Trustworthiness in healthcare question-answering (QA) systems is important for ensuring patient safety, clinical effectiveness, and user confidence. As large language models (LLMs) become increasingly integrated into medical settings, the reliability of their responses directly influences clinical decision-making and patient outcomes. However, achieving comprehensive trustworthiness in medical QA poses significant challenges due to the inherent complexity of healthcare data, the critical nature of clinical scenarios, and the multifaceted dimensions of trustworthy AI. In this survey, we systematically examine six key dimensions of trustworthiness in medical QA, i.e., Factuality, Robustness, Fairness, Safety, Explainability, and Calibration. We review how each dimension is evaluated in existing LLM-based medical QA systems. We compile and compare major benchmarks designed to assess these dimensions and analyze evaluation-guided techniques that drive model improvements, such as retrieval-augmented grounding, adversarial fine-tuning, and safety alignment. Finally, we identify open challenges—such as scalable expert evaluation, integrated multi-dimensional metrics, and real-world deployment studies—and propose future research directions to advance the safe, reliable, and transparent deployment of LLM-powered medical QA.
论文检索
输入标题、作者或关键词，从 1,386 篇学术成果中精准定位
Are LLMs Better than Reported? Detecting Label Errors and Mitigating Their Effect on Model Performance
PDF ↗NLP benchmarks rely on standardized datasets for training and evaluating models and are crucial for advancing the field. Traditionally, expert annotations ensure high-quality labels; however, the cost of expert annotation does not scale well with the growing demand for larger datasets required by modern models. While crowd-sourcing provides a more scalable solution, it often comes at the expense of annotation precision and consistency. Recent advancements in large language models (LLMs) offer new opportunities to enhance the annotation process, particularly for detecting label errors in existing datasets. In this work, we consider the recent approach of LLM-as-a-judge, leveraging an ensemble of LLMs to flag potentially mislabeled examples. We conduct a case study on four factual consistency datasets from the TRUE benchmark, spanning diverse NLP tasks, and on SummEval, which uses Likert-scale ratings of summary quality across multiple dimensions. We empirically analyze the labeling quality of existing datasets and compare expert, crowd-sourced, and LLM-based annotations in terms of the agreement, label quality, and efficiency, demonstrating the strengths and limitations of each annotation method. Our findings reveal a substantial number of label errors, which, when corrected, induce a significant upward shift in reported model performance. This suggests that many of the LLMs’ so-called mistakes are due to label errors rather than genuine model failures. Additionally, we discuss the implications of mislabeled data and propose methods to mitigate them in training to improve performance.
Pixels Versus Priors: Controlling Knowledge Priors in Vision-Language Models through Visual Counterfacts
PDF ↗Multimodal Large Language Models (MLLMs) perform well on tasks such as visual question answering, but it remains unclear whether their reasoning relies more on memorized world knowledge or on the visual information present in the input image. To investigate this, we introduce Visual CounterFact, a new dataset of visually-realistic counterfactuals that put world knowledge priors (e.g, red strawberry) into direct conflict with visual input (e.g, blue strawberry). Using Visual CounterFact, we show that model predictions initially reflect memorized priors, but shift toward visual evidence in mid-to-late layers. This dynamic reveals a competition between the two modalities, with visual input ultimately overriding priors during evaluation. To control this behavior, we propose Pixels Versus Priors (PvP) steering vectors, a mechanism for controlling model outputs toward either world knowledge or visual input through activation-level interventions. On average, PvP successfully shifts 99.3% of color and 80.8% of size predictions from priors to counterfactuals. Together, these findings offer new tools for interpreting and controlling factual behavior in multimodal models.
Large language models use high-dimensional latent spaces to encode and process textual information. Much work has investigated how the conceptual content of words translates into geometrical relationships between their vector representations. Fewer studies analyze how the cumulative information of an entire prompt becomes condensed into individual embeddings under the action of transformer layers. We use literary pieces to show that information about intangible, rather than factual, aspects of the prompt are contained in deep representations. We observe that short excerpts (10 - 100 tokens) from different novels separate in the latent space independently from what next-token prediction they converge towards. Ensembles from books from the same authors are much more entangled than across authors, suggesting that embeddings encode stylistic features. This geometry of style may have applications for authorship attribution and literary analysis, but most importantly reveals the sophistication of information processing and compression accomplished by language models.
