AI ENGINEERING NOTES
PAGE 05 • MODEL EVALUATION

LLM Benchmarking

Benchmarks • Core Capabilities • Model Selection
05 / 20
☰ Page Navigation
PAGE 05

LLM Benchmarking

Understanding Model Evals, Standardized Benchmarks, Custom Evaluation, and the Eight Core LLM Capabilities.

INTRODUCTION

LLM evaluation has two broad layers: Model Evals, which directly test an LLM's capabilities, and Application Evals, which test an LLM-powered application such as a RAG system or agent. This page moves from application-level evaluation into the model itself.

The central idea is simple: if an AI Engineer cannot measure a model's capabilities, it becomes difficult to choose the right model, compare new releases, assess safety, or justify deployment decisions.

AT A GLANCE
Model Eval
Measures an LLM's capabilities and behavior.
Benchmark
Standardized shared test for model comparison.
Custom Eval
Evaluation built around your actual task/data.
Core Capabilities
Eight major capability areas organize most benchmarks.

On This Page

  1. Why AI Engineers Need Model Evals
  2. What Is a Model Eval?
  3. The Four-Step Evaluation Process
  4. Benchmarks and Custom Evals
  5. Why Benchmarks Alone Are Not Enough
  6. The Eight Core Capabilities
  7. Interview Questions
  8. Quick Revision

1. Why Do AI Engineers Need Model Evals?

Model Evals exist to measure what an LLM can actually do. For an AI Engineer, however, the important question is not simply “Why do frontier labs evaluate models?” but “Why do I need model evaluation when my job is to build applications?”

Core principle: “If you can't measure, you can't improve.” Model Evals provide a mechanism for measuring different capabilities of an LLM.

Reason 1 — Compare Models and Select the Right One

Suppose a company is building a RAG application and must choose between two LLM providers. In a professional engineering discussion, “both are good” is not enough. The team needs concrete evidence showing which model performs better on the capabilities that matter for the application.

Model Evals provide comparable scores that support a model-selection decision.

Reason 2 — Track Whether New Models Actually Improve

Imagine an application is running on one model and a newer model is released. A manager may ask whether the team should migrate. Model evaluation provides numbers that can be used to compare the old and new models rather than relying on marketing claims or intuition.

Reason 3 — Evaluate Safety and Reliability

Model Evals can test whether a model is safe enough for a particular use case. The lecture specifically highlights concerns such as hallucination, jailbreak susceptibility, and responsible behavior.

Reason 4 — Decide Between Proprietary APIs and Self-Hosting

An AI Engineer may need to decide between using a proprietary model through an API and deploying an open-source model on company infrastructure. Capability, cost, and practical suitability must be compared. Model evaluation helps make this decision evidence-based.

Without Model Evals: model selection becomes largely guesswork. You are effectively “blind” to systematic capability differences.

2. What Is a Model Eval?

DEFINITION

A Model Eval is a systematic process of measuring an underlying model's capabilities, behavior, reliability, and operational characteristics under controlled conditions.

Simple Explanation

A model eval is essentially an exam for an LLM. You decide what capability you want to test, give the model a defined test, run it under controlled conditions, and then score and interpret the result.

Why There Is No Single “LLM IQ” Number

LLMs are general-purpose systems with many different capabilities. There is no single evaluation that fully describes all of them. Reasoning, coding, mathematics, long-context handling, multimodal understanding, tool use, safety, and instruction following can require different tests.

3. The Four-Step Model Evaluation Process

1. Select Capability
2. Bring a Test
3. Run Under Fixed Protocol
4. Score & Interpret

Step 1 — Decide What to Test

Choose the capability: reasoning, coding, safety, instruction following, or another target capability.

Step 2 — Bring a Test

Select or construct a test mechanism that measures the chosen capability. This can be a standardized benchmark or a custom evaluation set.

Step 3 — Run the Model Under a Fixed Protocol

Fix the relevant conditions so that the evaluation is repeatable. When comparing multiple models, comparable conditions are essential.

Step 4 — Score and Interpret

After the test, calculate the score and interpret what it means for model capability and the intended decision.

4. Two Types of Model Tests

TypeWhat It IsPrimary Purpose
BenchmarkA standardized, shared test used across models. Examples mentioned in the lecture include MMLU and SWE-bench.Compare models on common ground and understand general capabilities.
Custom EvaluationA dataset/test assembled from the actual task the application cares about.Measure suitability for a specific application rather than generic usefulness.
Key distinction: Benchmarks answer “How capable is this model generally?” Custom evals answer “How suitable is this model for our actual work?”

