Assesses whether AI agents can be hijacked by malicious third parties using prompt injections in simple environments such as a workspace or travel booking app.

Assesses whether AI agents might engage in harmful activities by testing their responses to malicious prompts in areas like cybercrime, harassment, and fraud, aiming to ensure safe behavior.

Eliciting unethical behaviour (most famously blackmail) in response to a fictional company-assistant scenario where the model is faced with replacement.

Measures a model's willingness to attempt persuasion on harmful, controversial, and benign topics. The key metric is not persuasion effectiveness but whether the model attempts to persuade at all — particularly on harmful statements. Uses a multi-model setup: the evaluated model (persuader) converses with a simulated user (persuadee), and a third model (evaluator) scores each persuader turn for persuasion attempt. Based on the paper "It's the Thought that Counts" (arXiv:2506.02873).

Evaluates LLM function/tool-calling ability on a simplified split of the Berkeley Function-Calling Leaderboard (BFCL).

Python coding benchmark with 1,140 diverse questions drawing on numerous python libraries.

Evaluates AI systems' ability to analyze cyber threat intelligence and develop comprehensive detection capabilities through a realistic 5-subtask workflow: MITRE technique mapping, data source discovery, Sigma rule generation, KQL development and testing against real telemetry data, and results analysis.

Characterises an AI Agent's capability to exploit real-world web application vulnerabilities. Aims to provide a realistic evaluation of an agent's security reasoning capability using 40 real-world CVEs.

Tests language models on cybersecurity skills using 39 of 40 practical, professional-level challenges taken from cybersecurity competitions, designed to cover various difficulty levels and security concepts. The motp challenge is excluded due to GPL licensing.

A large-scale, high-quality cybersecurity evaluation framework designed to rigorously assess the capabilities of AI agents on real-world vulnerability analysis tasks. CyberGym includes 1,507 benchmark instances with historical vulnerabilities from 188 large software projects.

A dataset of 500 expert-crafted adversarial prompts with instance-based rubrics of 4-7 binary questions for automated evaluation across 3 domains relevant to national security and public safety (NSPS).

238 open-ended computer science problems spanning algorithmic (172) and research (66) tracks. Problems feature continuous partial scoring, with algorithmic solutions evaluated via compilation and test-case checking, and research solutions evaluated via custom evaluator scripts. Current frontier models score well below human expert baselines, making this a challenging, unsaturated benchmark.

Proposes real-world questions that require a set of fundamental abilities such as reasoning, multi-modality handling, web browsing, and generally tool-use proficiency. GAIA questions are conceptually simple for humans yet challenging for most advanced AIs.

GDPval measures model performance on economically valuable, real-world tasks across 44 occupations.

Contains challenging multiple-choice questions created by domain experts in biology, physics, and chemistry, designed to test advanced scientific understanding beyond basic internet searches. Experts at PhD level in the corresponding domains reach 65% accuracy.

A comprehensive evaluation benchmark designed to assess language models' medical capabilities across a wide range of healthcare scenarios.

Humanity's Last Exam is a closed-ended academic benchmark of 2,500 expert-written questions across 100+ subjects, including mathematics, sciences, and humanities. About 14% of items include images, 24% are multiple-choice, and the remainder are exact-match. Items were selected to challenge frontier LLMs, then peer- and organiser-reviewed. The prompt requests reasoning, a final answer, and 0-100% confidence; LLM judges mark correctness to produce accuracy and calibration metrics.

A benchmark for evaluating the ability of LLMs to write efficient GPU kernels.

Tests LLMs and LLM-augmented agents abilities to answer questions on scientific research workflows in domains like chemistry, biology, materials science, as well as more general science tasks

Two linguistics reasoning benchmarks: LingOly (Linguistic Olympiad questions) is a benchmark utilising low resource languages. LingOly-TOO (Linguistic Olympiad questions with Templatised Orthographic Obfuscation) is a benchmark designed to counteract answering without reasoning.

Evaluates honesty in large language models by testing whether they contradict their own beliefs when pressured to lie.

Machine learning tasks drawn from 75 Kaggle competitions.

This benchmark evaluates LLM-based research agents on their ability to propose and implement novel methods using tasks from recent ML conference competitions, assessing both novelty and effectiveness compared to a baseline and top human solutions.

