PaperlayerView exact revision on arXiv

Source-linked AI summary

A$^2$Search: Ambiguity-Aware Question Answering with Reinforcement Learning

Fengji Zhang, Xinyao Niu, Chengyang Ying, Guancheng Lin, Zhongkai Hao, Zhou Fan, Chengen Huang, Jacky Keung, Bei Chen, Junyang Lin

arXiv:2510.07958v1cs.CLcs.AI

TL;DR

Single-gold QA benchmarks miss questions with multiple valid, evidence-supported answers, and manual ambiguity annotation is difficult to scale to multi-hop datasets. A2SEARCH automatically detects ambiguity and gathers verified alternatives, then trains with an answer-level F1 reward; it achieves state-of-the-art results across eight benchmarks, including a strong single-rollout result on multi-hop QA.

  • Problem

    Many QA benchmarks provide a single gold answer despite questions admitting multiple valid responses, while manual ambiguity annotation is difficult to scale to multi-hop datasets.

    Source: §1 · p. 1

  • Method

    A2SEARCH automatically detects ambiguous questions, gathers alternative answers through trajectory sampling and evidence verification, and uses GRPO with an answer-level F1 reward.

    Source: §1 · p. 2

  • Results

    A2SEARCH achieves state-of-the-art results across eight benchmarks; its four-multi-hop average is 48.4% AnsF1@1 under Exact Match for A2SEARCH-7B versus 46.2% for ReSearch-32B.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The findings suggest that real progress in QA requires explicitly embracing ambiguity.

    Source: §6 · p. 9

Abstract

from arXiv · show

Recent advances in Large Language Models (LLMs) and Reinforcement Learning (RL) have led to strong performance in open-domain question answering (QA). However, existing models still struggle with questions that admit multiple valid answers. Standard QA benchmarks, which typically assume a single gold answer, overlook this reality and thus produce inappropriate training signals. Existing attempts to handle ambiguity often rely on costly manual annotation, which is difficult to scale to multi-hop datasets such as HotpotQA and MuSiQue. In this paper, we present A$^2$Search, an annotation-free, end-to-end training framework to recognize and handle ambiguity. At its core is an automated pipeline that detects ambiguous questions and gathers alternative answers via trajectory sampling and evidence verification. The model is then optimized with RL using a carefully designed $\mathrm{AnsF1}$ reward, which naturally accommodates multiple answers. Experiments on eight open-domain QA benchmarks demonstrate that A$^2$Search achieves new state-of-the-art performance. With only a single rollout, A$^2$Search-7B yields an average $\mathrm{AnsF1}@1$ score of $48.4\%$ across four multi-hop benchmarks, outperforming all strong baselines, including the substantially larger ReSearch-32B ($46.2\%$). Extensive analyses further show that A$^2$Search resolves ambiguity and generalizes across benchmarks, highlighting that embracing ambiguity is essential for building more reliable QA systems. Our code, data, and model weights can be found at https://github.com/zfj1998/A2Search

1 INTRODUCTION

Open-domain QA faces a mismatch between questions with multiple evidence-supported answers and benchmarks that provide a single gold answer. A2SEARCH addresses this with automated alternative-answer generation and answer-level F1 training, achieving strong results across open-domain benchmarks.

1 INTRODUCTION

  • 27.6% of MuSiQue training examples admit multiple valid answers, although the benchmark supplies only one gold reference.Current RL pipelines reward the reference answer and implicitly penalize valid alternatives.Source: §1 · p. 1
  • A2SEARCH automatically detects ambiguous questions, gathers alternative answers through evidence verification, and trains with GRPO using an answer-level F1 reward.The reward accommodates multiple answers, and the agentic approach combines multi-step reasoning with tool use.Source: §1 · p. 2
  • 48.4% average AnsF1@1 under Exact Match for A2SEARCH-7B across four multi-hop datasets, versus 46.2% for ReSearch-32B.A2SEARCH-7B uses one rollout; the paper reports evaluation across eight open-domain QA benchmarks.Source: §1 · p. 2
  • The paper introduces an automated alternative-answer pipeline, establishes 3B and 7B RL baselines, and analyzes ambiguity detection and multi-answer generation.Its evaluation covers eight benchmarks, with A2SEARCH reported as state of the art across them.Source: §1 · p. 2

2 RELATED WORK

Prior open-domain QA work includes prompt-based and training-based approaches to using search tools. Most multi-hop benchmarks assume one correct answer, while manual ambiguity annotations are difficult to scale to multi-hop training.

