朗慧科技

Langhui AI STEM-Questions · Subject Hub

Technology & Engineering · University-Level Multimodal Benchmark

Covering Computer Science, Electronic Engineering, Mechanical Engineering, Materials Science, Civil Engineering, and Chemical Engineering — 6 sub-domains with a unified 5-model×5-pass pass@k ≤ 40% difficulty threshold.

25% of Total Bank 6 Sub-Domains 7 Diagram Types Code+Diagram Binding
中文

Subject Overview

25%
of Total Question Bank
Tied for 2nd largest
6
Sub-Domains
CS · EE · ME · Materials · Civil · ChemE
7
Diagram Types
Circuit · CAD · Flow · Phase · Structure · Code · Schematic

Engineering disciplines present unique multimodal challenges: circuit schematics, engineering CAD drawings, thermodynamic cycle diagrams, material phase diagrams, structural analysis diagrams, and code listings interleaved with visual outputs. These require models to reason across radically different visual languages — from geometric constraints in CAD to symbolic notation in circuit diagrams. Langhui Engineering benchmark systematically covers these high-value difficulty zones.

Sub-Domains & High-Value Difficulty Zones

Computer Science (40%)

Algorithm & Data Structure Visualization

BST operations, graph traversal, sorting animation traces, heap construction.

Code Debugging & Reasoning

Python/C++ bug localization, output prediction, complexity analysis with code+output screenshots.

Evidence: SOTA models fail to debug multi-file dependency bugs.

Computer Architecture

Pipeline diagrams, cache mapping, branch prediction schematics.

Electronic Engineering (20%)

Analog Circuit Analysis

Op-amp circuits, active/passive filters, Bode plot interpretation.

Digital Logic & VLSI

Karnaugh maps, state machine diagrams, timing diagrams, transistor-level schematics.

Signal Processing

Fourier transform visualizations, spectrograms, filter design.

Mechanical Engineering (15%)

Engineering Drawing & CAD

Orthographic projection, section views, dimensional tolerance interpretation.

Thermodynamics & Fluid Mechanics

P-V/T-S diagrams, Bernoulli principle diagrams, pipe flow analysis.

Mechanics of Materials

Stress-strain curves, beam deflection diagrams, Mohr's circle analysis.

Materials Science (10%)

Phase Diagrams

Fe-C, binary eutectic, ternary phase diagram interpretation and lever rule calculations.

Evidence: Fe-C phase diagram reasoning is a known blind spot for VLMs.

Crystal Structures

BCC/FCC/HCP identification, Miller indices, XRD pattern interpretation.

Civil Engineering (8%)

Structural Analysis

Shear/moment diagrams, truss analysis, influence lines.

Geotechnical Engineering

Soil classification charts, consolidation curves, slope stability diagrams.

Chemical Engineering (7%)

Process Flow Diagrams (PFD)

PFD/P&ID interpretation, unit operation identification, mass/energy balance.

Reaction Engineering

Reactor design diagrams, residence time distribution, kinetic curve analysis.

Difficulty Distribution

Level Stage Share Typical Question Types
L1 Lower UndergradIntroductory courses10%Basic circuit analysis, simple code output prediction
L2 Upper UndergradCore engineering courses50%Multi-stage circuit analysis, CAD dimension reasoning
L3 Graduate EntryFE/PE exam level30%Cross-domain synthesis, design optimization problems
L4 ResearchAdvanced / research10%Novel circuit design, paper-level algorithm analysis

Sample Questions

STEM-2026-ENG-0000456L2 Upper Undergrad

Python · Recursion Bug Localization

Prompt: A Python recursive function to compute the nth Fibonacci number is provided as a screenshot. The code has a subtle bug: it uses n-1 and n-1 (instead of n-1 and n-2) for the two recursive calls. Identify the bug, explain why it causes infinite recursion for n>1, and provide the corrected code.
Answer: The recursive call fib(n-1) is duplicated instead of fib(n-1) + fib(n-2). For n>1: each call spawns two identical calls to fib(n-1), never reducing the subproblem to the base case, resulting in infinite recursion until stack overflow. Correct: return fib(n-1) + fib(n-2).