DnDScore: Decontextualization and Decomposition for Factuality Verification in Long-Form Text Generation
PDF ↗The decompose-then-verify strategy for verification of Large Language Model (LLM) generations decomposes claims that are then independently verified. Decontextualization augments text (claims) to ensure it can be verified outside of the original context, enabling reliable verification. While decomposition and decontextualization have been explored independently, their interactions in a complete system have not been investigated. Their conflicting purposes can create tensions: decomposition isolates atomic facts while decontextualization inserts relevant information. Furthermore, a decontextualized subclaim presents a challenge to the verification step: what part of the augmented text should be verified as it now contains multiple atomic facts? We conduct an evaluation of different decomposition, decontextualization, and verification strategies and find that the choice of strategy matters in the resulting factuality scores. Additionally, we introduce DnDScore, a decontextualization aware verification method that validates subclaims in the context of contextual information.
Large Language Models (LLMs) have demonstrated near-human performance in summarization tasks based on traditional metrics such as ROUGE and BERTScore. However, these metrics do not adequately capture critical aspects of summarization quality, such as factual accuracy, particularly for long narratives (>100K tokens). Recent advances, such as LLM-as-a-Judge, address the limitations of metrics based on lexical similarity but still exhibit factual inconsistencies, especially in understanding character relationships and states. In this work, we introduce NarrativeFactScore (NFS), the first “Agent-as-a-Judge” framework that evaluates and refines factuality in narrative summarization. By leveraging a Character Knowledge Graph (CKG) extracted from input narrative, NarrativeFactScore evaluates the factuality and provides actionable guidance for refinement, such as identifying missing or erroneous facts. Our experimental results demonstrate that constructing the CKG enables reasoning with 1/3 of the factuality computation used in the prior approach, and achieve three times higher correlation with human judgments. Furthermore, refinement with actionable guidance improves the quality of the summary.
Iterative evaluation of LLMs during training is essential to ensure expected capability development, but can be time- and compute-intensive. While NLU tasks, where the model selects from fixed answer choices, are cheap to evaluate, essential capabilities like reasoning and code generation rely on the more time-consuming NLG (token-by-token generation) format. In this work, our aim is to decrease the computational burden of NLG benchmarks in order to enable monitoring crucial LLM capabilities during model training. We reformulate generative tasks into computationally cheaper NLU alternatives. We test the performance correlation between the original and reformulated tasks using 8 LMs of various sizes and 4 capabilities: mathematical reasoning, code generation, factual knowledge and reading comprehension. Our results show a strong correlation between task formats, supporting capability assessment via cheaper alternatives and achieving over 35x average reduction in evaluation time. Our project is available at: https://github.com/Fraunhofer-IIS/EvalShortcut
Retrieval Augmented Generation (RAG) frameworks can improve the factual accuracy of large language models (LLMs) by integrating external knowledge from retrieved documents, thereby overcoming the limitations of models’ static intrinsic knowledge. However, these systems are susceptible to adversarial attacks that manipulate the retrieval process by introducing documents that are adversarial yet semantically similar to the query. Notably, while these adversarial documents resemble the query, they exhibit weak similarity to benign documents in the retrieval set. Thus, we propose a simple yet effective **G**raph-based **R**eranking against **A**dversarial **D**ocument **A**ttacks (GRADA) framework aiming at preserving retrieval quality while significantly reducing the success of adversaries. Our study evaluates the effectiveness of our approach through experiments conducted on six LLMs: GPT-3.5-Turbo, GPT-4o, Llama3.1-8b-Instruct, Llama3.1-70b-Instruct, Qwen2.5-7b-Instruct and Qwen2.5-14b-Instruct. We use three datasets to assess performance, with results from the Natural Questions dataset demonstrating up to an 80% reduction in attack success rates while maintaining minimal loss in accuracy.