5. Case Study — Why Custom Evals Matter

Consider an application that reads incoming emails and routes them into categories such as Billing, Technical, and Refund. Two model choices are available:

PropertyModel AModel B
Model profileLarge, powerful, top-of-leaderboardSmaller, cheaper, mid-table on public benchmarks
Illustrative token cost~$15 / 1M tokens~$0.50 / 1M tokens
Classification accuracy94%91%
Urgency accuracy88%87%
Illustrative latency~4.1 s~9 s

The lecture's example then proposes building a golden dataset from roughly 200–500 historical emails, labeling each email according to the application's target categories, and evaluating both models on exactly that task.

Although Model A wins on generic public benchmarks, Model B is only slightly behind on this particular application task while being substantially cheaper. The conclusion is therefore that Model B can be the better value proposition for this specific workload.

Lesson: If the team had relied only on public benchmarks, it would likely have selected Model A. The custom evaluation exposed the application's actual trade-off.

6. The Eight Core LLM Capabilities

LLMs are general-purpose models. They can perform text generation, sentiment analysis, summarization, part-of-speech tagging, coding, and many other tasks. Because there is no single score that captures every capability, benchmarks are organized around specific capability areas.

01

Knowledge & Reasoning

Factual knowledge and the ability to connect multiple facts to reach conclusions.

02

Coding & Software Engineering

Code generation, debugging, testing, repository-level work, commands, and tool/API interaction.

03

Mathematics

Accurate symbolic and numerical reasoning across difficulty levels.

04

Long Context

Effective use of information contained in very large inputs.

05

Vision & Multimodal

Understanding images, videos, and information beyond text.

06

Agentic & Tool Use

Using tools, browsing, APIs, structured calls, and computer environments to perform actions.

07

Safety & Alignment

Responsible behavior, resistance to harmful/adversarial use, truthfulness, and security-related behavior.

08

Instruction Following

Following user-specified constraints, formats, styles, lengths, and clarification requirements.

7. Capability 1 — Knowledge & Reasoning

Definition

This capability combines the amount of factual knowledge a model has with its ability to connect facts and reason over them.

What Is Measured?

  • Factual recall: knowledge across areas such as biology, physics, chemistry, and history.
  • Multistep logical reasoning: connecting multiple facts in the correct sequence to reach a conclusion.
BENCHMARK EXAMPLE

MMLU

The lecture cites MMLU as a benchmark evaluating models across 57 subjects.

Example

A task asking the model to connect a long chain of historical or scientific events and explain how those events contributed to the structure of modern society tests both factual knowledge and reasoning.

Why It Matters

Strong performance is treated as an indicator of how intelligent or capable a model is. It is particularly relevant to research assistants, technical document analysis, professional assistants, and complex question answering.

8. Capability 2 — Coding & Software Engineering

This capability evaluates whether an LLM can perform useful programming and real software-engineering work, not merely produce code that looks plausible.

What Is Measured?

  • Generate functional code from natural-language requirements.
  • Generate test cases.
  • Improve code based on failing tests or errors.
  • Find and fix bugs in an existing codebase.
  • Perform multi-file and long-horizon engineering tasks.
  • Refactor an entire codebase based on a requirement.
  • Run multiple command-line operations.
  • Install packages, configure servers, and set up environments.
  • Use APIs and function calls effectively.
Real-world relevance: This capability is especially important for AI coding agents and software-engineering assistants.

9. Capability 3 — Mathematics

Mathematics can be viewed as a structured form of reasoning in which a sequence of operations leads to a solution.

Levels Mentioned

  • Grade-school mathematics.
  • Competition-level problem solving, including problems requiring creative thinking.
  • Undergraduate mathematics.
  • Research-level mathematical reasoning, including difficult or open-ended problems.

Real-World Relevance

Mathematical capability matters in scientific computing, financial modeling, engineering simulations, and data analysis.

10. Capability 4 — Long Context

Long-context evaluation measures whether a model can effectively use information from very large inputs, potentially containing hundreds of thousands of tokens.

What Is Tested?

  • Retrieve a small fact from a very large context.
  • Find details about a specific person or entity in a large document.
  • Summarize very large inputs.
  • Maintain and use the context of a large codebase when acting as a coding agent.
Important distinction: A stated context-window size does not automatically mean the model uses every part of that context equally well. Long-context evaluation tests actual effectiveness over large inputs.

11. Capability 5 — Vision & Multimodal

This capability moves beyond text and evaluates whether the model can understand information such as images and videos.

Why It Matters

Real-world interaction is multimodal. Examples include asking a model what can be prepared from items visible inside a refrigerator or asking it to identify a particular book from a visual environment.

