language:
- en
tags:
- structural-engineering
- truss-optimization
- reinforcement-learning
- reasoning
- design-benchmark
license: mit
DesignBench-Truss
Overview
DesignBench-Truss is the benchmark split (problems 090–099) of DesignBench, a dataset for teaching and evaluating large language models on iterative structural engineering design via Tree-of-Thought reasoning.
Each problem requires minimising the mass of a 2-D pin-jointed truss structure while satisfying structural safety constraints (factor of safety ≥ 1.5 for both buckling and yielding failure modes). A finite-element analysis oracle (trussme) provides deterministic feedback after every design modification.
Task Description
Given an initial truss design with known joint positions, member cross-sections, material properties, and loading conditions, a model must produce a sequence of discrete grammar actions that transform the design into a feasible, mass-minimised structure.
Grammar Actions
| Action | Signature | Description |
|---|---|---|
| SCALE_PARAM | SCALE_PARAM(member_id_or_all, param, factor) |
Scale one parameter of one (or all) members |
| SCALE_MULTI_PARAM | SCALE_MULTI_PARAM([member_ids], [param:factor, ...]) |
Scale multiple parameters simultaneously |
| ADD_MEMBER | ADD_MEMBER(joint1, joint2, material, shape_type, r, t) |
Add a new pipe member between two joints |
Data Format
benchmark.json
Flat list of examples — one per gold trace — with the following fields:
{
"problem_id": "auto_problem_090",
"trace_id": "auto_problem_090_trace_0",
"problem_text": "PROBLEM: Truss optimization ...",
"initial_state": {
"mass": 142.5,
"fos_buckling": 0.31,
"fos_yielding": 1.09,
"deflection": 0.038,
"is_feasible": false
},
"gold_action_sequence": [
"ADD_MEMBER(5, 6, 6061_T6_Aluminum, Pipe, 0.0497, 0.0056)",
"SCALE_PARAM(all_members, thickness, 1.28)"
],
"reaches_solution": true,
"trace_quality": 0.98
}
problems/
Raw problem JSON files containing joint positions, member topology, loads, and design constraints.
trees/
Full modification trees as JSON. Each tree contains all explored design states (nodes) and the grammar actions connecting them (edges).
traces/
Serialized gold traces — lists of node-sequences extracted from each tree.
Metrics
| Metric | Description |
|---|---|
| feasibility_rate | Fraction of episodes reaching is_feasible=True |
| mass_reduction | (initial_mass - final_mass) / initial_mass |
| grammar_validity_rate | Fraction of steps with parseable grammar actions |
| avg_steps_to_feasibility | Mean episode length when a feasible design is found |
| overall_score | Weighted composite: design_correctness (0.5) + structural_validity (0.25) + reasoning_quality (0.15) + grammar_compliance (0.10) |
Example Usage
import json
with open("benchmark.json") as f:
examples = json.load(f)
print(f"Loaded {len(examples)} benchmark examples")
for ex in examples[:3]:
print(ex["problem_id"], "->", ex["reaches_solution"],
"actions:", len(ex["gold_action_sequence"]))
Citation
@misc{designbench2025,
title = {DesignBench: A Benchmark for Iterative Structural Engineering Design},
author = {DesignBench Authors},
year = {2025},
note = {https://huggingface.co/datasets/DesignBench-Truss}
}