{"id":"cdaa4943-ef53-4135-b309-5d82bd043e88","arxiv_id":"2412.11344","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The paper introduces the first English annotated corpus for extracting factors that influence human trust in AI from scientific text, and shows supervised NER and RE models outperform prompt-based LLMs.","lead":"Researchers built a small annotated dataset that marks words in AI papers describing human, technology, and context factors that affect trust in AI. They benchmarked machine-learning and large-language-model systems on it, finding old-style supervised models still beat prompting-based LLMs on this task.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Test set is the 5-article pilot set used to develop the annotation guidelines, so the benchmark is not on unseen data and cannot support the central claim.","rationale":"The paper's central claim is that trust-in-AI factor extraction 'requires supervised learning which may not be currently feasible with prompt-based LLMs.' The only evidence for this is the benchmark comparison in Tables 2 and 3. If the test set is the same 5 articles used to develop and adjudicate the annotation guidelines, those numbers are not estimates of performance on fresh literature; the benchmark is internally inconsistent, not merely in tension with external consensus. The reader already noted that the test set is non-independent, but treated the weakest assumption as the reliability of inferring relations to a hidden 'trust' concept from a sentence and paragraph. I agree that relation-label reliability is a genuine concern (the initial RE kappa of 0.323 is low, and the reported 93.5 after resolution appears to be percentage agreement, not kappa). However, the non-independent test set is more load-bearing because it invalidates the quantitative basis for the strongest claim regardless of label quality: models are evaluated on documents that shaped the annotation scheme. The dataset may still be a useful resource, and the flaw is fixable by re-annotating a fresh held-out set, so the conditional verdict stands. The concrete check above would settle whether the observed gap survives a clean test set. A release of the data and code would also help, but the fresh test set is the necessary condition.","tokens_in":10931,"tokens_out":9122,"duration_ms":81576,"concrete_test":"Randomly select 10 new articles (or newly retrieved ones matching the same search queries) that were not among the 5 pilot documents and not used in any phase of guideline development. Have the two annotators independently annotate them using only the final guidelines, without discussion or LLM guidance, and compute Cohen's kappa. Then rerun the NER and RE experiments (Tables 2 and 3) on this clean held-out set. If the supervised models still beat GPT-4/ICL by a comparable margin, the central claim is provisionally supported; if the gap shrinks, reverses, or kappa drops well below the reported test-set agreement, the headline conclusion is an artifact of the non-independent test set.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The decisive flaw is that the reported test set is not held out. In §4, the annotation is run in five phases: phase i develops the guidelines on 5 pilot publications; phase ii has both annotators label the same 5 publications; phase iii resolves disagreements and updates the guidelines on those same 5 publications. In §5, the test set is defined as 'samples annotated by two annotators (phase ii and iii)', and Table 1 shows exactly 5 test documents. So the gold test labels come from the documents on which the annotation guidelines were iterated and disagreements adjudicated. The supervised models are trained on the other 22 documents (phases iv–v) and tested on these 5 pilot documents. This is a clear internal inconsistency: the test set participated in the development of the annotation scheme. The abstract conclusion that 'this problem requires supervised learning which may not be currently feasible with prompt-based LLMs' is therefore not supported by a valid out-of-sample benchmark. The small test size (233 sentences, 874 relations) further amplifies the risk that the observed RoBERTa-vs-GPT-4 gap is specific to these particular pilot documents.