Unsupervised hallucination detection aims to identify hallucinated content generated by large language models (LLMs) without relying on labeled data. While unsupervised methods have gained popularity by eliminating labor-intensive human annotations, they frequently rely on proxy signals unrelated to factual correctness. This misalignment biases detection probes toward superficial or non-truth-related aspects, limiting generalizability across datasets and scenarios. To overcome these limitations, we propose IRIS, an unsupervised hallucination detection framework, leveraging internal representations intrinsic to factual correctness. IRIS prompts the LLM to carefully verify the truthfulness of a given statement, and obtain its contextualized embedding as informative features for training. Meanwhile, the uncertainty of each response is considered a soft pseudolabel for truthfulness. Experimental results demonstrate that IRIS consistently outperforms existing unsupervised methods. Our approach is fully unsupervised, computationally low cost, and works well even with few training data, making it suitable for real-time detection.
Multilingual question answering (QA) systems must ensure factual consistency across languages, especially for objective queries such as What is jaundice?, while also accounting for cultural variation in subjective responses. We propose MIND, a user-in-the-loop fact-checking pipeline to detect factual and cultural discrepancies in multilingual QA knowledge bases. MIND highlights divergent answers to culturally sensitive questions (e.g., Who assists in childbirth?) that vary by region and context. We evaluate MIND on a bilingual QA system in the maternal and infant health domain and release a dataset of bilingual questions annotated for factual and cultural inconsistencies. We further test MIND on datasets from other domains to assess generalization. In all cases, MIND reliably identifies inconsistencies, supporting the development of more culturally aware and factually consistent QA systems.
Large Language Models (LLMs) have demonstrated impressive reasoning abilities, but their generated rationales often suffer from issues such as reasoning inconsistency and factual errors, undermining their reliability. Prior work has explored improving rationale quality via multi-reward fine-tuning or reinforcement learning (RL), where models are optimized for diverse objectives. While effective, these approaches train the model in a fixed manner and do not have any inference-time adaptability, nor can they generalize reasoning requirements for new test-time inputs. Another approach is to train specialized reasoning experts using reward signals and use them to improve generation at inference time. Existing methods in this paradigm are limited to using only a single expert and cannot improve upon multiple reasoning aspects. To address this, we propose MIXIE, a novel inference-time expert-mixing framework that dynamically determines mixing proportions for each expert, enabling contextualized and flexible fusion. We demonstrate the effectiveness of MIXIE on improving chain-of-thought reasoning in LLMs by merging commonsense and entailment reasoning experts finetuned on reward-filtered data. Our approach outperforms existing baselines on three question-answering datasets: StrategyQA, CommonsenseQA, and ARC, highlighting its potential to enhance LLM reasoning with efficient, adaptable expert integration.
The reliability of large language models remains a critical challenge, particularly due to their susceptibility to hallucinations and factual inaccuracies during text generation. Existing solutions either underutilize models’ self-correction with preemptive strategies or use costly post-hoc verification. To further explore the potential of real-time self-verification and correction, we present Dynamic Self-Verify Decoding (DSVD), a novel decoding framework that enhances generation reliability through real-time hallucination detection and efficient error correction. DSVD integrates two key components: (1) parallel self-verification architecture for continuous quality assessment, (2) dynamic rollback mechanism for targeted error recovery. Extensive experiments across five benchmarks demonstrate DSVD’s effectiveness, achieving significant improvement in truthfulness (Quesetion-Answering) and factual accuracy (FActScore). Results show the DSVD can be further incorporated with existing faithful decoding methods to achieve stronger performance. Our work establishes that real-time self-verification during generation offers a viable path toward more trustworthy language models without sacrificing practical deployability.