2 RELATED WORK

3 METHODOLOGY

A2SEARCH builds multi-answer training data by sampling and checking alternative answers, then trains a search-capable QA policy with GRPO and an answer-level F1 reward.

3.1 ALTERNATIVE ANSWER GENERATION

  • The alternative-answer pipeline samples search trajectories, filters out reference-equivalent or uninformative answers, verifies evidence, and clusters equivalent answers into groups.It retains one representative per semantic cluster and keeps other variants as aliases.Source: §3.1 · pp. 3–4

3.2 REINFORCEMENT LEARNING FRAMEWORK.

  • The training set adds verified alternatives to each original reference, and an outcome reward combines format validity with AnsF1 to reward answer coverage while penalizing over-generation.Invalid formats receive 0; valid outputs with no matched references receive 0.1; otherwise the reward depends on AnsF1 and α.Source: §3.2 · pp. 4–5
  • GRPO updates the policy from groups of sampled rollouts, estimating its baseline from the group rather than using a separately trained critic.The objective discards the KL penalty term and uses a clipped normalized-advantage objective.Source: §3.2 · p. 4
  • Search-tool rollouts alternate reasoning, tool calls, and answers; tool responses are masked out of the policy loss because they come from the external tool.A rollout ends at an end-of-sequence token or the maximum length.Source: §3.2 · pp. 4–5

4 TRAINING DATA CONSTRUCTION

The authors construct alternative-answer training data from sampled QA trajectories, filter and verify candidate answers, and group semantic variants before measuring ambiguity across the resulting dataset.

  • Five search models generate 16 trajectories per question, yielding about 3.99 million trajectories across 49,938 source questions.The sources are MuSiQue, 2Wiki, and randomly sampled NQ questions.Source: §4 · p. 5
  • Filtering retains 208,829 trajectories across 33,997 questions, while answer grouping consolidates 28.6% of candidate answers as semantic variants.The retained trajectories constitute 5.2% of the initial trajectories and average 6.1 distinct trajectories per question.Source: §4 · pp. 5–6
  • At verification threshold η = 3, human agreement reaches 96%, leaving 19,529 verified trajectories; alternatives are identified for 19.0% of the source questions.Stricter thresholds improve agreement but reduce the number of remaining trajectories.Source: §4 · pp. 5–6
  • 27.6% of MuSiQue questions contain alternatives, compared with 19.3% of NQ and 7.2% of 2Wiki questions.MuSiQue has 5,498 such questions, versus 2,899 in NQ and 1,076 in 2Wiki.Source: Figure 3 · p. 6, §4 · p. 6

5 EXPERIMENTS

Experiments evaluate A2SEARCH on eight open-domain QA benchmarks and report strong performance with a single rollout. The analyses examine training stability, generalization, rollout efficiency, ambiguity prevalence, and representative alternative answers.

5.1 EXPERIMENTAL SETUP

  • The training dataset contains 49,938 questions, approximately 19% of which have alternative answers.Evaluation covers eight open-domain benchmarks, including four multi-hop datasets, general QA datasets, and AmbigQA.Source: §5 · p. 6, §5.1 · p. 6
  • The evaluation compares prompt-based, RL-trained search, SinSearch, and AmbigQA-trained AbgSearch baselines using AnsF1 and Recall with Exact Match or LMJudge hits.Training uses Qwen2.5 3B, 7B, Base, and Instruct variants; greedy decoding is used for AnsF1@1.Source: §5.1 · pp. 6–7

5.2 EXPERIMENTAL RESULTS

  • With one greedy rollout, A2SEARCH achieves the best AnsF1 on most benchmarks and recall comparable to or above baselines’ @3 results; it exceeds larger 32B models on several multi-hop benchmarks.SinSearch, trained on the same setup without alternative answers, remains competitive but falls short of A2SEARCH.Source: §5.2 · p. 8
  • Across four training epochs, AnsF1 and Recall increase without collapse; the 3B model’s entropy fluctuates around 0.2 while the 7B model remains within a healthy range.The training-dynamics figure presents smoothed curves.Source: §5.2 · pp. 8–9, Figure 4 · p. 8
  • Table 3 presents representative multi-hop questions with benchmark reference answers alongside evidence-supported alternatives identified by A2SEARCH.The full trajectories for these examples are provided in the appendix.Source: Table 3 · p. 9, §5.2 · p. 9
  • Additional analyses report generalization to Qwen2.5-Base, reaching @3 recall with fewer tool calls on average: 2.16 for 3B and 4.14 for 7B.Verified alternative answers occur in 19.7% of MuSiQue, 11.1% of HotpotQA, 8.6% of 2Wiki, and 8% of Bamboogle questions.Source: §5.2 · p. 9

6 CONCLUSION

A2SEARCH uses annotation-free reinforcement learning to identify ambiguous questions and generate alternative answers through trajectory sampling and evidence verification. The paper reports state-of-the-art results with one rollout and concludes that explicitly embracing ambiguity supports more reliable QA.