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Trust in AI, a new English annotated dataset of scientific publications on antecedents of human trust in AI, and formulates two information extraction tasks: named entity recognition of factor and application spans, and relation extraction between factor mentions and the hidden concept of trust. The authors describe a multi-phase annotation process with two annotators and LLM-guided annotation, and benchmark supervised models (RoBERTa, Seq2seq-BERT, BiaffineNER) against prompt-based LLMs (GPT-3.5 for NER, GPT-4 for RE). Based on these benchmarks, the abstract concludes that this extraction problem requires supervised learning and may not currently be feasible with prompt-based LLMs.","tokens_in":11150,"tokens_out":3459,"duration_ms":31602,"significance":"If the benchmark were valid, the paper would provide a useful new resource for the trust-in-AI community and an informative comparison of supervised and prompt-based IE methods in a specialized scientific domain. The dataset itself, with expert-informed annotation guidelines, is a potentially valuable contribution, and the authors are transparent about the annotation procedure and the LLM-guided component. However, the central empirical claim rests on a test set that is not held out from the guideline-development process, which undermines the benchmark conclusion as currently presented. The dataset release (on request) and the detailed baselines are strengths that could make this a solid contribution after the evaluation is made sound.","major_comments":[{"comment":"For the NER few-shot experiments, the manuscript says \"we adopt the prompt template provided by EasyInstruct\" but does not give the exact prompt for the few-shot setting beyond the zero-shot instruction. The prompt shown appears to be for the NER task with a single sentence, while the ICL experiments use various sample selection methods (BM25, KATE, etc.). Please include an example of the full few-shot prompt (with demonstrations) to make the experiments reproducible.","section":"Appendix B (Experimental Setups)"}],"minor_comments":[{"comment":"Please clarify the caption of Table 4 by spelling out the relation-type order in the bracketed distributions.","section":"Table 4"}],"recommendation":"major_revision","confidential_remarks":"To the editor: The central empirical claim is not currently supported because the test set is the same five publications used to develop the annotation guidelines (Section 4 and 5). This is a load-bearing issue, not a stylistic one. The paper also misreports agreement statistics (Section 4), which further weakens confidence in the gold labels. The dataset and supervised baselines are potentially valuable, and a revised version with a proper held-out test set and corrected agreement reporting could meet the bar for publication. However, as it stands, the benchmark cannot support the abstract's conclusion. I recommend major revision with a request for a genuinely independent test set or a clearly justified alternative evaluation protocol."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing to know: this paper delivers the first annotated English dataset for extracting trust-in-AI antecedent factors, and that contribution is real. The annotation process involved domain experts, the guidelines were iterated carefully, and the task formulation — NER plus relation extraction where relations point to a hidden 'trust' concept — is sensible. The dataset fills a genuine gap and could be a useful resource for both NLP and trust-in-AI researchers.\n\nThe soft spots are mostly in the evaluation, and one is load-bearing. The test set is the same five publications used in phases i–iii to develop and revise the annotation guidelines. So the reported RoBERTa-vs-GPT-4 gap is measured on data that the guidelines were tuned on. That is not a valid out-of-sample benchmark, and it undermines the abstract's claim that 'this problem requires supervised learning which may not be currently feasible with prompt-based LLMs.' The test set is also tiny (233 sentences), so the gap could easily be specific to those five documents.\n\nSeparately, the paper conflates Cohen's kappa with percent agreement. Early kappa values are correctly reported as 0.395 and 0.323, but after resolution they report 93.3 and 93.5 and then call them 'kappa values over 90.' Those are clearly percentages, not kappas. This needs a straightforward correction.\n\nTwo more things. The dataset is not publicly released — it's 'available on request' from the first author, which limits its usefulness and makes reproducibility harder. And the LLM comparison is narrow: GPT-4 only for zero-shot relation extraction, GPT-3.5 for few-shot NER, with no prompt variation. So the conclusion that prompt-based LLMs are not viable is broader than what the experiments can support.