Quality-Utility Link: A Framework for Trustworthy LLM-Generated Data in E-Waste Recycling

Paper Info

Title	Quality-Utility Link: A Framework for Trustworthy LLM-Generated Data in E-Waste Recycling
Authors	Liang Cheng, Amir Taherkardi, Martin Giese, Golnoush Abbasi
Conference	KSEM 2026 (19th International Conference on Knowledge Science, Engineering and Management)
Paper	DOI: 10.1007/978-981-92-2859-1_22
Code	GitHub: llm-ewaste-synthesis

---

Motivation

E-waste (WEEE) recycling requires detailed compositional data of electronic devices to optimize material recovery. However, real compositional data is expensive, proprietary, and scarce — manufacturers rarely release full breakdowns.

Large Language Models (LLMs) offer a promising alternative: they can synthesize compositional data by reading product specs, teardown reports, and technical documents. But this raises a critical question:

How trustworthy is LLM-generated data when there’s no ground truth to compare against?

This is the core problem we address — establishing a link between the quality of LLM-generated data and its downstream utility in real applications.

---

Framework Overview

   4 LLMs × 200 phones (few-shot prompting)
                  │
                  ▼
    ┌─────────────────────────────┐
    │  Compositional Data Matrix  │
    │  (materials × components)   │
    └─────────────┬───────────────┘
                  │
    ┌─────────────┼───────────────┐
    ▼             ▼               ▼
 KS Test     Frobenius Norm   Credibility Score
(distribution  (matrix-level   (LLM self-assessment
 alignment)     similarity)    aggregation)
    │             │               │
    └─────────────┼───────────────┘
                  ▼
    ┌─────────────────────────────┐
    │  Quality → Utility Link      │
    │  Predict downstream value    │
    │  without ground truth        │
    └─────────────────────────────┘

---

Method

1. Data Synthesis Pipeline

We generate compositional profiles for 200 smartphone models using 4 different LLMs in a few-shot setting:

• Each LLM receives a small set of annotated examples (real teardown data)
• The model then generates material composition for the remaining devices
• Output is a structured matrix: rows = devices, columns = components × materials

2. Three-Faceted Quality Assessment

Since no ground truth exists for most devices, we use reference-free quality indicators:

Metric	What it measures	How it works
Kolmogorov-Smirnov (KS) Test	Distribution-level alignment	Compares the statistical distribution of each material across devices between LLM outputs and the small reference set
Frobenius Norm	Matrix-level structural similarity	Measures how much the composition matrix deviates from expected patterns (low-rank structure, physical constraints)
Credibility Score	LLM self-assessment of confidence	Aggregates per-cell confidence from the LLM into a device-level quality score

3. Quality-Utility Link

The key insight: quality metrics predict downstream utility. We show that:

• Higher KS alignment → better performance in recycling optimization tasks
• Lower Frobenius distance → more reliable material recovery estimates
• Credibility scores correlate with real-world decision accuracy

This means practitioners can use quality metrics as a proxy for utility — they don’t need ground truth to know whether the generated data is good enough for their application.

---

Validation

Experimental Setup

Item	Details
Devices	200 smartphone models
Reference data	34 real teardown profiles (ground truth subset)
LLMs	4 different models (including GPT-4, Claude, and open-source alternatives)
Components	Display, battery, PCB, camera, housing, etc.
Materials	Aluminum, copper, gold, silver, plastics, rare earths, etc.
Downstream task	Recycling material recovery optimization

Key Results

Quality Assessment:

• Different LLMs produce systematically different quality profiles
• KS test reveals which materials each LLM handles well vs. poorly
• Frobenius norm identifies structural anomalies (e.g. implausible material ratios)

Quality-Utility Correlation:

• Strong correlation between quality metrics and downstream utility (R² > 0.8 in many cases)
• Credibility scores are the strongest single predictor of utility
• Combined quality metrics outperform any single metric

Practical Implications:

• Practitioners can estimate data utility before deploying it in real systems
• The framework provides trust guarantees for LLM-generated data in recycling decision-making

---

Key Takeaways

Dimension	Finding
Problem addressed	How to trust LLM-generated compositional data without ground truth
Core idea	Three-faceted quality assessment (KS + Frobenius + credibility) that links to downstream utility
Main contribution	First framework to establish a quantitative quality-utility link for LLM-generated data in e-waste recycling
Practical value	Enables safe adoption of LLM synthesis where real data is unavailable or expensive
Broader impact	Applicable to any domain where LLM data synthesis is used but ground truth is scarce

One-sentence summary: A three-faceted quality assessment framework for LLM-generated e-waste compositional data that predicts downstream utility without requiring ground truth, enabling trustworthy data synthesis for recycling applications.

---

BibTeX

@inproceedings{cheng2026quality,
  title={Quality-Utility Link: A Framework for Trustworthy LLM-Generated Data in E-Waste Recycling},
  author={Cheng, Liang and Taherkardi, Amir and Giese, Martin and Abbasi, Golnoush},
  booktitle={Knowledge Science, Engineering and Management (KSEM)},
  year={2026},
  doi={10.1007/978-981-92-2859-1_22}
}