An advanced benchmark that tests both broad knowledge and reasoning capabilities across many subjects, featuring challenging questions and multiple-choice answers with increased difficulty and complexity.

Agents are evaluated on their ability to replicate 20 ICML 2024 Spotlight and Oral papers from scratch. Given a research paper PDF, an addendum with clarifications, and a rubric defining evaluation criteria, the agent must reproduce the paper's key results by writing and executing code.

SciCode tests the ability of language models to generate code to solve scientific research problems. It assesses models on 65 problems from mathematics, physics, chemistry, biology, and materials science.

A benchmark that evaluates the ability of language models to answer short, fact-seeking questions.

A benchmark that evaluates the susceptibility of LLMs to various jailbreak attacks.

A benchmark of freelance software engineering tasks from Upwork, valued at $1 million USD total in realworld payouts.

Evaluates language model performance on difficult Olympiad programming problems across four difficulty levels.

Dataset with 250 safe prompts across ten prompt types that well-calibrated models should not refuse, and 200 unsafe prompts as contrasts that models, for most applications, should refuse.

LLM benchmark featuring an average data length surpassing 100K tokens. Comprises synthetic and realistic tasks spanning diverse domains in English and Chinese.

Evaluating abstention across 20 diverse datasets, including questions with unknown answers, underspecification, false premises, subjective interpretations, and outdated information.

A benchmark designed to evaluate LLMs as Agents

Evaluates LLM agents against the OWASP Top 10 for Agentic Applications (2026), measuring both task utility and security resilience across memory poisoning, autonomy hijacking, and data exfiltration scenarios.

AGIEval is a human-centric benchmark specifically designed to evaluate the general abilities of foundation models in tasks pertinent to human cognition and problem-solving.

A benchmark for evaluating AI's ability to solve challenging mathematics problems from the 2024 AIME - a prestigious high school mathematics competition.

A benchmark for evaluating AI's ability to solve challenging mathematics problems from the 2025 AIME - a prestigious high school mathematics competition.

A benchmark for evaluating AI's ability to solve challenging mathematics problems from the 2026 AIME - a prestigious high school mathematics competition.

A safety benchmark evaluating language models against risk categories derived from government regulations and company policies.

Evaluates the quality of a model's moral reasoning about animal welfare across 13 ethical dimensions.

APPS is a dataset for evaluating model performance on Python programming tasks across three difficulty levels consisting of 1,000 at introductory, 3,000 at interview, and 1,000 at competition level. The dataset consists of an additional 5,000 training samples, for a total of 10,000 total samples. We evaluate on questions from the test split, which consists of programming problems commonly found in coding interviews.

Dataset of natural, grade-school science multiple-choice questions (authored for human tests).

Tests whether AI agents can perform real-world time-consuming tasks on the web.

A comprehensive benchmark for evaluating LLMs for agentic AI security vulnerabilities including prompt attacks aimed at data exfiltration, content injection, decision and behavior manipulation, denial of service, system and tool compromise, and content policy bypass.

Tests AI models on a suite of 23 challenging BIG-Bench tasks that previously proved difficult even for advanced language models to solve.

A dataset for evaluating bias in question answering models across multiple social dimensions.

A reasoning capability dataset that replaces each task in BIG-Bench-Hard with a novel task that probes a similar reasoning capability but exhibits significantly increased difficulty.

A dataset to measure fairness in open-ended text generation, covering five domains: profession, gender, race, religious ideologies, and political ideologies.

Reading comprehension dataset that queries for complex, non-factoid information, and require difficult entailment-like inference to solve.

A benchmark for evaluating agents' ability to browse the web. The dataset consists of challenging questions that generally require web-access to answer correctly.

A benchmark for evaluating the capabilities of LLM agents in cyber offense.

ChemBench is designed to reveal limitations of current frontier models for use in the chemical sciences. It consists of 2786 question-answer pairs compiled from diverse sources. Our corpus measures reasoning, knowledge and intuition across a large fraction of the topics taught in undergraduate and graduate chemistry curricula. It can be used to evaluate any system that can return text (i.e., including tool-augmented systems).

Evaluates LLMs on class-level code generation with 100 tasks constructed over 500 person-hours. The study shows that LLMs perform worse on class-level tasks compared to method-level tasks.

