Allen Institute for AI: AI benchmarks & research

AI2D is a dataset of 4,903 illustrative diagrams from grade school natural sciences (such as food webs, human physiology, and life cycles) with over 15,000 multiple choice questions and answers. The benchmark evaluates diagram understanding and visual reasoning capabilities, requiring models to interpret diagrammatic elements, relationships, and structure to answer questions about scientific concepts represented in visual form.

A dataset of 7,787 genuine grade-school level, multiple-choice science questions assembled to encourage research in advanced question-answering. The dataset is partitioned into a Challenge Set and Easy Set, where the Challenge Set contains only questions answered incorrectly by both retrieval-based and word co-occurrence algorithms. Covers multiple scientific domains including biology, physics, earth science, and chemistry, requiring scientific reasoning, causal understanding, and conceptual knowledge beyond simple fact retrieval. Includes a supporting corpus of over 14 million science sentences.

The AI2 Reasoning Challenge (ARC) Challenge Set is a multiple-choice question-answering benchmark containing grade-school level science questions that require advanced reasoning capabilities. ARC-C specifically contains questions that were answered incorrectly by both retrieval-based and word co-occurrence algorithms, making it a particularly challenging subset designed to test commonsense reasoning abilities in AI systems.

ARC-E (AI2 Reasoning Challenge - Easy Set) is a subset of grade-school level, multiple-choice science questions that requires knowledge and reasoning capabilities. Part of the AI2 Reasoning Challenge dataset containing 5,197 questions that test scientific reasoning and factual knowledge. The Easy Set contains questions that are answerable by retrieval-based and word co-occurrence algorithms, making them more accessible than the Challenge Set.

BLINK: Multimodal Large Language Models Can See but Not Perceive. A benchmark for multimodal language models focusing on core visual perception abilities. Reformats 14 classic computer vision tasks into 3,807 multiple-choice questions paired with single or multiple images and visual prompting. Tasks include relative depth estimation, visual correspondence, forensics detection, multi-view reasoning, counting, object localization, and spatial reasoning that humans can solve 'within a blink'.

DROP (Discrete Reasoning Over Paragraphs) is a reading comprehension benchmark requiring discrete reasoning over paragraph content. It contains crowdsourced, adversarially-created questions that require resolving references and performing discrete operations like addition, counting, or sorting, demanding comprehensive paragraph understanding beyond paraphrase-and-entity-typing shortcuts.

A challenging commonsense natural language inference dataset that uses Adversarial Filtering to create questions trivial for humans (>95% accuracy) but difficult for state-of-the-art models, requiring completion of sentence endings based on physical situations and everyday activities

Instruction Following Benchmark evaluating model's ability to follow complex instructions

OpenBookQA is a question-answering dataset modeled after open book exams for assessing human understanding. It contains 5,957 multiple-choice elementary-level science questions that probe understanding of 1,326 core science facts and their application to novel situations, requiring combination of open book facts with broad common knowledge through multi-hop reasoning.

PIQA (Physical Interaction: Question Answering) is a benchmark dataset for physical commonsense reasoning in natural language. It tests AI systems' ability to answer questions requiring physical world knowledge through multiple choice questions with everyday situations, focusing on atypical solutions inspired by instructables.com. The dataset contains 21,000 multiple choice questions where models must choose the most appropriate solution for physical interactions.

PointArena is a comprehensive platform for evaluating multimodal pointing across diverse reasoning scenarios. It includes Point-Bench, a curated dataset of ~1,000 pointing tasks across five categories: Spatial (positional references), Affordance (functional part identification), Counting (attribute-based grouping), Steerable (relative pointing), and Reasoning (open-ended visual inference). The benchmark evaluates language-guided pointing capabilities in vision-language models.

PopQA is an entity-centric open-domain question-answering dataset consisting of 14,000 QA pairs designed to evaluate language models' ability to memorize and recall factual knowledge across entities with varying popularity levels. The dataset probes both parametric memory (stored in model parameters) and non-parametric memory effectiveness, with questions covering 16 diverse relationship types from Wikidata converted to natural language using templates. Created by sampling knowledge triples from Wikidata and converting them to natural language questions, focusing on long-tail entities to understand LMs' strengths and limitations in memorizing factual knowledge.

QASPER is a dataset of 5,049 information-seeking questions and answers anchored in 1,585 NLP research papers. Questions are written by NLP practitioners who read only titles and abstracts, while answers require understanding the full paper text and provide supporting evidence. The dataset challenges models with complex reasoning across document sections for academic document question answering. Each question seeks information present in the full text and is answered by a separate set of NLP practitioners who also provide supporting evidence to answers.

ScienceQA is the first large-scale multimodal science question answering benchmark with 21,208 multiple-choice questions covering 3 subjects (natural science, language science, social science), 26 topics, 127 categories, and 379 skills. The benchmark includes both text and image modalities, featuring detailed explanations and Chain-of-Thought reasoning to diagnose multi-hop reasoning ability.

ScienceQA Visual is a multimodal science question answering benchmark consisting of 21,208 multiple-choice questions from elementary and high school science curricula. The dataset covers 3 subjects (natural science, language science, social science), 26 topics, 127 categories, and 379 skills. 48.7% of questions include image context requiring multimodal reasoning. Questions are annotated with lectures (83.9%) and explanations (90.5%) to support chain-of-thought reasoning for science question answering.

The first large-scale benchmark for commonsense reasoning about social situations. Contains 38,000 multiple choice questions probing emotional and social intelligence in everyday situations, testing commonsense understanding of social interactions and theory of mind reasoning about the implied emotions and behavior of others.

WildBench is an automated evaluation framework that benchmarks large language models using 1,024 challenging, real-world tasks selected from over one million human-chatbot conversation logs. It introduces two evaluation metrics (WB-Reward and WB-Score) that achieve high correlation with human preferences and uses task-specific checklists for systematic evaluation.

WinoGrande: An Adversarial Winograd Schema Challenge at Scale. A large-scale dataset of 44,000 pronoun resolution problems designed to test machine commonsense reasoning. Uses adversarial filtering to reduce spurious biases and provides a more robust evaluation of whether AI systems truly understand commonsense or exploit statistical shortcuts. Current best AI methods achieve 59.4-79.1% accuracy, significantly below human performance of 94.0%.

ZebraLogic is an evaluation framework for assessing large language models' logical reasoning capabilities through logic grid puzzles derived from constraint satisfaction problems (CSPs). The benchmark consists of 1,000 programmatically generated puzzles with controllable and quantifiable complexity, revealing a 'curse of complexity' where model accuracy declines significantly as problem complexity grows.