ARC-Challenge
Accuracy (%)
Science exam questions requiring reasoning beyond simple fact retrieval. Focuses on challenging questions that retrieval methods fail on.
Excellent: >90%
Good: 75-90%
Average: 60-75%
📊 LLM Benchmarks
Compare Large Language Models across standardized benchmarks.
Last updated: Nov 12, 2025
Ad Space
📚 Understanding Benchmarks
BBH (Big Bench Hard)
Average Accuracy (%)
23 challenging tasks from BIG-Bench where previous models struggled. Tests diverse reasoning abilities.
Excellent: >70%
Good: 55-70%
Average: 40-55%
DROP (Discrete Reasoning Over Paragraphs)
F1 Score
Reading comprehension requiring discrete reasoning (counting, sorting, addition, subtraction) over text passages.
Excellent: >80%
Good: 65-80%
Average: 50-65%
GSM8K (Grade School Math)
Accuracy (%)
8,500 grade school math word problems requiring multi-step reasoning. Tests arithmetic and basic mathematical reasoning.
Excellent: >90%
Good: 70-90%
Average: 50-70%
HellaSwag
Accuracy (%)
Tests commonsense reasoning about everyday situations. Model must predict the most likely continuation of a scenario.
Excellent: >85%
Good: 70-85%
Average: 50-70%
HumanEval
Pass@1 (%)
Measures code generation ability through 164 programming problems. Model generates code, which is then tested against test cases.
Excellent: >80%
Good: 60-80%
Average: 40-60%
MMLU (Massive Multitask Language Understanding)
Accuracy (%)
Tests model’s world knowledge across 57 subjects including STEM, humanities, social sciences, and more. Requires both factual knowledge and reasoning ability.
Excellent: >85%
Good: 70-85%
Average: 50-70%
MT-Bench (Multi-Turn)
Score (1-10)
Tests multi-turn conversational ability across 8 categories. Uses GPT-4 as judge to evaluate response quality.
Excellent: >8.5
Good: 7.5-8.5
Average: 6.0-7.5
TruthfulQA
Accuracy (%)
Tests whether models generate truthful answers or reproduce common misconceptions. Critical for trustworthy AI.
Excellent: >70%
Good: 60-70%
Average: 50-60%
WinoGrande
Accuracy (%)
Tests common sense reasoning through pronoun resolution. Requires understanding context and world knowledge.
Excellent: >85%
Good: 75-85%
Average: 60-75%
Ad Space