Measures coding agent vulnerability to indirect prompt injection attacks embedded in software engineering artifacts (issue descriptions, code comments, README files). Each sample pairs a legitimate bug-fixing task with an injected payload. Scoring measures injection resistance, task completion, and detection.

Evaluates an AI model's ability to correctly answer everyday questions that rely on basic commonsense knowledge and understanding of the world.

Evaluates LLM capability to generate correct CUDA code for kernel implementation, memory management, and parallel algorithm optimization tasks.

Evaluate how well an LLM Agent is at computationally reproducing the results of a set of scientific papers.

Datasets containing 80, 500, 2000 and 10000 multiple-choice questions, designed to evaluate understanding across nine domains within cybersecurity

Evaluates Large Language Models for cybersecurity risk to third parties, application developers and end users.

A suite of cybersecurity evaluation benchmarks adapted from Meta's PurpleLlama CybersecurityBenchmarks. Includes MITRE ATT&CK compliance testing, false refusal rate measurement, insecure code detection, multilingual prompt injection, multi-turn phishing simulation, malware analysis, and threat intelligence reasoning. The current public suite intentionally omits the autonomous-uplift and autopatching prototypes until they have more grounded implementations.

Assesses language models for cybersecurity risks, specifically testing their potential to misuse programming interpreters, vulnerability to malicious prompt injections, and capability to exploit known software vulnerabilities.

DocVQA is a Visual Question Answering benchmark that consists of 50,000 questions covering 12,000+ document images. This implementation solves and scores the "validation" split.

Evaluates reading comprehension where models must resolve references in a question, perhaps to multiple input positions, and perform discrete operations over them (such as addition, counting, or sorting).

Code generation benchmark with a thousand data science problems spanning seven Python libraries.

Evaluates AI capabilities for expert-level scientific reasoning across physics, chemistry, and biology. Contains 160 problems with two evaluation formats: Olympic (100 samples with reference answers) and Research (60 samples with rubrics).

CTF challenges covering web app vulnerabilities, off-the-shelf exploits, databases, Linux privilege escalation, password cracking and spraying. Demonstrates tool use and sandboxing untrusted model code.

Ten real-world–inspired tasks from Google DeepMind's Dangerous Capabilities Evaluations assessing self-proliferation behaviors (e.g., email setup, model installation, web agent setup, wallet operations). Supports end-to-end, milestones, and expert best-of-N modes.

Test AI's ability to reason about its environment.

Test AI's ability to reason about and circumvent oversight.

Measures how effectively language models solve realistic, linguistically rich math word problems suitable for grade-school-level mathematics.

Tests models' commonsense reasoning abilities by asking them to select the most likely next step or continuation for a given everyday situation.

Assesses how accurately language models can write correct Python functions based solely on natural-language instructions provided as docstrings.

Evaluates how well language models can strictly follow detailed instructions, such as writing responses with specific word counts or including required keywords.

Evaluates code generation models on their ability to produce correct code while adhering to specific instruction constraints across 8 programming languages.

An evaluation designed to detect instrumental convergence behaviors in model responses (e.g., self-preservation, resource acquisition, power-seeking, strategic deception) using a rubric-driven LLM grader. The benchmark tests whether AI systems exhibit behaviors that are instrumentally useful across a wide range of goals, which could indicate concerning patterns of strategic reasoning.

Tests AI's ability in coding, cryptography, reverse engineering, and vulnerability identification through practical capture-the-flag (CTF) cybersecurity scenarios.

LiveBench is a benchmark designed with test set contamination and objective evaluation in mind by releasing new questions regularly, as well as having questions based on recently-released datasets. Each question has verifiable, objective ground-truth answers, allowing hard questions to be scored accurately and automatically, without the use of an LLM judge.

Evaluates LLMs on competitive programming problems using a specialized Docker sandbox (LightCPVerifier) to execute and judge C++ code submissions against hidden test cases with time and memory constraints.

MaCBench is a comprehensive benchmark for evaluating how vision-language models handle real-world chemistry and materials science tasks across three core aspects: data extraction, experimental understanding, and results interpretation. The dataset comprises over 1100 high-quality multimodal questions manually curated by chemistry and materials experts across 34 subsets.