\n\nWhat is good survives these issues: the dataset itself, if released, would be worth building on. The benchmarks need to be redone with a proper held-out split, the kappa confusion fixed, and the LLM baselines widened. The paper deserves peer review — the contribution is real and the problems are addressable — but it is not ready in its current form. I would engage with it as a dataset paper, not as a benchmark paper, until the evaluation is fixed.","headline":"The dataset is a useful new resource, but the benchmark's test set was used to develop the annotation guidelines, so the headline claim about supervised vs. prompt-based LLMs is not supported as is.","tokens_in":11674,"tokens_out":1561,"would_cite":false,"duration_ms":14148,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Prompt-based LLMs cannot yet replace supervised learning for extracting trust-in-AI factors from scientific papers, a new benchmark dataset shows.","keywords":["trust in AI","information extraction","named entity recognition","relation extraction","large language models","scientific literature annotation","few-shot learning","dataset benchmark"],"falsifier":"Re-annotate the test set's relation labels with access to the full article rather than only the sentence and paragraph; if a substantial share of labels change, the dataset's gold relations do not stand. Alternatively, run the same zero-shot GPT-4 prompt with the full article as context; if it matches or beats RoBERTa-Large's 81.7 ROC-AUC on relation identification, the paper's conclusion that prompt-based LLMs are not currently feasible for this problem would be overturned.","tokens_in":10739,"feed_emoji":"🤖","tokens_out":6538,"duration_ms":53316,"temperature":0.7,"pith_summary":"This paper asks whether large language models can extract the factors that drive human trust in AI from the scientific literature, and answers that, at present, they cannot do so reliably through prompting alone. To test this, the authors build a new annotated dataset, called Trust in AI, covering 27 English publications, and define two extraction tasks: recognizing mentions of applications and of human, technological, and contextual trust factors, and determining how those factors relate to the concept of trust. They benchmark supervised models and prompt-based LLMs, finding that supervised models clearly outperform zero- and few-shot LLMs on both tasks. Their central claim is that this problem requires supervised learning, and that prompt-based LLMs are not yet a feasible substitute for expert annotation in this domain. The dataset is meant to be the first step toward structured, queryable knowledge of trust antecedents for researchers building and evaluating AI systems.","feed_headline":"GPT-4 trails supervised models on trust-in-AI extraction","feed_subtitle":"A new annotated benchmark shows prompt-based LLMs can't yet replace human annotation for factor-trust relations.","key_machinery":"The central object is the Trust in AI dataset and its annotation scheme. Each sentence is paired with its surrounding paragraph as context; entity spans are labeled as application, human factor, technology factor, or context factor; and each relation connects one factor (or a pair of factors) to the concept of trust, which itself is not a mention in the text, so the relation is to a hidden concept. The evaluation is carried by the dataset's split into training, development, and test sets, by a two-stage relation extraction design (first decide whether a relation exists, then classify its type), and by the contrast between supervised fine-tuning and prompt-based LLM inference. The paper's argument stands on the measured gap between these two families of methods on the same annotation scheme.","core_discovery":"The paper's central discovery is a negative empirical result with a new resource attached. It constructs the Trust in AI dataset, which annotates 27 English scientific papers at the sentence level with application names and trust-factor mentions (human, technology, context), plus relations linking one or two factors to the hidden concept of trust, labeled unspecified, null, positive, negative, interaction, or no relation. On this dataset, supervised fine-tuned models such as RoBERTa reach substantially higher scores than prompt-based LLMs: for relation identification, RoBERTa-Large reaches a ROC-AUC of 81.7 while zero-shot GPT-4 reaches 52.2, near the random baseline of 50.5, and few-shot in-context learning for entity recognition reaches only 18.3 micro F1 against 61.2 for RoBERTa. The authors interpret this as evidence that extracting trust antecedents from scientific text is a genuinely hard structured-prediction problem, not one that current prompting techniques can solve with a handful of examples.","pith_inferences":["The paper leaves open whether full-document context rather than sentence-plus-paragraph would let prompt-based LLMs close the gap; a testable extension is to prompt with the complete article and the annotation guideline.","The relation-to-hidden-trust formulation could generalize to other latent constructs in scientific text, such as risk, fairness, or usability, where the target concept is discussed without being named in every sentence.","If supervised learning remains necessary, the cost of annotation in this domain argues for active-learning or human-in-the-loop pipelines that focus annotation effort on the sentences a model is least certain about; this follows from the paper's findings but is not proposed in the paper."],"forward_implications":["A practical consequence for behavioural and computer science teams is that building a structured database of