Source-linked AI summary
ACL-Verbatim: hallucination-free question answering for research
Gábor Recski, Szilveszter Tóth, Nadia Verdha, István Boros, Ádám Kovács
TL;DR
AI-assisted research tools can hallucinate, making reliable evidence retrieval from scientific literature difficult. ACL-Verbatim applies extractive VerbatimRAG to ACL papers and finds that customized models trained with synthetic data outperform zero-shot LLM extraction on the task.
Problem
LLM-based question answering over scientific literature remains difficult to trust because outputs can be inaccurate, irrelevant, or nonsensical.
Method
ACL-Verbatim applies extractive VerbatimRAG to ACL Anthology papers and trains query-conditioned token classifiers to map questions to evidence spans.
Results
A small customized extractive model trained with synthetic data outperforms zero-shot LLM-based extraction on the benchmark, while achieving higher precision through abstention on irrelevant chunks.
Takeaways & Limitations
The results support combining VerbatimRAG with task-oriented extractive training for efficient, high-precision question answering over research papers.
Takeaways & Limitations
Conclusions are limited by the small manually annotated dataset and the complexity of annotation, which constrained agreement measurement and rigorous adjudication.
Abstract
from arXiv · showhide
Academic researchers need efficient and reliable methods for collecting high-quality information from trusted sources, but modern tools for AI-assisted research still suffer from the tendency of Large Language Models (LLMs) to produce factually inaccurate or nonsensical output, commonly referred to as hallucinations. We apply the extractive question answering system VerbatimRAG to research papers in the ACL Anthology, directly mapping user queries to verbatim text spans in retrieved documents. We contribute a novel ground truth dataset for the task of mapping user queries to relevant text spans in research papers, and use it to train and evaluate a variety of extractive models. Human annotation is performed by NLP researchers and is based on synthetic user queries generated using a custom pipeline based on the ScIRGen methodology, paired with chunks of research papers retrieved by VerbatimRAG. On this benchmark, a 150M-parameter ModernBERT token classifier trained on silver supervision from our pipeline achieves the best word-level F1 (53.6), ahead of the strongest evaluated LLM extractor (48.7).
1 Introduction
Researchers need reliable evidence from large scientific-literature collections, but finding relevant information is difficult. LLM-based tools, including RAG systems, can generate hallucinated answers that require tedious independent fact-checking, motivating ACL-Verbatim.
- Motivation: Finding relevant evidence in large scientific-paper collections remains difficult despite researchers’ reliance on scientific literature as a trusted information source.Modern AI tools are presented as a way to improve information-search efficiency.
- Motivation: RAG systems remain prone to LLM-related hallucinations, including factually inaccurate, irrelevant, or nonsensical output.The passage identifies hallucinations as a major limitation even when retrieval is combined with generative AI.
- Motivation: LLM answers cannot be trusted without independent fact-checking, but verification is tedious and often omitted.The passage attributes omission partly to algorithm appreciation, in which people favor algorithmic judgment despite opaque processes.
- Contribution: The paper presents ACL-Verbatim to address the reliability problems associated with AI-assisted research.The supplied passage introduces ACL-Verbatim immediately after discussing hallucinations and inadequate verification.
2 Related work
Prior work identifies hallucination and poor explainability as central risks of LLM-generated answers, including in retrieval-augmented systems. VerbatimRAG addresses these risks by returning extractive, verbatim spans from source documents, while the ACL Anthology provides the paper resource used here.
- LLM limitations: LLM question-answering systems can produce factually incorrect, inconsistent, or nonsensical outputs, making their answers unreliable without independent fact-checking.This undermines their use in critical medical, legal, and financial domains.
- RAG limitations: RAG reduces intrinsic hallucinations by grounding models in external sources, but extrinsic hallucinations remain when models override retrieved information with prior knowledge.RAG systems therefore continue to hallucinate, limiting their use in complex, high-risk domains.
- Explainability: LLMs also lack reliable explainability, because post-hoc explanations can fail and intuitive self-explanations can wrongly inflate users’ trust.These risks arise when neural language models generate the final output presented to users.
- VerbatimRAG: VerbatimRAG tackles hallucination by combining retrieval with extraction and returning only text spans taken verbatim from source documents.Its extraction approaches include prompted LLMs and smaller fine-tuned models such as Provence.
- ACL Anthology: The ACL Anthology is a public resource underlying the experiments and dataset, with prior corpora adding metadata that enables advanced NLP research analysis.Examples include NLP Scholar, NLPExplorer, and ACL-OCL.
3 Corpus creation
The corpus pipeline converts ACL Anthology papers into indexed markdown chunks and builds query–span annotations through retrieval and human relevance judgments. Creating this ground truth is labor-intensive and requires substantial domain expertise because relevance and span selection can be difficult and subjective.