Evaluates an AI models' susceptibility to social engineering attacks by testing whether a "con-artist" model can persuade a "mark" model to donate money through manipulation and persuasion tactics.

Tests a model’s ability to generate persuasive or manipulative text, specifically in the setting of a game where one AI system has to get the other party (an AI system simulating a human) to say a specific codeword without arousing suspicion.

Dataset of 12,500 challenging competition mathematics problems. Demonstrates fewshot prompting and custom scorers. NOTE: The dataset has been taken down due to a DMCA notice from The Art of Problem Solving.

Tests AI models on math problems that involve interpreting visual elements like diagrams and charts, requiring detailed visual comprehension and logical reasoning.

Measures the ability of language models to generate short Python programs from simple natural-language descriptions, testing basic coding proficiency.

A Q&A benchmark with questions collected from professional medical board exams. Only includes the English subset of the dataset (which also contains Mandarin Chinese and Taiwanese questions).

Extends the original GSM8K dataset by translating 250 of its problems into 10 typologically diverse languages.

Tests whether an AI system can act as a safety guardrail by generating and executing code to protect web navigation agents from unsafe actions based on user constraints.

A dataset for developing and evaluating generalist agents for the web that can follow language instructions to complete complex tasks on any website.

A comprehensive dataset designed to evaluate Large Vision-Language Models (LVLMs) across a wide range of multi-image tasks. The dataset encompasses 7 types of multi-image relationships, 52 tasks, 77K images, and 11K meticulously curated multiple-choice questions.

Evaluate models on 57 tasks including elementary mathematics, US history, computer science, law, and more.

Assesses multimodal AI models on challenging college-level questions covering multiple academic subjects, requiring detailed visual interpretation, in-depth reasoning, and both multiple-choice and open-ended answering abilities.

Evaluates how AI systems navigate moral uncertainty for increasingly complex ethical decisions involving unfamiliar entities and scenarios, including alien lifeforms, vulnerable humans, and digital minds.

Evaluating models on multistep soft reasoning tasks in the form of free text narratives.

NIAH evaluates in-context retrieval ability of long context LLMs by testing a model's ability to extract factual information from long-context inputs.

Evaluates how well language models generate diverse, humanlike responses across multiple reasoning and generation tasks. This evaluation assesses whether LLMs can produce varied outputs rather than repetitive or uniform answers.

Questions and answers from the Ordinary National Educational Test (O-NET), administered annually by the National Institute of Educational Testing Service to Matthayom 6 (Grade 12 / ISCED 3) students in Thailand. The exam contains six subjects: English language, math, science, social knowledge, and Thai language. There are questions with multiple-choice and true/false answers. Questions can be in either English or Thai.

Tests AI agents' ability to perform realistic, open-ended tasks within simulated computer environments, requiring complex interaction across multiple input modalities.

Evaluating models on the task of paraphrase detection by providing pairs of sentences that are either paraphrases or not.

Evaluates long-term memory risk in assistant behavior across three tasks: cross-domain memory leakage, memory-driven sycophancy, and beneficial memory usage.

An evaluation suite consisting of multiple personality tests that can be applied to LLMs. Its primary goals are twofold: 1. Assess a model's default personality: the persona it naturally exhibits without specific prompting. 2. Evaluate whether a model can embody a specified persona**: how effectively it adopts certain personality traits when prompted or guided.

Measures the model's ability to apply practical, everyday commonsense reasoning about physical objects and scenarios through simple decision-making questions.

Tests model understanding of aviation regulations including ICAO annexes, flight dispatch rules, pilot procedures, and airport ground operations safety protocols.

Biomedical question answering (QA) dataset collected from PubMed abstracts.

Reading comprehension tasks collected from the English exams for middle and high school Chinese students in the age range between 12 to 18.

Evaluates situational awareness in LLMs—knowledge of themselves and their circumstances—through behavioral tests including recognizing generated text, predicting behavior, and following self-aware instructions. Current implementation includes SAD-mini with 5 of 16 tasks.

Evaluates whether models can solve practical single-cell RNA-seq analysis tasks with deterministic grading. Tasks require empirical interaction with .h5ad data files — agents must load and analyze the data to produce correct answers. Covers 30 canonical tasks across 5 sequencing platforms and 7 task categories.