6 CONCLUSION

  • A2SEARCH automatically identifies ambiguous questions, discovers alternative answers through trajectory sampling and evidence verification, and trains with an answer-level F1 reward.The reward accommodates multiple references.Source: §6 · p. 9
  • A2SEARCH attains state-of-the-art results with a single rollout and generalizes effectively to AmbigQA.Source: §6 · p. 9
  • Analyses indicate that the model learns to detect ambiguity and generate multiple valid answers.Source: §6 · p. 9
  • The authors conclude that progress in QA requires explicitly embracing ambiguity, treating alternative answers as first-class signals to improve reliability and alignment with human expectations.Source: §6 · p. 9

A.1 TRAINING SETUP

The base-model experiment uses Qwen2.5-Base with prompts tailored for base models while keeping the training configuration consistent with A2SEARCH and SinSearch. These models generate multiple answers but can undergo early entropy collapse, limiting exploration and lowering validation performance.

A.1 TRAINING SETUP

  • The Qwen2.5-Base experiment keeps the A2SEARCH and SinSearch training configuration, changing to prompts tailored for base models.Source: §A.1 · p. 14
  • Base models can generate multiple answers for ambiguous questions, but early entropy collapse restricts exploration and lowers validation performance relative to instruction-tuned counterparts.Source: §A.1 · p. 14
  • The authors introduce entropy regularization with adaptive entropy control to mitigate this issue.Source: §A.1 · p. 14

A.2 ENTROPY CONTROL

Entropy regularization is added to base-model RL, with adaptive control adjusting its weight to avoid entropy collapse. Controlled training maintains higher entropy and improves validation AnsF1 and Recall.

A.2 ENTROPY CONTROL

  • The RL objective adds an entropy regularizer computed as average token-level entropy along each rollout.The entropy weight λ is dynamically adjusted around a target: increased below the target and decreased above it.Source: §A.2 · p. 14
  • Moderate entropy is important: too little leads to collapse, while too much causes unstable learning.Targets are 0.3 for 3B-Base and 0.25 for 7B-Base; λmax is 1e-2 and δ is 2e-3.Source: §A.2 · p. 14
  • Entropy control stabilizes training and improves validation AnsF1 and Recall.Figure 5 compares validation performance with the model’s entropy value under entropy control.Source: §A.2 · p. 14, Figure 5 · p. 14

A.3 RESULTS

Base models trained with A2SEARCH outperform SinSearch-Base across both sizes and all QA benchmarks, while remaining broadly comparable to A2SEARCH. The authors attribute the remaining gap to instruct models’ additional tool-use posttraining.

A.3 RESULTS

  • A2SEARCH-Base consistently outperforms SinSearch-Base across both model sizes and all QA benchmarks, with only minor degradation versus A2SEARCH.Table 4 reports Exact Match results across eight open-domain QA benchmarks using AnsF1/Recall@k; A2SEARCH and A2SEARCH-Base use one rollout.Source: §A.3 · p. 15, Table 4 · p. 15

B.1 DETAILED DESCRIPTION OF THE FILTERING STEP

The filtering pipeline retains distinct candidate answers and sends them for evidence verification. Majority voting balances human agreement against coverage when selecting supported trajectories.

B.1 DETAILED DESCRIPTION OF THE FILTERING STEP

  • The five-model ensemble achieves higher Recall@80 than individual models’ Recall@16, providing broader reference-answer coverage and more diverse trajectories.Recall@k measures whether at least one of k sampled trajectories yields an answer semantically equivalent to the reference.Source: Table 5 · p. 15, §B.1 · pp. 15–16
  • Filtering removes reference-equivalent answers and question-level sets with no reference match, while retaining non-canonical answers for verification when canonical and alternative answers coexist.This coarse filtering aims to preserve valid alternatives distinct from the reference answer.Source: §B.1 · p. 16
  • Answer-level de-duplication lowercases candidates, removes punctuation, and trims whitespace, keeping one trajectory per unique normalized answer.Source: §B.1 · p. 16
  • After filtering and de-duplication, 208,829 trajectories across 33,997 questions remain, or 5.2% of the original trajectories and 68.1% of source questions.The retained set averages 6.1 distinct trajectories per question and becomes the input to verification.Source: §B.1 · p. 16