Large Language Models (LLMs) are widely used for temporal prediction, but their reliance on pretraining data raises contamination concerns, as accurate predictions on pre-cutoff test data may reflect memorization rather than reasoning, leading to an overestimation of their generalization capability. With the recent emergence of prompting-based unlearning techniques, a natural question arises: Can LLMs be prompted to simulate an earlier knowledge cutoff? In this work, we investigate the capability of prompting to simulate earlier knowledge cutoff in LLMs. We construct three evaluation datasets to assess the extent to which LLMs can forget (1) direct factual knowledge, (2) semantic shifts, and (3) causally related knowledge. Results demonstrate that while prompt-based simulated knowledge cutoffs show effectiveness when directly queried with the information after that date, they struggle to induce forgetting when the forgotten content is not directly asked but causally related to the query. These findings highlight the need for more rigorous evaluation settings when applying LLMs for temporal prediction tasks. The full dataset and evaluation code are available at https://github.com/gxx27/time_unlearn.
The quality of supervised fine-tuning (SFT) data is crucial for the performance of large multimodal models (LMMs), yet current data enhancement methods often suffer from factual errors and hallucinations due to inadequate visual perception. To address this challenge, we propose VERITAS, a pipeline that systematically integrates vision priors and multiple state-of-the-art LMMs with statistical methods to enhance SFT data quality. VERITAS leverages visual recognition models (RAM++) and OCR systems (PP-OCRv4) to extract structured vision priors, which are combined with images, questions, and answers. Three LMMs (GPT-4o, Gemini-2.5-Pro, Doubao-1.5-pro) evaluate the original answers, providing critique rationales and scores that are statistically fused into a high-confidence consensus score serving as ground truth. Using this consensus, we train a lightweight critic model via Group Relative Policy Optimization (GRPO), enhancing reasoning capabilities efficiently. Each LMM then refines the original answers based on the critiques, generating new candidate answers; we select the highest-scoring one as the final refined answer. Experiments across six multimodal benchmarks demonstrate that models fine-tuned with data processed by VERITAS consistently outperform those using raw data, particularly in text-rich and fine-grained reasoning tasks. Our critic model exhibits enhanced capability comparable to state-of-the-art LMMs while being significantly more efficient. We release our pipeline, datasets, and model checkpoints to advance research in multimodal data optimization.
LLM-Driven Completeness and Consistency Evaluation for Cultural Heritage Data Augmentation in Cross-Modal Retrieval
PDF ↗Cross-modal retrieval is essential for interpreting cultural heritage data, but its effectiveness is often limited by incomplete or inconsistent textual descriptions, caused by historical data loss and the high cost of expert annotation. While large language models (LLMs) offer a promising solution by enriching textual descriptions, their outputs frequently suffer from hallucinations or miss visually grounded details. To address these challenges, we propose C^3, a data augmentation framework that enhances cross-modal retrieval performance by improving the completeness and consistency of LLM-generated descriptions. C^3 introduces a bidirectional validation mechanism to assess semantic coverage using both visual cues and language-model outputs. Furthermore, to mitigate factual inconsistencies, we formulate a Markov Decision Process to supervise Chain-of-Thought reasoning, guiding consistency verification through adaptive query control. Experiments on the cultural heritage dataset CulTi and general benchmarks MSCOCO and Flickr30K demonstrate that C^3 achieves state-of-the-art performance in both fine-tuned and zero-shot settings.
FacLens: Transferable Probe for Foreseeing Non-Factuality in Fact-Seeking Question Answering of Large Language Models
PDF ↗Despite advancements in large language models (LLMs), non-factual responses still persist in fact-seeking question answering. Unlike extensive studies on post-hoc detection of these responses, this work studies non-factuality prediction (NFP), predicting whether an LLM will generate a non-factual response prior to the response generation. Previous NFP methods have shown LLMs’ awareness of their knowledge, but they face challenges in terms of efficiency and transferability. In this work, we propose a lightweight model named Factuality Lens (FacLens), which effectively probes hidden representations of fact-seeking questions for the NFP task. Moreover, we discover that hidden question representations sourced from different LLMs exhibit similar NFP patterns, enabling the transferability of FacLens across different LLMs to reduce development costs. Extensive experiments highlight FacLens’s superiority in both effectiveness and efficiency.
