HotpotQA vs cua

Provides 113,000 question-answer pairs where each question requires chaining reasoning across 2+ Wikipedia articles to derive the answer. The dataset includes explicit supporting fact annotations identifying which sentences from source documents are necessary for answering, enabling training of models that can both answer questions and justify their reasoning through evidence selection. Built on Wikipedia snapshots with crowdsourced annotation of answer spans and supporting sentences.

Unique: Combines answer prediction with supporting fact annotation in a single dataset, enabling joint training of answer generation and evidence selection. Unlike SQuAD (single-document) or MS MARCO (ranking-focused), HotpotQA explicitly requires models to perform intermediate reasoning steps and identify which sentences enable the final answer, making it the first large-scale dataset to measure both answer correctness AND reasoning transparency.

vs alternatives: Uniquely measures explainability through supporting fact prediction rather than just answer accuracy, forcing models to learn which evidence matters rather than memorizing answer patterns from single documents.

Enables evaluation of whether QA systems can decompose complex questions into sub-questions, retrieve relevant documents for each step, and chain reasoning across multiple sources. The dataset structure (questions requiring 2+ hops) forces models to learn retrieval-then-reasoning patterns rather than end-to-end memorization. Supports both open-domain (retrieve from full Wikipedia) and distractor-based (retrieve from provided candidates) evaluation modes.

Unique: Explicitly structures questions to require intermediate reasoning steps (e.g., 'Who directed film X?' → find film → find director → extract name), forcing evaluation of whether systems learn compositional reasoning vs pattern matching. Supporting fact annotations enable measuring retrieval quality independently from answer correctness, unlike SQuAD where retrieval is implicit.

vs alternatives: Uniquely decouples retrieval evaluation from answer evaluation through supporting fact metrics, revealing whether models retrieve correct evidence even when they produce wrong answers — a diagnostic capability absent from single-document QA benchmarks.

Provides ground-truth supporting fact annotations (sentence-level indices from source documents) enabling training and evaluation of models that predict which evidence is necessary for answering. This enables measuring explainability as a quantitative metric (supporting fact F1/precision/recall) rather than qualitative assessment. Models can be trained jointly on answer prediction and supporting fact prediction, or separately for interpretability analysis.

Unique: First large-scale QA dataset to include sentence-level supporting fact annotations, enabling quantitative measurement of explainability through supporting fact F1 rather than subjective evaluation. This shifts explainability from a qualitative property to a measurable metric that can be optimized during training.

vs alternatives: Enables explainability as a first-class optimization target (supporting fact F1) rather than an afterthought, unlike SQuAD or MS MARCO where evidence selection is implicit and unmeasured.

Provides a curated set of distractor documents (Wikipedia articles that are topically related but don't contain supporting facts) alongside correct source documents, enabling controlled evaluation of reading comprehension and reasoning without requiring full retrieval. Models receive a fixed set of candidate documents and must identify which contain relevant information and extract answers, isolating reasoning capability from retrieval quality.

Unique: Provides curated distractor documents (topically related but non-supporting) rather than random negatives, enabling more realistic evaluation of document relevance judgment. Distractors are selected to be challenging (e.g., same topic, different entity) rather than trivial, forcing models to perform fine-grained reasoning.

vs alternatives: Offers a middle ground between single-document SQuAD (no retrieval challenge) and open-domain evaluation (expensive retrieval), enabling controlled reasoning assessment with realistic document selection difficulty.

Serves as a standardized benchmark for measuring both answer correctness and reasoning transparency through supporting fact prediction. The dataset includes train/dev/test splits with consistent evaluation protocols, enabling reproducible comparison of QA systems on their ability to produce correct answers AND identify supporting evidence. Supports multiple evaluation metrics (answer F1, supporting fact F1, combined scores) for comprehensive system assessment.

Unique: Combines answer evaluation with supporting fact evaluation in a single benchmark, forcing systems to be evaluated on both correctness AND transparency. Unlike SQuAD (answer-only) or information retrieval benchmarks (ranking-only), HotpotQA measures the full pipeline of reasoning, retrieval, and justification.

vs alternatives: Uniquely standardizes evaluation of reasoning transparency alongside answer accuracy, enabling reproducible comparison of systems on their ability to justify answers — a capability absent from single-metric benchmarks.

Questions are generated from Wikipedia articles and require reasoning over real-world entities, relationships, and facts. This grounds reasoning in a concrete knowledge domain (Wikipedia) rather than synthetic or template-based questions, enabling evaluation of whether systems can handle real-world complexity. Questions span diverse topics (people, places, films, organizations) and reasoning patterns (attribute lookup, entity linking, relationship chaining).

Unique: Questions are grounded in real Wikipedia entities and relationships rather than synthetic templates, requiring models to handle actual knowledge base complexity (entity disambiguation, relationship chaining, fact lookup). This makes reasoning evaluation more realistic than template-based datasets.

vs alternatives: Grounds reasoning in a real, large-scale knowledge base (Wikipedia) rather than synthetic examples, enabling evaluation of whether systems can handle real-world entity linking and relationship reasoning.

Captures desktop screenshots and feeds them to 100+ integrated vision-language models (Claude, GPT-4V, Gemini, local models via adapters) to reason about UI state and determine appropriate next actions. Uses a unified message format (Responses API) across heterogeneous model providers, enabling the agent to understand visual context and generate structured action commands without brittle selector-based logic.

Unique: Implements a unified Responses API message format abstraction layer that normalizes outputs from 100+ heterogeneous VLM providers (native computer-use models like Claude, composed models via grounding adapters, and local model adapters), eliminating provider-specific parsing logic and enabling seamless model swapping without agent code changes.

vs alternatives: Broader model coverage and provider flexibility than Anthropic's native computer-use API alone, with explicit support for local/open-source models and a standardized message format that decouples agent logic from model implementation details.

Provisions isolated execution environments across macOS (via Lume VMs), Linux (Docker), Windows (Windows Sandbox), and host OS, with unified provider abstraction. Handles VM/container lifecycle (creation, snapshot management, cleanup), resource allocation, and OS-specific action handlers (keyboard/mouse events, clipboard, file system access) through a pluggable provider architecture that abstracts platform differences.

Unique: Implements a pluggable provider architecture with unified Computer interface that abstracts OS-specific action handlers (macOS native events via Lume, Linux X11/Wayland via Docker, Windows input simulation via Windows Sandbox API), enabling single agent code to target multiple platforms. Includes Lume VM management with snapshot/restore capabilities for deterministic testing.

vs alternatives: More comprehensive OS coverage than single-platform solutions; Lume provider offers native macOS VM support with snapshot capabilities unavailable in Docker-only alternatives, while unified provider abstraction reduces code duplication vs. platform-specific agent implementations.