B.2 DETAILED DESCRIPTION OF THE VERIFICATION STEP

  • Four LLM verifiers label candidates supported, partially supported, or not supported; majority voting with threshold η reconciles their differing label tendencies.The partially supported label is intended to prevent borderline cases from being over-classified as supported.Source: §B.2 · p. 16
  • At η = 3, voting yields 96% human agreement while retaining 9.4% of trajectories (19,529); η = 4 reaches 99% agreement but retains only 4.6%.The authors adopt η = 3 to balance agreement and coverage; Table 7 reports supported counts and proportions by threshold.Source: Table 7 · p. 16, §B.2 · p. 17

C.1 ANSWER COUNT DISTRIBUTION

A2SEARCH produces multiple answers more often on benchmarks with greater ambiguity, especially MuSiQue, while usually returning one answer overall. The authors report that it avoids over-generating when a question has a single correct answer.

C.1 ANSWER COUNT DISTRIBUTION

  • A2SEARCH-7B gives one answer in about 75.1% of cases, two in 15.4%, and three or more in roughly 9.5%; its average is 2.26 on MuSiQue versus 1.31 on Bamboogle and 1.33 on 2Wiki.Table 8 presents answer-count distributions across benchmarks.Source: §C.1 · p. 17, Table 8 · p. 17
  • A2SEARCH-3B averages 1.23 answers per question, compared with 1.51 for 7B; 81.5% of its questions receive one answer and 13.6% receive two.MuSiQue has the highest 3B average answer count at 1.40.Source: §C.1 · p. 17
  • The authors report that A2SEARCH adapts answer counts to dataset ambiguity, capturing multiple answers when needed while refraining from over-generating for single-answer questions.Source: §C.1 · p. 17

C.2 ROLLOUT EFFICIENCY

A2SEARCH uses tool calls efficiently: more calls do not guarantee higher recall, and A2SEARCH reaches strong recall with fewer rollouts than several baselines.

C.2 ROLLOUT EFFICIENCY

  • 4.14 calls accompany A2SEARCH-7B's highest Recall@1 of 51.2% at k = 1 and rptc of 0.12, surpassing other 7B baselines.The section characterizes its additional calls as productive rather than uninformative exploration.Source: Table 9 · p. 18, §C.2 · p. 18
  • rptc measures recall per tool call, relating each model's recall gain to tool-call inference cost.Table 9 reports average tool calls, recall at different rollout depths, and rptc across four multi-hop QA benchmarks.Source: §C.2 · pp. 17–18, Table 9 · p. 18
  • 2.16 calls give A2SEARCH-3B Recall@1 of 44.7% and rptc of 0.21, on par with or above Search-R1-32B's 0.22.The paper reports this in the single-rollout, temperature = 0 setting.Source: §C.2 · p. 18
  • A2SEARCH-7B exceeds 51% recall with one rollout, while ReSearch-7B uses two or three rollouts to reach 45–50% Recall@k.ReSearch, Search-R1, and AFM-MHQA often use multiple stochastic rollouts to approach similar recall.Source: §C.2 · p. 18

C.3 AMBIGUITY RATIO OF QA BENCHMARKS

The authors estimate that valid alternative answers occur across all seven examined benchmarks, with rates varying by dataset. Multi-hop datasets show especially high rates for MuSiQue and HotpotQA.

C.3 AMBIGUITY RATIO OF QA BENCHMARKS

  • 19.7% of MuSiQue and 11.1% of HotpotQA questions exhibit valid alternative answers, the highest rates among the multi-hop datasets.The estimates use six sampled rollouts per question at temperature 0.6 from five baseline models.Source: §C.3 · p. 18
  • 14.7% of PopQA questions and 7.0% of TriviaQA questions exhibit valid alternative answers; PopQA is highest and TriviaQA relatively less ambiguous in general QA.Source: §C.3 · p. 18
  • Valid alternative answers occur in 8.6% of 2Wiki, 8.0% of Bamboogle, and 12.7% of NQ questions.These are the reported ambiguity estimates for the three benchmarks.Source: §C.3 · p. 18

D EXPERIMENTAL SETUP

The appendix estimates multi-answer scores by averaging over uniformly sampled subsets of rollouts, rather than selecting only the best or worst samples. Experiments use temperature 0.6 to encourage rollout diversity.