- Data collection: The corpus draws on over 120,000 ACL Anthology research papers, whose metadata and full-text PDFs are programmatically accessible under permissive licensing.The source covers computational linguistics and natural language processing.
- Preprocessing: 114 475 PDFs were converted into markdown with Docling, producing a released dataset while skipping fewer than 100 papers because of conversion errors.Text content such as headers, lists, tables, and captions is rendered in markdown, while some figures and formulas are replaced by placeholders.
- Segmentation: A custom markdown chunker parses section structure, segments at section boundaries, prefixes section titles, and avoids splitting tables or code blocks.The chunks are indexed with the VerbatimRAG library.
- Annotation dataset: The ground-truth pipeline samples 333 papers, selects one chunk per paper, generates queries, retrieves the top 5 chunks per query, and annotates relevance before highlighting answer spans.Irrelevant chunks receive no further annotation; relevant chunks require annotators to mark the most relevant text spans, or captions for relevant tables and figures.
- Annotation challenges: 906 queries and 4530 query-chunk pairs were sampled, but human annotation covered only the first 20 queries and 100 chunks because the task was complex.Annotators faced domain-specific, difficult, and subjective judgments, so the researchers used NLP researchers with diverse backgrounds rather than requiring parsing specialists.
4 Extraction experiments
The extraction experiments evaluate multiple span-extraction systems on a 100-row manually annotated benchmark, using word-level precision and recall as the primary metric. They also train a query-conditioned ModernBERT token classifier from silver supervision and release a multi-domain sibling model.
- Benchmark: 100 query–chunk pairs comprise the benchmark, including 47 relevant chunks with 78 gold evidence spans and 53 irrelevant chunks without gold spans.Metrics are reported across all 100 rows.
- Compared systems: Four LLM span extractors—Mistral Small 2603, Nemotron-120B-A12B, GLM-5, and Qwen 3.6 35B—are compared with default and paragraph-oriented prompts for three models.The systems receive a question and retrieved chunk and must return verbatim evidence spans.
- Student model: The student model is a query-conditioned binary token classifier over an 8192-token ModernBERT backbone, trained to decode evidence labels into character spans.The input concatenates the question and chunk, and sliding-window inference is used.
- Evaluation: Word-level precision and recall are the primary evaluation metrics because they assess whether extracted spans highlight the correct words without depending on exact span boundaries.The metrics compare words covered by gold and extracted spans across all 100 benchmark rows.
- Additional release: The released multi-domain sibling model combines ACL silver data with RAGBench and Squeez for retrieval-augmented question answering and coding-agent tool-output pruning domains.The model is KRLabsOrg/verbatim-rag-modern-bert-v210, trained on KRLabsOrg/verbatim-spans11.
5 Results
On the 100-row manually annotated benchmark, the reranker-initialized ModernBERT student achieves the best Word-F1, surpassing the strongest LLM extractors while using far fewer parameters. ModernBERT models also remain competitive on ACL gold, with the ACL-specialized model attaining the highest word-level precision and often abstaining on irrelevant chunks.
- Model comparison: 53.63 Word-F1 is achieved by the reranker-initialized ModernBERT student, ahead of GLM-5 at 48.71, Mistral Small at 46.94, and Qwen with paragraph prompt at 46.73.The comparison uses 100 rows from the manually annotated benchmark.
- Model comparison: 3 to 4 orders of magnitude fewer parameters are used by the ModernBERT student than by the compared LLM extractors.The parameter reduction accompanies the student’s higher Word-F1.
- ACL gold performance: 46.29 ACL gold Word-F1 is achieved by the generic multi-domain ModernBERT model, which outperforms the public extractive baselines.The model remains competitive despite not being specialized only for the ACL Anthology.
- Precision and recall: The ACL-specialized model achieves the highest word-level precision and often abstains on irrelevant chunks, unlike the LLM extractors.LLM extractors, particularly with paragraph-oriented prompts, achieve higher recall.
6 Conclusion
The paper applies VerbatimRAG to more than 100K ACL Anthology research papers, introduces a manually annotated extraction dataset, and evaluates customized extractive models trained with synthetic data. The authors release the pipeline openly and propose combining VerbatimRAG with task-oriented extractive training for efficient hallucination-free question answering across domains.
- The study applies the VerbatimRAG architecture to over 100K research papers in the ACL Anthology.
- The authors contribute a manually annotated dataset for the core extraction task.
- Experiments show that a small customized encoder-decoder architecture trained with synthetic data outperforms zero-shot LLM-based extraction at a fraction of the cost.
- All components of the pipeline are released as open-source software.
- The authors present this combination of VerbatimRAG and task-oriented extractive-model training as a blueprint for efficient, high-performing hallucination-free question answering across domains.
Limitations
The study’s conclusions are constrained by the small manually annotated dataset and by the annotation task’s complexity, which limited evaluation rigor and annotation quality controls.