Evaluation Focus

  • Image understanding
  • Video understanding
  • Cross-modal reasoning

12. Capability 6 — Agentic & Tool Use

Agentic AI requires models that can do more than generate text. They must interact with tools and external environments.

What Is Tested?

  • Web browsing.
  • Structured tool calling.
  • API interaction.
  • Computer or desktop interaction.
  • Executing multi-step actions using available tools.
User Goal
LLM Reasoning
Tool Call
External Action
Result

This capability is increasingly important as agentic applications become more common.

13. Capability 7 — Safety & Alignment

Safety and alignment evaluate whether a model can be trusted to behave responsibly.

Areas Mentioned in the Lecture

  • Harmful-content generation.
  • Resistance to adversarial attacks.
  • Truthfulness versus excessive agreeableness or sycophancy.
  • Cybersecurity-related skills and behavior.
  • Capabilities involving cryptography, reverse engineering, and digital forensics.
Why frontier labs care: Safety has regulatory, reputational, and product consequences. A serious safety incident can create substantial risk for a model provider.

14. Capability 8 — Instruction Following

Instruction following measures whether the model does what the user requested and follows the requested constraints.

Examples

  • If the user requests a bullet list, does the model produce a bullet list?
  • If the user asks for fewer than 200 words, does it respect the limit?
  • If the user requests a friendly response, does it follow that style?
  • If the instruction is ambiguous, does the model ask an appropriate clarifying question?
Product relevance: Poor instruction following directly affects user satisfaction. If the model repeatedly ignores user requirements, users may stop trusting the product.

15. Benchmark vs Custom Evaluation — Final Comparison

DimensionBenchmarkCustom Evaluation
Test designStandardized/sharedDesigned around your task
Primary goalGeneral capability comparisonApplication-specific suitability
ComparabilityHigh across modelsHigh within your controlled setup
ExampleMMLU, SWE-benchHistorical labeled application data
Best question answered“How good is this model generally?”“Which model is best for our workload?”

16. Interview Questions

Q1. What is a Model Eval?

A systematic process for measuring an underlying model's capabilities, behavior, reliability, and operational characteristics under controlled conditions.

Q2. Why does an AI Engineer need Model Evals?

To compare models, select the right model, assess new model releases, evaluate safety/reliability, and support decisions such as API usage versus self-hosting.

Q3. What are the four steps of a model evaluation?

Select the capability → bring a test → run under a fixed protocol → score and interpret.

Q4. What is a benchmark?

A standardized shared test used to compare models on common ground.

Q5. Why are custom evals needed if benchmarks already exist?

Because public benchmarks measure generic capabilities. A custom eval measures performance on the actual task, data, constraints, cost, and latency requirements of your application.

Q6. Name the eight core capability categories discussed.

Knowledge & Reasoning; Coding & Software Engineering; Mathematics; Long Context; Vision & Multimodal; Agentic & Tool Use; Safety & Alignment; Instruction Following.

Q7. Why is a fixed protocol important?

It makes evaluation repeatable and ensures that models being compared are tested under comparable conditions.

Q8. Why can a smaller model be better than a stronger model?

If it provides sufficiently similar task performance while reducing cost and/or other operational burdens, it can be a better value proposition for the workload.

17. Quick Revision

  • Model Evals directly test LLM capabilities; Application Evals test LLM-powered applications.
  • AI Engineers need Model Evals for model comparison, selection, release tracking, safety assessment, and deployment strategy.
  • A Model Eval follows four steps: Capability → Test → Fixed Protocol → Score & Interpret.
  • Two major test approaches are standardized benchmarks and custom evaluations.
  • Benchmarks provide common-ground comparison; custom evals provide application-specific evidence.
  • The lecture's Zomato example demonstrates why benchmark leadership does not automatically mean best application fit.
  • The eight core capabilities are: Knowledge & Reasoning, Coding & SWE, Mathematics, Long Context, Vision & Multimodal, Agentic & Tool Use, Safety & Alignment, Instruction Following.
  • There is no single universal “LLM IQ” score that captures every capability.
  • Long context should be evaluated by actual information use, not only by the advertised context-window size.
  • Instruction following is important because it directly affects user experience and product trust.
KEY TAKEAWAY

Think Like an AI Engineer

Do not ask only, “Which model is strongest?” Ask: “Which model performs well on the capabilities my application actually needs, under the cost and operational constraints I have?” Benchmarks help you understand general capability; custom evaluations help you make the final application decision.

AI Engineering Notes • LLM Evaluation Series • Page 05