D.1 AnsF1@k ESTIMATION ALGORITHM

  • AnsF1@k estimation averages precision, recall, and F1 across all size-k subsets drawn uniformly without replacement from k′ sampled trajectories.This avoids the biased picture produced by reporting only the best or worst case among k′ samples.Source: §D.1 · p. 18
  • Experiments use temperature 0.6, which encourages stronger rollout diversity and increases the chance of capturing multiple answers.Source: §D.1 · p. 19
  • Algorithm 1 counts matched predictions and distinct reference answers in each subset, then averages the resulting precision, recall, and F1 to estimate AnsF1@k and Recall@k.The inputs include the matched reference answer for each prediction or ⊥, the total number of reference answers, and k.Source: §D.1 · p. 19

E.1 THE ROLE OF α IN REWARD DESIGN

The appendix examines reward weighting, rollout size, and evaluation sampling temperature. Its ablations describe trade-offs between precision, recall, validation performance, and efficiency.

E.1 THE ROLE OF α IN REWARD DESIGN

  • α = 0.4 achieves validation stability and a reasonable precision–recall balance, between the single-answer tendency at α = 0.8 and answer enumeration at α = 0.2.Figure 6 tracks validation performance and answer count; a larger α penalizes imperfect AnsF1 more, while a smaller α weakens the precision incentive.Source: §E.1 · pp. 19–20, Figure 6 · p. 20

E.2 ROLLOUT SIZE

  • Larger rollout sizes improve validation Precision, Recall, and AnsF1, but the gap between sizes 16 and 32 is small; the authors use 16 for efficiency.The A2SEARCH-7B ablation compares rollout sizes 8, 16, and 32 over two epochs.Source: Figure 7 · p. 20, §E.2 · p. 20

E.3 SAMPLING TEMPERATURE

  • ReSearch-7B recall peaks at 31.1 at T = 0.6 without significant F1 loss, and average recall across three baselines peaks at T = 0.6.The authors select T = 0.6 as a trade-off between recall and F1 while allowing each baseline to perform near its best.Source: Table 10 · p. 21, §E.3 · p. 21
  • Search-R1-7B has relatively stable performance across temperatures, while AFM-MHQA-7B's F1 and recall decline more sharply as temperature rises.Source: §E.3 · p. 21

E.4 EVALUATING WITH LMJUDGE

This section evaluates A2SEARCH using LMJudge as a semantic alternative to exact-match evaluation and describes the SinSearch-Base prompt’s tagged sections for reasoning, tool use, and the final answer.

E.4 EVALUATING WITH LMJUDGE

  • Because EM can reward paraphrased variants without showing true ambiguity resolution, the evaluation adds Qwen2.5-32B-Instruct judgments of semantic matches and recomputes AnsF1 and Recall.During training, AnsF1 counts distinct reference answers, so multiple synonyms matching one reference yield one hit and reduced precision.Source: §E.4 · p. 21
  • Under LMJudge, A2SEARCH achieves the best AnsF1 on multi-hop QA, while agentic search models outperform prompting- and RAG-based baselines, especially on multi-hop datasets.The reported conclusions are consistent with those under EM.Source: §E.4 · pp. 21–22, Table 11 · p. 22
  • With one greedy rollout, A2SEARCH matches or exceeds baseline Recall@3, beats larger 32B models on several benchmarks, and surpasses AbgSearch on AmbigQA despite no AmbigQA training.AbgSearch performs well on AmbigQA but does not generalize as well across benchmarks.Source: Table 11 · p. 22, §E.4 · p. 22, Table 12 · p. 23

Prompt for LMJudge

  • The LMJudge prompt asks for a rationale and a correct-or-incorrect judgment based on whether a prediction is semantically equivalent to at least one ground-truth answer.It accepts semantic matches rather than requiring exact wording.Source: Prompt for LMJudge · p. 23

Prompt for Evidence-based Verification

  • The evidence-consistency judge evaluates each answer claim against retrieved tool results, ignoring internal thinking and outside knowledge, and permits concise answers that select one reasonable interpretation.It labels claims supported, partially supported, or unsupported and returns a JSON verdict.Source: Prompt for Evidence-based Verification · p. 24

Prompt for Grouping Semantically Identical Answers

Prompt for Training SinSearch

  • The SinSearch training prompt directs the model to search relevant Wikipedia pages, branch across interpretations, and synthesize and justify multiple answers when ambiguity exists.Its output uses structured reasoning and answer tags.Source: Prompt for Training SinSearch · p. 25

Prompt for Training A2SEARCH and AbgSearch

Prompt for training SinSearch-Base

  • The prompt places internal reasoning in <think>, search queries and results in <search> and <result>, and the exact final answer in a boxed LaTeX expression within <answer>.The assistant may invoke Wikipedia search during reasoning when factual information is needed.Source: Prompt for training SinSearch-Base · p. 27

G ROLLOUT CASES OF A2SEARCH

This section presents A2SEARCH’s reason-and-search trajectories across questions about multi-hop relationships, nationality, work history, music, and testosterone synthesis. The cases trace retrieved evidence through to answers, including one example that explores alternative interpretations.

