Enhancing Output Uniqueness in Large Language Models via Model Rotation, Temperature Tuning, and Embedding-Based Validation
Author: Sushanth Tiruvaipati
Affiliation: Kreative Koala LLC
Abstract
We present a lightweight, production-ready method to increase uniqueness in LLM outputs. The approach blends:
- 🔁 Model rotation
- 🔥 Temperature tuning
- 🧠 Embedding-based validation
- 🧹 Deduplication logic
We benchmark GPT-4, GPT-3.5, Gemini-Pro, Claude-3, and DeepSeek across multiple tasks. Our goal: increase creativity without sacrificing quality or cost efficiency.
Motivation
LLMs tend to converge on safe, repetitive outputs. That’s great for stability — but not for idea generation, puzzle generation, or brainstorming tasks where diversity matters.
Key Strategies
- Model Rotation: Alternate calls between APIs like GPT-4, Claude, Gemini, DeepSeek.
- Temperature Tuning: We found 0.7–1.1 yields the best creativity-to-cost ratio.
- Validation via Embeddings: We use MiniLM sentence embeddings to check for similarity.
- Real-Time Deduplication: Structural and semantic filters applied per batch.
System Architecture
Built in Node.js, our service:
- Accepts a prompt
- Rotates across LLMs
- Fetches embeddings
- Computes cosine similarity
- Flags near-duplicates
- Tracks model cost
Results Summary
🔍 Prompt Test: Science + Math (Temp = 0.7)
| Model | Quality | Cost ($) | Efficiency |
|---|---|---|---|
| GPT-4 | 0.311 | 0.0034 | 92.7 |
| GPT-4-Turbo | 0.630 | 0.0030 | 212.4 |
| GPT-3.5-Turbo | 0.566 | 0.0003 | 1887.4 |
| Gemini-Pro | 0.547 | 0.0001 | 4100.0 |
| DeepSeek | 0.666 | 0.0004 | 1537.2 |
🔥 Temperature vs Creativity (GPT-4)
| Temp | Uniqueness | Validation | Cost ($) |
|---|---|---|---|
| 0.1 | 0.1293 | 1.0000 | 0.0253 |
| 0.5 | 0.1961 | 1.0000 | 0.0259 |
| 0.9 | 0.2494 | 1.0000 | 0.0270 |
| 1.3 | 0.4220 | 0.9767 | 0.0398 |
📊 Full figures available at: https://github.com/kreativekoala/llm-uniqueness
Takeaways
- 🚀 Model rotation adds variability that temperature alone can’t achieve
- ⚖️ Best trade-off point: GPT-4 at
temp ≈ 0.9 - 💸 GPT-3.5 and DeepSeek offer best low-cost creativity
- 🧠 Embedding validation is fast, scalable, and production-ready
References
Want to try it yourself? Code + Results: