ResearchQA is introduced, a benchmark of 6,211 single-paper question-answer pairs from 494 open-access papers spanning eight domains and four question types: lookup, comprehension, multi-hop, and adversarial, designed for citation-grounded evaluation.
Abstract
Large language models are increasingly used to assist scientific reading, but existing evaluation methods often fail to detect whether answers are supported by verifiable citations. We introduce ResearchQA, a benchmark of 6,211 single-paper question-answer pairs from 494 open-access papers spanning eight domains and four question types: lookup, comprehension, multi-hop, and adversarial. ResearchQA is designed for citation-grounded evaluation: it permits multiple valid supporting passages for a claim and rewards grounded refusal when the source paper does not support an answer. We evaluate eight leading closed- and open-weight models in a citation-grounded chat-with-paper setting using a deterministic citation matcher and an LLM-based rubric evaluator. Citation-based metrics separate systems more clearly than LLM-evaluator scores: section coverage and citation accuracy vary substantially across models, while evaluator scores remain tightly compressed. We further find that open-weight models approach the best closed-model citation accuracy while achieving 3 to 6 times lower per-example latency. We release the benchmark, evaluation harness, and evaluator prompt.
By evaluating paper retrieval, evidence grounding, and answer accuracy separately, LitTraceQA provides a testbed for scientific QA systems that produce verifiable answers rather than unsupported summaries.
The spread of scientific knowledge depends on how researchers discover and cite prior work. Large language models (LLMs) now add a new layer to this process, but their alignment with human citation practices across domains remains unclear. Here, we compare human citations with GPT-4ogenerated reference suggestions produced from paper metadata and abstracts. Analyzing 274, 951 generated references for 10, 000 focal papers, we find that LLMs systematically reinforce the Matthew effect by favoring highly cited papers, with field-specific variation in the rate at which generated references match real papers in bibliometric databases. Generated references diverge from groundtruth reference lists by favoring more recent papers, shorter titles, and smaller author teams. Yet they remain semantically aligned with focal-paper content at levels comparable to human references, reproduce similar local citation-network structure, and reduce author self-citations. These results show that LLMs can generate content-relevant bibliographic suggestions from parametric knowledge alone, but that they also amplify dominant citation patterns. As such tools become routine in research workflows, they may reshape how scientific communities discover, prioritize, and build on prior work.
A. Algaba, Vincent Holst, Floriano Tori et al.· Quantitative Science Studies· 1 citation
Reinforcement learning increasingly relies on an LLM judge to score each rubric criterion, and that judge acts as the reward model during training. Before such a signal can be trusted, we need to know how capable the judge must be and how biased it is. We study this calibration question for citation quality in deep-research systems, where a search-grounded LLM must support each claim it writes with a cited source. Citation quality is a structured rubric task in which each attribution-citation pair is judged along two dimensions that require an LLM, source relevance and factual support. On an adversarial long-form benchmark, we score 8 off-the-shelf LLM judges from 3 model families against gold labels over 1,248 rubric decisions, all of which were human-reviewed and 378 of which were hard cases adjudicated from judge disagreements. Cheaper judges remain competitive across both dimensions, with GPT-5-mini attaining the strongest source-relevance pass-class F1 at 0.908 ($\kappa$=0.636), while on factual support the judges are statistically indistinguishable (overlapping confidence intervals), so no single model dominates. At comparable F1, the judges still differ substantially in pass-rate drift, false positive rate, and false negative rate. Scalar F1 obscures this directional bias, yet it is exactly what a downstream reinforcement learning loop would reinforce. Calibrating the judge is therefore a prerequisite for using citation rubrics as reward signals, and our results show that this calibration does not require the most expensive available model.
E. Leung, Elias Lumer, Corey Feld et al.· 0 citations
Long-form question answering increasingly relies on retrieved evidence to make LLM outputs verifiable, with inline citations tracing claims to source documents. However, existing systems often attach citations that are topically related but insufficient to support their claims. We identify attribution ambiguity as a structural challenge: end-to-end generation must implicitly resolve combinatorial claim--document assignments, obscuring evidential boundaries and increasing the risk of evidence-boundary overrun, where claims exceed cited support. To address this challenge, we propose CAGE (Cognitive Attribution Graphs for Citation Generation), a two-stage framework that introduces an explicit cognitive attribution map before answer generation. CAGE first trains a plug-and-play Cognitive Map Induction Model to construct answer-centered support subgraphs, aligning each semantic answer unit with supporting documents through explicit relations. A Structured Citation Reasoning Model then realizes these units as sentence-level claims with map-aligned citations. Experiments on ASQA, ELI5, and ExpertQA show that CAGE achieves state-of-the-art performance, demonstrating the effectiveness of attribution-space contraction and map-guided citation generation.
Zhichao Yan, Shizhao Li, Jiapu Wang et al.· 1 citation
In this paper, we examine unsupervised, content-based collaboration recommendations using publication text in scholarly settings. We compare three families of methods: a TF-IDF baseline, topic-based models (LDA and BERTopic, including clone variants), and embedding-based retrieval using SciBERT with Faiss. To evaluate model behavior beyond simple lexical matching, we introduce a constrained setting where publication overlap between researchers is partially removed while still using historical co-authorship as proxy ground truth for post-hoc evaluation. Results show clear differences across methods. TF-IDF performs best under full information but drops significantly as overlap is reduced. In contrast, topic-based and embedding-based approaches show more stable performance, suggesting they capture broader distributional similarities, rather than relying only on direct lexical overlap. We also examine explainability through two perspectives: intrinsic topic-based explanations and post-hoc, retrieval-based explanations generated using language models. These provide complementary trade-offs between transparency and human readability.
Md Asaduzzaman Noor, John W. Sheppard, Jason A. Clark· 0 citations
Financial question answering is typically evaluated by answer correctness, yet in SEC filings a plausible and even numerically correct answer can be grounded in the wrong evidence. Similar facts and disclosures recur across sections of a filing, across reporting periods of the same firm, and across comparable firms. FinRank targets this provenance-sensitive retrieval problem by requiring systems to identify evidence for the intended entity, reporting period, and disclosure context. The benchmark contains 1185 manually authored question-answer records over the 10-K and 10-Q filings of 22 companies. Each record includes a reference answer, gold supporting passages, and hand-curated hard negatives drawn from confusable passages within filings, across reporting periods, and across comparable firms. FinRank evaluates passage retrieval, reranking, and hard-negative discrimination as separately measured tasks. Baseline results demonstrate the difficulty of this setting: among the evaluated systems, even a 7B instruction-tuned embedder reaches only 44.8% Recall@10 on the pooled evidence corpus; sub-billion-parameter encoders gain at most 3.5 points over BM25, a finance-adapted embedder trails BM25 by 9.7 points, and pairwise accuracy falls by 13.0-20.5 percentage points when random negatives are replaced with the curated hard negatives. FinRank provides an evidence-first benchmark for developing financial question answering systems that are not only accurate but also grounded in the correct disclosure.
Sasan Mansouri, D. Saad, M. Wahrenburg et al.· 0 citations
