LLM Evaluations Notes (Part 3)
In this part, we'll cover why evaluating LLM applications is fundamentally different from testing traditional software. This is one of the most important concepts for AI Engineers and is frequently discussed in interviews.
Chapter 6: Why LLM Evaluation is Hard
English Translation
After explaining why evaluation is important, the speaker asks:
"If evaluation is so important, why doesn't everyone evaluate their LLM applications?"
The answer is simple:
Because evaluating LLM applications is not straightforward.
Compared to traditional software, LLM-based applications are much more difficult to test due to their probabilistic nature and multiple quality dimensions.
Traditional Software vs LLM Applications
Traditional Software
Imagine building a calculator.
Input:
2 + 2
Output:
4
No matter who uses it or when they use it:
2 + 2 = 4
always.
This property is called Determinism.
Deterministic System
Definition
A deterministic system always produces the same output for the same input.
Input
│
▼
Program
│
▼
Same Output
Example:
5 × 10
Output
50
Every single time.
Examples
Calculator
10 / 2 = 5
Sorting Algorithm
Input
[5,3,2]
Output
[2,3,5]
Binary Search
Same input
↓
Same output
Every time.
Characteristics of Deterministic Systems
✔ Predictable
✔ Repeatable
✔ Easy to Test
✔ Easy to Debug
✔ Easy to Measure
LLM Applications
Now consider ChatGPT.
Prompt:
Explain Machine Learning.
Today's answer:
Machine Learning is a branch of AI...
Tomorrow's answer:
Machine Learning enables computers to learn...
Next week:
ML is a field where algorithms improve using experience...
All answers are correct.
But they are different.
Why?
Because LLMs are Probabilistic Models.
Probabilistic System
Definition
A probabilistic system can generate multiple valid outputs for the same input.
Prompt
│
▼
LLM
│
├─────────────► Output 1
│
├─────────────► Output 2
│
└─────────────► Output 3
Same Prompt
↓
Different Answers
↓
All can be correct.
Why Does This Happen?
LLMs don't "look up" one fixed answer.
Instead, they predict the next most probable token repeatedly until the response is complete.
At every step, the model chooses from many possible words.
Example
Prompt
The capital of France is
Possible next tokens
Paris
99.7%
Lyon
0.2%
Marseille
0.1%
The model selects based on probability.
For creative prompts, probabilities are much closer, producing more varied outputs.
Interview Question
Why are LLMs called probabilistic models?
Because they generate text by predicting the probability distribution of the next token rather than executing fixed rules. Therefore, the same prompt can produce different valid responses.
Consequence for Testing
Suppose you ask
Explain Neural Networks.
Version 1
Uses 120 words.
Version 2
Uses 300 words.
Version 3
Uses examples.
Version 4
Uses equations.
Which one is correct?
Actually...
All of them.
So unlike software testing,
there isn't always a single expected output.
Biggest Challenge
Traditional Testing
Expected Output
│
▼
Compare
│
▼
Pass / Fail
LLM Evaluation
Prompt
│
▼
Many Possible Correct Answers
│
▼
Need Quality Measurement
Instead of asking
Is it correct?
we ask
How good is this answer?
Chapter 7: Single Metric vs Multi-Dimensional Evaluation
Traditional software usually has one important metric.
Example
Calculator
Question
10 × 5
Answer
50
Correct?
YES
Done.
Only One Metric
Correct
or
Incorrect
Nothing else matters.
LLM Evaluation
Imagine building a RAG chatbot.
Question
What is Transformer Architecture?
Suppose the answer is technically correct.
Can we deploy it immediately?
No.
We need to evaluate many more aspects.
Multiple Evaluation Dimensions
Prompt
│
▼
LLM Response
│
▼
┌────────────────────┐
│ Factuality │
│ Completeness │
│ Groundedness │
│ Tone │
│ Relevance │
│ Safety │
│ Latency │
│ Cost │
└────────────────────┘
1. Factuality
Question
Is the answer factually correct?