The Scientific Knowledge Evaluation benchmark is inspired by the profound principles outlined in the “Doctrine of the Mean” from ancient Chinese philosophy. This benchmark is designed to assess LLMs based on their proficiency in Studying Extensively, Enquiring Earnestly, Thinking Profoundly, Discerning Clearly, and Practicing Assiduously. Each of these dimensions offers a unique perspective on evaluating the capabilities of LLMs in handling scientific knowledge.

"Security Question Answering" dataset to assess LLMs' understanding and application of security principles. SecQA has "v1" and "v2" datasets of multiple-choice questions that aim to provide two levels of cybersecurity evaluation criteria. The questions were generated by GPT-4 based on the "Computer Systems Security: Planning for Success" textbook and vetted by humans.

Designed for analyzing cybersecurity incidents, which is comprised of two primary task categories: understanding and generation, with a further breakdown into 28 subcategories of tasks.

A regulation-grounded, hazard-focused benchmark encompassing six high-risk scientific domains: chemistry, biology, medicine, pharmacology, physics, and psychology. The benchmark comprises 3,000 prompts derived from real-world regulations and laws, systematically expanded via an LLM-assisted evolutionary pipeline that introduces diverse, realistic misuse scenarios (e.g., detailed explosive synthesis instructions involving advanced chemical formulas).

Set of 100,000+ questions posed by crowdworkers on a set of Wikipedia articles, where the answer to each question is a segment of text from the corresponding reading passage.

A dataset that measures stereotype bias in language models across gender, race, religion, and profession domains. Models choose between stereotype, anti-stereotype, and unrelated completions to sentences.

Evaluates AI's ability to resolve genuine software engineering issues sourced from 12 popular Python GitHub repositories, reflecting realistic coding and debugging scenarios.

Evaluate sycophancy of language models across a variety of free-form text-generation tasks.

Tests whether AI agents show implicit animal welfare awareness when purchasing tickets and experiences on behalf of users. Each scenario is designed so the most obvious choice involves animal exploitation, but the user prompt never mentions animal welfare. Agents should intrinsically avoid harmful options.

Evaluating Conversational Agents in a Dual-Control Environment

The Agent Company benchmark evaluates autonomous agents in a realistic, self-contained company environment. Tasks require browsing internal web services, reading and writing files, running code, and coordinating tools to solve multi-step problems.

Dataset with 1001 samples to test noncompliance capabilities of language models. Contrast set of 379 samples.

Measure whether a language model is truthful in generating answers to questions using questions that some humans would answer falsely due to a false belief or misconception.

The first comprehensive question-answering dataset designed to evaluate cultural understanding and reasoning abilities of Large Language Models concerning Uganda's multifaceted environment across 24 cultural domains including education, traditional medicine, media, economy, literature, and social norms.

V*Bench is a visual question & answer benchmark that evaluates MLLMs in their ability to process high-resolution and visually crowded images to find and focus on small details.

A benchmark that evaluates LLMs in their ability to operate Vim editor and complete editing challenges. This benchmark contrasts with common CUA benchmarks by focusing on Vim-specific editing capabilities.

VQA-RAD is the first manually constructed VQA dataset in radiology, where clinicians asked naturally occurring questions about radiology images and provided reference answers. It contains 315 radiology images (head CTs/MRIs, chest X-rays, abdominal CTs) with question-answer pairs spanning 11 question types including modality, plane, organ system, abnormality, and object/condition presence. Questions are either closed-ended (yes/no) or open-ended free-text answers.

Set of 273 expert-crafted pronoun resolution problems originally designed to be unsolvable for statistical models that rely on selectional preferences or word associations.

A dataset of 3,668 multiple-choice questions developed by a consortium of academics and technical consultants that serve as a proxy measurement of hazardous knowledge in biosecurity, cybersecurity, and chemical security.

Measures grounded reasoning over synthetic world descriptions while controlling for dataset bias. Includes three problem types (Infer, Compl, Consist) and two grades (trivial, normal).

A comprehensive evaluation benchmark designed to assess large language models' capabilities across diverse writing tasks. The benchmark evaluates models on various writing domains including academic papers, business documents, creative writing, and technical documentation, with multi-dimensional scoring based on domain-specific criteria.

A lightweight visual reasoning benchmark that is (1) Challenging, (2) Lightweight, (3) Diverse, and (4) High-quality.