Question Answering

Question answering returns an answer conditioned on a question and optional evidence. The key distinction is whether the answer is extracted, generated, retrieved, or computed.

A retrieval-augmented setup can be written as:

$$d_{1:k}=\mathrm{Retrieve}(q,\mathcal{D})$$ $$\hat{a}=\mathrm{Generate}(q,d_{1:k})$$

where $q$ is the question, $\mathcal{D}$ is the evidence corpus, and $d_{1:k}$ are retrieved documents.

Types

• Closed-book QA: answer from model parameters.
• Open-book QA: answer from retrieved evidence.
• Extractive QA: select a span from context.
• Generative QA: synthesize an answer.
• Multimodal QA: answer over image, video, table, molecule, or structure inputs.

Evidence Contract

For public and research workflows, QA should define what evidence is allowed:

$$\hat{a}=f_{\theta }(q,E),E\subseteq {\mathcal{D}}_{\mathrm{allowed}}$$

where $E$ may include retrieved documents, tool outputs, source files, figures, tables, or user-provided context. If the answer cannot be supported by $E$, the correct behavior may be abstention rather than generation.

Answer Types

The output space may be:

$$\mathcal{Y}\in \{\text{span},\text{short answer},\text{long answer},\text{yes/no},\text{list},\text{calculation},\text{citation set},\text{tool result}\}$$

Each answer type needs a different metric and validity check. Exact match is reasonable for short factual answers, but weak for synthesis, calculations, or evidence-grounded explanations.

Retrieval and Answer Coupling

A QA failure can happen in retrieval or generation:

$$\mathrm{Pr}(\text{correct answer})\le \mathrm{Pr}(\text{evidence retrieved})$$

If necessary evidence is absent from retrieved context, generation quality cannot fix the task. Evaluate retrieval recall, answer correctness, and citation support separately.

Checks

• What evidence is allowed?
• Is the answer directly supported by the context?
• Are citations or source spans required?
• Does the model abstain when evidence is missing?
• Is evaluation measuring exact answer, reasoning, retrieval quality, or citation quality?
• Is unsupported answer generation counted as failure even when fluent?
• Is the question answerable from the allowed evidence set?
• Are retrieval misses separated from reasoning or synthesis errors?
• Are citations checked for support, not only presence?

Related

• Task specification
• Retrieval-augmented generation
• Evidence-grounded generation
• Citation grounding
• Retrieval
• Text
• Multimodal learning
• Failure mode taxonomy
• Verification loop