Example
Who discovered Gravity?
Correct
Isaac Newton
Wrong
Albert Einstein
2. Completeness
Question
Did the answer cover everything?
Example
Question
Explain RAG.
Poor
RAG combines search and LLM.
Good
Retrieval
Embedding
Vector Database
Retriever
Prompt Construction
LLM Generation
Much more complete.
3. Groundedness
Question
Did the answer come from the retrieved documents?
Suppose your PDF says
Maximum leave = 20 days.
Chatbot replies
Maximum leave = 40 days.
Even if it sounds convincing,
it is not grounded.
Groundedness checks whether the response is supported by the provided context.
4. Tone
Example
Healthcare chatbot.
User
I have cancer.
Bad response
That's unfortunate.
Better response
I'm sorry to hear that. Please consult a qualified medical professional for personalized advice.
Same information.
Different tone.
Tone matters.
5. Latency
How long does it take to answer?
User
│
▼
Prompt
│
▼
LLM
│
▼
Response
If users wait 30 seconds,
they will likely abandon the application.
6. Cost
Every API call costs money.
Example
GPT-4
↓
Long Response
↓
More Tokens
↓
Higher Cost
Evaluation includes checking whether the system is financially sustainable.
Other Metrics
Production systems may also evaluate:
- Fluency
- Relevance
- Helpfulness
- Toxicity
- Bias
- Privacy
- Robustness
- Consistency
- Hallucination Rate
- Citation Quality
- Faithfulness
- Safety
Evaluation Pyramid
Safety
Hallucination
Groundedness
Completeness
Relevance
Correctness
Each layer contributes to overall quality.
Why Different Applications Need Different Metrics
Customer Support Bot
Important metrics:
- Accuracy
- Tone
- Latency
Medical Assistant
Important metrics:
- Factuality
- Hallucination
- Safety
Coding Assistant
Important metrics:
- Correctness
- Executability
- Efficiency
RAG Chatbot
Important metrics:
- Retrieval Quality
- Groundedness
- Faithfulness
- Citation Quality
AI Agent
Important metrics:
- Planning
- Tool Usage
- Task Completion
- Recovery from Errors
Interview Questions
1. Why is LLM evaluation harder than software testing?
Because LLMs are probabilistic systems. The same input can produce multiple valid outputs, so evaluation must assess response quality across multiple dimensions rather than comparing against one fixed expected answer.
2. What is a deterministic system?
A deterministic system always produces the same output for the same input.
Example: Calculator, Binary Search, Sorting Algorithms.
3. What is a probabilistic system?
A probabilistic system generates outputs based on probability distributions, allowing multiple valid responses for the same input.
Example: ChatGPT, Claude, Gemini.
4. What are the major evaluation dimensions of an LLM response?
Common dimensions include:
- Factuality
- Completeness
- Groundedness
- Tone
- Relevance
- Safety
- Latency
- Cost
- Faithfulness
- Robustness
Summary (Part 3)
- Traditional software is deterministic: the same input always produces the same output.
- LLMs are probabilistic: the same prompt can generate multiple valid responses.
- Traditional software is typically evaluated with a simple pass/fail criterion.
- LLM applications require multi-dimensional evaluation, considering not only correctness but also factuality, completeness, groundedness, tone, latency, cost, safety, and more.
- Different AI applications prioritize different evaluation metrics depending on their domain and purpose.
In Part 4, we'll dive into the learning roadmap introduced in the transcript:
- What LLM Evals actually are.
- The LLM Evaluation Landscape (frameworks, metrics, techniques, and tools).
- LLM benchmarks such as MMLU, HellaSwag, HumanEval, GSM8K, and others.
- Building a custom evaluation pipeline with golden datasets and rubrics.
- Evaluation strategies for RAG systems, AI agents, safety, and operational monitoring.