G ROLLOUT CASES OF A2SEARCH

  • The section presents full reason-and-search trajectories for A2SEARCH’s answers to ambiguous questions listed in Table 3.The reproduced text omits some special Unicode characters, which may cause minor character loss.Source: G ROLLOUT CASES OF A2SEARCH · p. 28
  • For a MuSiQue question linking Islamic philosophy’s most influential figure to a claim about great thinkers, A2SEARCH returned Oliver Leaman and George Sarton.It considered Mulla Sadra and Avicenna as possible figures, then identified Leaman’s description of Mulla Sadra and Sarton’s description of Avicenna.Source: G ROLLOUT CASES OF A2SEARCH · pp. 28, 31
  • A second MuSiQue trajectory begins by decomposing a question about the owner of the record label for the performer of “What Kind of Love.”The supplied passage ends after the model starts identifying the performer, so it does not show the resulting answer.Source: G ROLLOUT CASES OF A2SEARCH · p. 31

Rollout cases

  • The “What Kind of Love” trace identifies Rodney Crowell and returns Warner Music Group and Sony Music Entertainment as owners of his labels.The trace connects him to Warner Bros. Records and Columbia Records, respectively.Source: G ROLLOUT CASES OF A2SEARCH · p. 34
  • For the question about Prince Nikolaus Wilhelm of Nassau’s mother, the reference answer gives Germany as her country of origin.The trace identifies his mother as Princess Pauline of Württemberg, born in Stuttgart, Kingdom of Württemberg.Source: G ROLLOUT CASES OF A2SEARCH · pp. 34–35

G ROLLOUT CASES OF A2SEARCH

  • The Pauline rollout concludes that Prince Nikolaus Wilhelm of Nassau’s mother was from Württemberg and therefore German.The response identifies his mother as Princess Pauline of Württemberg and infers her nationality from her Württemberg title and origin.Source: G ROLLOUT CASES OF A2SEARCH · p. 36
  • The Hayranidil Kadın rollout identifies her husband as Sultan Abdulaziz and concludes he died at raan Palace in Constantinople.The trace links Hayranidil Kadın to Abdulaziz, whose death passage names raan Palace in Constantinople.Source: G ROLLOUT CASES OF A2SEARCH · pp. 36, 38
  • The Wine of Morning rollout identifies Katherine Stenholm as director and answers that she worked at Bob Jones University.Stenholm is described as founding director of Unusual Films, Bob Jones University’s production company, and as continuing to teach at BJU until 2001.Source: G ROLLOUT CASES OF A2SEARCH · pp. 38–40

Stenholm work-history case

  • Katherine Stenholm worked at Bob Jones University as founding director of its Unusual Films production company and taught there until 2001.The rollout concludes that her workplace included Bob Jones University.Source: G ROLLOUT CASES OF A2SEARCH · p. 40

Music-similarity case

  • Ceephax Acid Crew and Squarepusher are both associated with drum and bass, the similarity identified in the HotpotQA reference answer.Ceephax is described as a drum-and-bass electronic musician, while drum and bass is among Squarepusher’s musical influences.Source: G ROLLOUT CASES OF A2SEARCH · p. 41

G ROLLOUT CASES OF A2SEARCH

  • The passage describes androstenedione biosynthesis through pathways from 17-hydroxypregnenolone or 17-hydroxyprogesterone, with production in the adrenal glands and gonads.The first pathway proceeds through DHEA; the second converts 17-hydroxyprogesterone directly to androstenedione.Source: G ROLLOUT CASES OF A2SEARCH · p. 46
  • Testosterone is identified as derived from cholesterol and androstenedione, and androstenedione is described as a precursor of testosterone.Androstenedione is converted to testosterone by 17-hydroxysteroid dehydrogenase.Source: G ROLLOUT CASES OF A2SEARCH · pp. 46–47
Loading 2510.07958v1…