Benchmark	Focus	Links
MMLU	Evaluates broad knowledge and reasoning across many academic and professional subjects.	Paper
Humanity's Last Exam	Evaluates frontier models on expert-level academic reasoning and knowledge across more than 100 subjects. It is a multimodal benchmark with 2,500 challenging closed-ended questions designed to avoid benchmark saturation.	Website Paper Dataset
HellaSwag	Measures commonsense reasoning by selecting the most plausible ending for a grounded scenario.	Project Page Paper
GSM8K	Evaluates grade-school math word-problem solving and multi-step numerical reasoning.	Paper Dataset
TruthfulQA	Tests factual truthfulness and resistance to generating common misconceptions.	Paper Repository
ARC (Challenge)	Assesses science question answering and reasoning over grade-school level knowledge.	Paper Dataset
BIG-bench Hard (BBH)	Stress-tests difficult reasoning tasks selected from BIG-bench for stronger discrimination.	Paper Repository
MATH	Evaluates advanced mathematical reasoning on competition-style problems.	Paper Repository
BIG-bench	Evaluates broad and diverse emergent capabilities across many challenging language tasks.	Repository Paper
MT-Bench	Measures multi-turn instruction-following quality and chat performance, often with LLM-as-a-judge.	Paper Implementation
GPQA	Evaluates graduate-level domain knowledge and difficult reasoning in science-focused questions.	Paper
DROP	Tests reading comprehension with discrete reasoning over passages (counts, arithmetic, and comparisons).	Dataset Paper
SQuAD	Evaluates extractive question answering and reading comprehension on Wikipedia passages.	Dataset Paper
HumanEval	Measures code generation quality by checking functional correctness of generated Python functions.	Repository
SWE-bench	Tests software engineering ability by solving real GitHub issues in open-source repositories.	Leaderboard Paper
DeepSWE	Evaluates frontier coding agents on original, long-horizon software engineering tasks. It is designed to differentiate top models with contamination-free tasks, diverse repositories and languages, and behavior-based verification.	Website Methodology Blog Repository
ProgramBench	Evaluates whether language models and coding agents can rebuild programs from scratch given only a compiled binary and its documentation, reproducing the original program's behavior with a complete codebase.	Website Paper Repository
MCP-Atlas	Evaluates AI agents' tool-use capabilities on real-world tasks across 36 Model Context Protocol servers, using a standardized Dockerized environment and LLM-as-judge scoring with pass rate and coverage metrics.	Leaderboard Paper Repository
Tool-Decathlon	Evaluates language agents' general tool use in diverse, realistic, and long-horizon task environments. Toolathlon features 600+ tools based on real-world software environments and tasks that require extended sequences of tool calls.	Website Paper Repository
Terminal-Bench	Evaluates AI agents in terminal environments, measuring terminal mastery across long-horizon tasks in software engineering, machine learning, security, data science, and related domains.	Website Leaderboard Terminal-Bench 2.0
NL2RepoBench	Evaluates LLMs and coding agents on long-horizon 0-to-1 software tasks that require generating a complete, runnable code repository from scratch. The benchmark consists of 104 tasks, each with its own testing environment.	Repository README
MMMU	Assesses multimodal understanding and reasoning with image-plus-text tasks across disciplines.	Project Page Paper
HELM	Provides holistic evaluation across scenarios and metrics such as accuracy, calibration, and robustness.	Dashboard Paper
LMArena	LMArena (formerly known as Chatbot Arena) is a popular, crowdsourced AI evaluation platform where users compare different artificial intelligence models through blind, head-to-head battles.	Website YouTube tutorials