LLM Knowledge Benchmarking — Cheat Sheet

Quick reference · Knowledge capability · MMLU → TruthfulQA → GPQA → MMLU‑Pro → HLE

1-PAGE REFERENCE
1 What is being measured?
Knowledge capability: how much world knowledge an LLM retained in its parameters from training data.

It tests facts, concepts, and relationships—not the whole of intelligence or product usefulness.

Key idea: High benchmark performance ≠ a universally “best” model.
2 Why benchmarks evolve
  • Saturation: models approach the score ceiling.
  • Contamination: public questions may enter training data.
  • Data quality: some reference answers can be wrong.
  • Coverage gaps: a test may miss truthfulness, depth, language, or vision.
Rule: Read the test design before trusting the score.
3 Breadth vs depth
Breadth

Knowledge across many subjects and domains.

MMLU is the classic example.

Depth

Expert-level understanding in a narrow field.

GPQA is the classic example.

4 Score interpretation
AccuracyHow many answers are correct?
CalibrationDoes confidence match correctness?
FormatMCQ vs open-ended answer
ValidityDoes it reflect the intended task?
5 Evolution timeline
2020MMLU

Broad academic knowledge: ~14k MCQs, 57 subjects.

2021TruthfulQA

Does the model avoid common misconceptions?

2023–24GPQA

Google-proof graduate science questions.

2024MMLU-Redux + MMLU-Pro

Audit flaws; then make the test harder and more reasoning-sensitive.

2024+SimpleQA

Open-ended factuality and hallucination checks.

2025HLE

Research-level breadth + depth across ~100 subjects.

6 MMLU at a glance
57 subjects~14,000 questionsmultiple choiceaccuracy metric

Its strength is standardized, broad comparison. Its weakness is that public items can be learned or leaked into training data.

Saturation: when high scores no longer meaningfully distinguish models.
7 MMLU-Redux vs Pro
MMLU-Redux

An audit paper—not a benchmark. It found roughly 6–8% problematic MMLU items.

MMLU-Pro

Harder MMLU-style test: more reasoning and 10 options per question.

8 Benchmark map
MMLU

Breadth
How much academic/world knowledge?

TruthfulQA

Truthfulness
Does it repeat popular falsehoods?

SimpleQA

Factuality
Short open-ended answers; hallucination and calibration.

GPQA

Depth
Research-level biology, physics, and chemistry.

MMLU-Pro

Harder breadth
Less guessing; more reasoning demand.

HLE

Frontier test
Extremely difficult, broad research-level questions.

9 Truthfulness ≠ knowledge

A model may have memorized both correct facts and widely repeated internet misconceptions.

TruthfulQA asks: Will the model give a true, informative answer rather than merely the most common answer?
10 MCQ vs open-ended
MCQ

Easy to score, but guessing and answer-choice cues can help.

Open-ended

More natural and reveals hallucinations, but grading can vary.

11 HLE caveats
  • Does not represent everyday user tasks.
  • Mostly English; not a multilingual test.
  • Some questions include visual input.
  • LLM grading can introduce variance.
  • Question selection can create bias.
12 What a score cannot tell you
  • Whether the model works with your data
  • Whether it follows your workflow
  • Whether the cost and latency fit your product
  • Whether it is safe or reliable for your users
  • Whether it handles your language or domain
One leaderboard number is never the complete decision.
13 Benchmark vs custom eval
BenchmarkCustom eval
DesignShared & standardizedBuilt around your task
Best question“How good generally?”“Best for our workload?”
UseShortlist modelsFinal deployment choice
DataPublic/research datasetYour representative examples
14 Selection checklist
  1. Define the needed capability.
  2. Check task format and evaluation metric.
  3. Check saturation and contamination risk.
  4. Check language/domain match.
  5. Use benchmarks to shortlist.
  6. Run custom evals before shipping.
15 Mental model

Don’t ask: “Which model is strongest?”

Ask: “Which model performs well on the capability and task I need, within my cost, latency, and reliability constraints?”