anonymous2026nips/llm-answer-set-qa

Answer-Set Consistency Benchmark (ASCB)

Overview

The Answer-Set Consistency Benchmark (ASCB) evaluates whether language models provide mutually consistent answers to related factual enumeration questions. Unlike conventional QA datasets, ASCB focuses on whether generated answer sets satisfy known set-theoretic relations rather than solely on factual accuracy.

ASCB contains 600 English question quadruples (2,400 questions) across primarily static, objective factual domains.

Source	Quadruples	Questions
QALD	150	600
LC-QuAD 2.0	150	600
QAWiki	150	600
Synthetic	150	600
Total	600	2,400

Dataset structure

Each example contains four related questions (Q1, Q2, Q3, Q4) satisfying:

• answers(Q1) = answers(Q2) — equivalence;
• answers(Q3) ⊂ answers(Q1) — containment;
• answers(Q4) ⊂ answers(Q1) — containment;
• answers(Q3) ∩ answers(Q4) = ∅ — disjointness; and
• answers(Q4) = answers(Q1) \ answers(Q3) — set difference.

Conceptually, each record includes a quadruple identifier, its source, and the four questions. Related questions should remain grouped when creating experimental splits to avoid leakage.

ASCB does not provide exhaustive gold answer sets. It is designed to evaluate relations between model-generated sets, not answer accuracy by itself.

Dataset creation

Questions were selected to have objective, non-empty, finite sets of answers and to avoid yes/no questions, single-value answers, vague categories, and multiple answer dimensions. The benchmark design targeted approximately 2–100 answers per question; the preliminary automated filter used a stricter 2–50 range.

• QALD and LC-QuAD 2.0: Candidate KGQA questions and their structured queries were screened for suitability. An LLM-assisted pipeline proposed equivalent and subset questions, followed by extensive manual editing.
• QAWiki: Existing equivalent or containment pairs were completed manually to form quadruples.
• Synthetic: GPT-4.1 proposed 500 candidate sets, which were reviewed, substantially rewritten, deduplicated, and reduced to 150 quadruples.

The LLM-assisted pipeline validated Q1, generated an equivalent Q2, proposed complementary subsets Q3 and Q4, and checked the complete quadruple. Automated output was used only as a candidate source and was never accepted without human review.

Validation and data quality

Three authors manually reviewed, corrected, and pruned the dataset. Validation focused on:

• equivalence of Q1 and Q2;
• strict containment of Q3 and Q4 within Q1;
• disjointness and complete coverage (Q3 ∪ Q4 = Q1);
• objective and unambiguous wording;
• fluent English and topic diversity; and
• removal of duplicates and malformed questions.

Structured queries were evaluated where available to confirm enumeration behavior and approximate answer cardinality. Manual review was essential because LLM-generated subsets did not always form a valid partition and were corrected or discarded.

Recommended use

Ask each question independently, parse responses as sets, normalize aliases and formatting, and test whether the sets satisfy the encoded relations. Report abstentions and empty answers separately because they may produce trivial consistency. Jaccard similarity can complement exact relation satisfaction.

ASCB supports research on answer-set consistency, relation classification, paraphrase robustness, set reasoning, and relation-aware prompting. It should not be treated as an exhaustive knowledge base or as proof of factual reliability.

License and citation

If you use ASCB, please cite the accompanying paper. We will replace the placeholder below with the final bibliographic record once available:

@inproceedings{ascb2026,
  title     = {Answer-Set Consistency Benchmark},
  author    = {Anonymous authors},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2026}
}