Adapting cultural values in Large Language Models (LLMs) presents significant challenges, particularly due to biases and data limitations. Previous work aligns LLMs with different cultures using survey data, primarily from the World Values Survey (WVS). However, it remains unclear whether this approach effectively captures cultural nuances or produces distinct cultural representations for tasks like offensiveness classification. In this paper, we systematically investigate WVS-based training for cultural value adaptation and find that relying solely on survey data can homogenize cultural norms and interfere with factual knowledge. To address these issues, we propose augmenting WVS with encyclopedic and scenario-based cultural narratives from Wikipedia and NormAd. Our experiments across multiple cultures show that this approach captures more enhances differentiated cultural values and improves downstream classification performances.
VeriFact: Enhancing Long-Form Factuality Evaluation with Refined Fact Extraction and Reference Facts
PDF ↗Large language models (LLMs) excel at generating long-form responses, but evaluating their factuality remains challenging due to complex inter-sentence dependencies within the generated facts. Prior solutions predominantly follow a decompose-decontextualize-verify pipeline but often fail to capture essential context and miss key relational facts. In this paper, we introduce VeriFact, a factuality evaluation framework designed to enhance fact extraction by identifying and resolving incomplete and missing facts to support more accurate verification results. Moreover, we introduce FactRBench , a benchmark that evaluates both precision and recall in long-form model responses, whereas prior work primarily focuses on precision. FactRBench provides reference fact sets from advanced LLMs and human-written answers, enabling recall assessment. Empirical evaluations show that VeriFact significantly enhances fact completeness and preserves complex facts with critical relational information, resulting in more accurate factuality evaluation. Benchmarking various open- and close-weight LLMs on FactRBench indicate that larger models within same model family improve precision and recall, but high precision does not always correlate with high recall, underscoring the importance of comprehensive factuality assessment.
The increasing acceptance of large language models (LLMs) as an alternative to knowledge sources marks a significant paradigm shift across various domains, including time-sensitive fields such as law, healthcare, and finance. To fulfill this expanded role, LLMs must not only be factually accurate but also demonstrate consistency across temporal dimensions, necessitating robust temporal reasoning capabilities. Despite this critical requirement, efforts to ensure temporal consistency in LLMs remain scarce including noticeable absence of endeavors aimed at evaluating or augmenting LLMs across temporal references in time-sensitive inquiries. In this paper, we seek to address this gap by introducing a novel benchmark entitled temporal referential consistency, accompanied by a resource TEMP-ReCon designed to benchmark a wide range of both open-source and closed-source LLMs with various linguistic contexts characterized by differing resource richness (including English, French, and Romanian). The findings emphasis that LLMs do exhibit insufficient temporal referent consistency. To address this, we propose , a reasoning path alignment-based model that aims to enhance the temporal referential consistency of LLMs. Our empirical experiments substantiate the efficacy of UnTRaP compared to several baseline models.
REACT: Representation Extraction And Controllable Tuning to Overcome Overfitting in LLM Knowledge Editing
PDF ↗Large language model editing methods frequently suffer from overfitting, wherein factual updates can propagate beyond their intended scope, overemphasizing the edited target even when it’s contextually inappropriate. To address this challenge, we introduce REACT (Representation Extraction And Controllable Tuning), a unified two-phase framework designed for precise and controllable knowledge editing. In the initial phase, we utilize tailored stimuli to extract latent factual representations and apply Principal Component Analysis with a simple learnbale linear transformation to compute a directional “belief shift” vector for each instance. In the second phase, we apply controllable perturbations to hidden states using the obtained vector with a magnitude scalar, gated by a pre-trained classifier that permits edits only when contextually necessary. Relevant experiments on EVOKE benchmarks demonstrate that REACT significantly reduces overfitting across nearly all evaluation metrics, and experiments on COUNTERFACT and MQuAKE shows that our method preserves balanced basic editing performance (reliability, locality, and generality) under diverse editing scenarios.