing the scores |
|
|
| Every block directory is a valid run directory for |
| [`assembly-world-agent`](https://github.com/AssemblyWorld/assembly-world-agent): |
|
|
| ```sh |
| # one block |
| uv run --locked --extra episodes python scripts/evaluate_run.py \ |
| results/ikea-manual/gpt-6-astra/ikea-manualbook --output /tmp/ikea |
| # one Table 1 system, aggregated with the frozen benchmark rules |
| uv run --locked --extra episodes python scripts/evaluate_run.py \ |
| results/assemblyworldbench/claude-fable-5-1/{partnet-none,partnet-final-image,ikea-manualbook,assemblybench-manualbook,fantastic-breaks-none} \ |
| --benchmark results/assemblyworldbench/benchmark/benchmark.json --output /tmp/fable |
| ``` |
|
|
| The evaluator rebuilds the ground truth from the pinned source dataset revisions on |
| Hugging Face. Scores use `assembly-evaluation-v2`: SCD (whole-shape Chamfer ×1000), PA |
| (fraction of parts within Chamfer 0.01 of their matched part) and SR (all parts correct). |
| Re-scoring reproduces the recorded `metrics.jsonl` rows. |
|
|
| ## Known limitations |
|
|
| - AssemblyBench: the reported aggregate excludes sample 6772 (the model refused the task) |
| and uses the selected retry of sample 6994. The original 280-sample aggregate used the |
| first attempt of 6994, which is not included, so only the 279-sample number is reproducible. |
| - The four collaborator-run PartNet blocks (chair and table without reference and with a |
| final image) keep only the attempt selected for each sample; their superseded attempts |
| were not transferred, so retry counts are unknown. |
| - A few runs ended with a CLI error after the episode was exported (`result.json.status = |
| failed`, `archive.status = saved`); they are scored like any other sample. |
| - Model reasoning text is not exported by the agent CLIs; only reasoning token counts appear. |
|
|
| ## Terms |
|
|
| The final episodes embed source part geometry, and transcripts embed the reference images |
| given to the agent. These assets keep their original rights; see `LICENSE.md`. Source |
| datasets: [`AssemblyWorld/partnet-manualpa`](https://huggingface.co/datasets/AssemblyWorld/partnet-manualpa), |
| [`AssemblyWorld/ikea-manual`](https://huggingface.co/datasets/AssemblyWorld/ikea-manual), |
| [`AssemblyWorld/assemblybench`](https://huggingface.co/datasets/AssemblyWorld/assemblybench), |
| [`AssemblyWorld/fantastic-breaks`](https://huggingface.co/datasets/AssemblyWorld/fantastic-breaks). |
| Please also cite Mo et al., *PartNet*, 2019; Zhang et al., *Manual-PA*, 2025; |
| Wang et al., *IKEA-Manual*, 2022; Li et al., *AssemblyBench*, 2026; Lamb et al., |
| *Fantastic Breaks*, 2023. |
|
|