trust antecedents from the literature currently requires a human-annotated training corpus, not just a prompt.","Because the dataset fixes the task definitions, future supervised models can be compared against the reported baselines, and the claim that prompts are insufficient becomes testable.","Relation labels are highly imbalanced and mostly 'no relation', so any practical extraction system needs a reliable relation-identification step before type classification, which is the direct reason the paper adopts its two-step design.","The low LLM-human agreement (Cohen's kappa 12.85 for NER and 5.59 for RE) implies that using LLMs to pre-annotate in this domain would still require substantial manual correction.","Supervised models themselves remain weak on human and contextual factor types, so the annotation scheme exposes a hard subproblem that is not yet solved even with fine-tuning."],"supporting_citations":[{"why":"Supplies the review of empirical research on human trust in AI that motivates the need for structured extraction.","marker":"(Glikson and Woolley, 2020)"},{"why":"Identifies over 450 distinct antecedents of trust, defining the domain scope the dataset must organize.","marker":"(Saßmannshausen et al., 2023)"},{"why":"Provides the in-context learning sample-selection methods and the finding that LLM extraction underperforms on scientific writing, used for the few-shot baselines.","marker":"(Bölücü et al., 2023)"},{"why":"Prior result that GPT-3 in-context learning underperforms for biomedical information extraction, the comparison the paper's LLM benchmarks extend to trust in AI.","marker":"(Gutiérrez et al., 2022)"},{"why":"Supplies the RoBERTa encoder used in the supervised NER and RE baselines.","marker":"(Liu et al., 2019)"},{"why":"Supplies the span-based classifier stacked on RoBERTa for the NER baseline.","marker":"(Zhong and Chen, 2021)"},{"why":"Supplies the Seq2seq-BERT baseline for nested NER.","marker":"(Straková et al., 2019)"},{"why":"Supplies the BiaffineNER dependency-parsing baseline for NER.","marker":"(Yu et al., 2020)"},{"why":"Defines the kappa statistic used to measure inter-annotator agreement in both annotation stages.","marker":"(Cohen, 1960)"},{"why":"Supplies the small-expert-annotation approach that the dataset construction follows.","marker":"(Karimi et al., 2015)"}],"fun_headline_variants":["Supervised models trounce GPT-4 on trust-in-AI extraction","Prompt-based LLMs fail to extract trust factors from literature","Supervised learning outperforms prompting on trust-factor extraction","GPT-4 trails RoBERTa on trust-in-AI relation extraction","Trust-in-AI extraction: few-shot prompting lags supervised models"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The gold relations are annotated under the assumption that a factor mention's relationship to trust can be judged reliably from a single sentence and its paragraph, even when the word 'trust' never appears, and the initial 32.3% agreement between annotators on relations shows that judgment is not straightforward.","fun_headline_variants_meta":{"raw":{"variants":["Supervised models trounce GPT-4 on trust-in-AI extraction","Prompt-based LLMs fail to extract trust factors from literature","Supervised learning outperforms prompting on trust-factor extraction","GPT-4 trails RoBERTa on trust-in-AI relation extraction","Trust-in-AI extraction: few-shot prompting lags supervised models"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000409,"raw_usage":{"total_tokens":2094,"prompt_tokens":890,"completion_tokens":1204,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":506,"completion_tokens_details":{"reasoning_tokens":1126}},"tokens_in":506,"tokens_out":1204,"duration_ms":8104,"temperature":1.0,"reasoning_tokens":1126,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T15:02:02.700994+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-annotate the test set's relation labels with access to the full article rather than only the sentence and paragraph; if a substantial share of labels change, the dataset's gold relations do not stand. Alternatively, run the same zero-shot GPT-4 prompt with the full article as context; if it matches or beats RoBERTa-Large's 81.7 ROC-AUC on relation identification, the paper's conclusion that prompt-based LLMs are not currently feasible for this problem would be overturned.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the review of empirical research on human trust in AI that motivates the need for structured extraction."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the span-based classifier stacked on RoBERTa for the NER baseline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the kappa statistic used to measure inter-annotator agreement in both annotation stages."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the small-expert-annotation approach that the dataset construction follows."}],"review_version":1}