5-Model Evaluation

GPT-5.1
3/5
Claude Opus 4.6
2/5
Gemini-3.1-Pro
2/5
Qwen3.6-Plus
1/5
DeepSeek-V4
1/5
STEM-2026-ENG-0000457L3 Graduate Entry

Circuits · 2nd-Order Active Low-Pass Filter

Prompt: A Sallen-Key 2nd-order active low-pass filter schematic is shown. Given R1 = R2 = 10kΩ, C1 = 100nF, C2 = 47nF, calculate the cutoff frequency fc and the quality factor Q. Determine whether the response is Butterworth, Bessel, or Chebyshev.
Answer: fc = 1/(2π√(R1R2C1C2)) = 1/(2π√(10k·10k·100n·47n)) ≈ 1/(2π√(4.7e-10)) ≈ 1/(2π·2.168e-5) ≈ 7,340 Hz. Q = √(R1R2C1C2)/(C2(R1+R2)) = √(4.7e-10)/(47n·20k) = 2.168e-5/9.4e-4 ≈ 0.023. This is heavily overdamped — closer to a Bessel characteristic.

5-Model Evaluation

GPT-5.1
1/5
Claude Opus 4.6
0/5
Gemini-3.1-Pro
1/5
Qwen3.6-Plus
0/5
DeepSeek-V4
0/5
STEM-2026-ENG-0000458L3 Graduate Entry

Materials · Fe-C Phase Diagram — Eutectoid Steel

Prompt: The Fe-Fe3C phase diagram (partial, 0-6.67 wt% C) is shown. An alloy with 0.77 wt% C (eutectoid composition) is slowly cooled from 850°C to room temperature. Identify all phase transformations, the resulting microstructure, and calculate the relative amounts of each phase at room temperature.
Answer: At 850°C: 100% austenite (γ). At 727°C (eutectoid): γ(0.77%C)→α(0.022%C)+Fe3C(6.67%C) via pearlite formation. At room temperature: ~100% pearlite (α+Fe3C lamellae). Lever rule: Wα = (6.67-0.77)/(6.67-0.022) = 88.7%, WFe3C = 11.3%.

Step-by-Step Reasoning
  1. At 850°C, 0.77wt%C → single-phase γ (austenite) region.
  2. Cool to 727°C: eutectoid isotherm. γ(0.77%) decomposes via eutectoid reaction.
  3. Eutectoid: γ→α+Fe3C forms pearlite (alternating lamellae).
  4. Lever rule at room temperature: Wα = (6.67-0.77)/(6.67-0.022) ≈ 88.7%.
  5. WFe3C = (0.77-0.022)/(6.67-0.022) ≈ 11.3%.

5-Model Evaluation

GPT-5.1
2/5
Claude Opus 4.6
1/5
Gemini-3.1-Pro
1/5
Qwen3.6-Plus
0/5
DeepSeek-V4
0/5

Delivery & API

JSONL Bulk Download: Full fields + image URLs + 5-model evaluation + reasoning chains.
REST API: GET /DataAssetsAPI/stem-questions/engineering. For enterprise API keys, contact lk@langhuiai.com.

FAQ

How is the difficulty threshold enforced?

Every question must achieve ≤40% pass rate across 5 models × 5 passes (GPT-5.1 / Claude Opus 4.6 / Gemini-3.1-Pro / Qwen3.6-Plus / DeepSeek-V4). Human expert accuracy ≥95%, dual-review.

What sub-domains does the Engineering benchmark cover?

6 sub-domains: Computer Science (40%), Electronic Engineering (20%), Mechanical Engineering (15%), Materials Science (10%), Civil Engineering (8%), Chemical Engineering (7%).

What diagram types does the Engineering benchmark emphasize?

Circuit schematics, CAD engineering drawings, flowcharts, phase diagrams, crystal structures, code+output screenshots, and process flow diagrams (PFD/P&ID).

Does the CS portion include programming questions?

Yes — code debugging, output prediction, complexity analysis, and algorithm visualization. All code questions include screenshot-based code listings to test visual code understanding, not just text-based reasoning.

Is there overlap with public engineering benchmarks?

Zero overlap. <0.5% pHash+n-gram fingerprint overlap with MMMU-Engineering, EngineeringQA, and 6 other public benchmarks.