- The small manually annotated dataset limits the validity of both quantitative and qualitative evaluation.
- The annotation task’s high complexity limited the ability to measure agreement between multiple annotators.
- The study could not establish a rigorous adjudication process or detailed, objective annotation guidelines.
A Detailed model comparison
Table 2 compares extractor configurations from Table 1 using containment and coverage metrics. Evaluation covers all 100 benchmark rows, including 53 irrelevant query–chunk pairs as negative examples.
- A Detailed model comparison: Table 2 reports containment and coverage metrics for the extractor configurations evaluated in Table 1.The comparison focuses on two metric categories across the evaluated configurations.
- A Detailed model comparison: Evaluation uses the full 100-row benchmark rather than a subset.All reported metrics are computed across the complete benchmark.
- A Detailed model comparison: 53 irrelevant query–chunk pairs are included as negative examples in the evaluation.These irrelevant pairs are part of the 100 benchmark rows used for computing every metric.
B Threshold selection for the student model
The student model’s span decisions depend on an inference probability threshold, evaluated at t ∈ {0.2, 0.3, 0.4, 0.5} with fixed post-processing. Word-F1 is highest at t=0.2 because higher thresholds trade recall for precision and reduce F1.
- Threshold evaluation: The binary token classifier’s span decisions depend on an inference probability threshold evaluated at t ∈ {0.2, 0.3, 0.4, 0.5}.Table 3 reports all-row gold benchmark scores for the GTE-reranker student.
- Threshold evaluation: Post-processing drops spans shorter than 10 characters and merges neighbouring spans separated by at most 20 characters.These settings are held constant across threshold values.
- Threshold selection: The best Word-F1 is obtained at t=0.2.Increasing the threshold improves precision but reduces recall, lowering F1 on this benchmark.
C Prompts for query generation
Appendix C specifies prompts for classifying question types, generating questions from paper text, and rewriting questions into search queries. The prompts constrain outputs to concise, task-specific formats, including valid JSON for classification and question-only responses for generation.
- C.1 Question-type classification prompt: The classification prompt frames the task as generating questions and answers for finding relevant information within a specific domain.
- C.1 Question-type classification prompt: It asks the model to choose the question type that best fits the field information and the user’s purpose.
- C.1 Question-type classification prompt: The listed types include verification, disjunctive, concept completion, example, feature specification, quantification, definition, and comparison questions.
- C.1 Question-type classification prompt: Given research-paper text, the classifier must return the three most appropriate question types, naming each type without additional information.
- C.1 Question-type classification prompt: The classification output must be only valid JSON consisting of an array of objects, with no markdown or explanations.
- C.2 Question generation prompt: The generation prompt supplies a paper-text chunk, a question type, its description, and an example, then requests one answerable question from that text.
- C.3 Query rewriting prompt: The rewriting prompt presents a user question and asks for a search query that could be used to find its answer.
D Prompts for extraction
The appendix specifies two extraction prompts that require verbatim, question-relevant evidence returned as document-indexed JSON arrays. The default prompt emphasizes sentence-level spans and relevance ordering, while the paragraph-style prompt expands passages to include contextual and interpretive material.
- D.1 Default VerbatimRAG extraction prompt: The default prompt extracts EXACT verbatim text spans from multiple documents that answer the question.It uses question and document placeholders as inputs and requests verbatim spans from each document.
- D.1 Default VerbatimRAG extraction prompt: Default extraction forbids paraphrasing or additions, preserves wording, capitalization, and punctuation, and orders spans by relevance within each document.The prompt also requires complete sentences or paragraphs for context.
- D.1 Default VerbatimRAG extraction prompt: The default output maps document IDs to relevance-ordered span arrays, using an empty array when no relevant information exists.The specified JSON structure includes separate arrays for each document.
- D.2 Paragraph-style extraction prompt: The paragraph-style prompt extracts verbatim supporting passages from each document that answer the question.A supporting passage is defined as the complete portion a researcher would highlight to justify the answer.
- D.2 Paragraph-style extraction prompt: Paragraph-style passages include direct answer sentences, preceding setup, concluding interpretations, and table captions when relevant.These components capture the referenced topic, methodology, figure, implications, or table context.
- D.2 Paragraph-style extraction prompt: The paragraph-style prompt prefers a single continuous paragraph and splits into multiple spans only when relevant content is non-adjacent.This avoids fragmented excerpts from the same paragraph.
- D.2 Paragraph-style extraction prompt: Its rules require exact text without edits, preserved capitalization and punctuation, empty arrays for unsupported documents, and relevance-ordered spans.The output remains JSON mapping document IDs to arrays of supporting passages.
- D.2 Paragraph-style extraction prompt: The paragraph-style template supplies question and document placeholders before requesting supporting passages from each document.This mirrors the default prompt’s document-by-document extraction interface.