Contextual Candor: Enhancing LLM Trustworthiness Through Hierarchical Unanswerability Detection

Antonio Carlos Rivera; Steven Robinson

REVIEW 4 major objections 5 minor 28 references

Contextual Candor: Enhancing LLM Trustworthiness Through Hierarchical Unanswerability Detection

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper's RUL method trains LLMs to detect unanswerable questions and produce helpful refusals, reporting ranking-level accuracy of 0.910 and a refusal rate of 0.920 on its own ECA dataset.

desk verdict A coherent idea about training LLMs to refuse unanswerable queries, but every empirical claim rests on an unreleased dataset, so the paper cannot be verified as-is. read the letter →

arxiv 2506.01104 v1 pith:RTKGXIQC submitted 2025-06-01 cs.CL

Steven Robinson , Antonio Carlos Rivera This is my paper

classification cs.CL

keywords ReinforcedUnanswerabilityLearningunanswerablequestiondetectionrefusalresponsegenerationRLHFconversationalansweringhierarchicalanswerabilitylabelsECAdatasetattention-weightedaggregation

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that a large language model can be trained to recognize when a question cannot be answered from the given context, and to refuse helpfully, rather than needing a separate classifier bolted on after generation. Its method, RUL, adds a classification head to the LLM and trains it in two stages: supervised fine-tuning on the paper's ECA dataset, which carries hierarchical answerability labels and ground-truth refusal responses, followed by reinforcement learning from human feedback to make refusals clearer and more useful. The reported result is better unanswerability detection at sentence, paragraph, and ranking levels than BERT and prompting baselines, plus a 0.920 refusal rate for unanswerable questions. A sympathetic reader would care because the ability to say "I don't know" is one of the core missing pieces in making conversational AI trustworthy.

What carries the argument

RUL's load-bearing component is the unanswerability prediction head, which maps the LLM's pooled representation $h_{CLS}$ to a probability through a sigmoid, $\hat{y} = \sigma(W_{cls} h_{CLS} + b_{cls})$, turning the generative model into a joint classifier-generator. Around it, attention-weighted aggregation computes paragraph-level scores $\hat{y}_P = \sum_k \alpha_k \hat{y}_k$ and ranking-level scores $\hat{y}_D = \sum_m \beta_m \hat{y}_{P_m}$ from sentence-level predictions, so the model can weigh which pieces of evidence matter at each granularity. The two-stage training combines a binary cross-entropy classification loss with a negative log-likelihood generation loss, then optimizes refusal quality through RLHF with a KL-regularized policy objective. The hierarchy and the two-stage loss together carry the argument; ablations attribute 3.5 ranking-level accuracy points to the attention mechanism and 2 points to RLHF.

What would settle it

Run RUL and the same baselines on a publicly available unanswerable question-answering benchmark, such as SQuAD 2.0, using the paper's full training recipe; alternatively, release the ECA dataset with annotation guidelines and compute inter-annotator agreement on its hierarchical labels. If the reported ranking-level accuracy (0.910) and refusal rate (0.920) fall to baseline levels on public data, or if independent annotators cannot reproduce the ECA labels, the central claim is not supported.

Watch

Extended reading notes

Core claim

RUL claims that a language model can be made to detect unanswerable questions intrinsically rather than through an external module. The model appends a sigmoid classification head to the LLM's pooled representation, and hierarchical attention aggregates sentence-level scores into paragraph-level and ranking-level answerability scores. Stage 1 supervises both classification and generation on the ECA dataset, whose ground-truth refusals teach the model to say why it cannot answer; Stage 2 uses reinforcement learning from human feedback to make refusals more helpful. The reported outcome is that RUL reaches 0.840 sentence-level, 0.945 paragraph-level, and 0.910 ranking-level detection accuracy, refuses 0.920 of unanswerable queries, and is rated 4.6 out of 5 for refusal helpfulness by human annotators.

Load-bearing premise

The load-bearing premise is that the paper's own Enhanced-CAsT-Answerability (ECA) dataset, which is provisionally named and never released, has valid hierarchical answerability labels and ground-truth refusals; if those labels do not match real user judgments, every reported accuracy and refusal-rate number becomes uninterpretable.

Editorial extensions

If this is right

On the paper's ECA test set, RUL raises ranking-level unanswerability detection accuracy to 0.910, compared with 0.829 for the BERT mean-aggregation baseline and 0.712 for the best generative baseline.
RUL generates refusals for 0.920 of unanswerable questions, versus 0.685 for the SQuAD 2.0 fine-tuned LLM baseline, directly reducing the chance of a hallucinated answer.
Answerable questions do not suffer: RUL's answer F1 is 0.785, above the 0.721 of the best generative baseline.
Ablations attribute a 3.5-point ranking-level accuracy gain to attention-weighted aggregation and a further 2-point gain to the RLHF stage, indicating both components carry the result.
Human annotators rate RUL's refusals 4.6 out of 5 for helpfulness and appropriateness versus 2.5 out of 5 for the SQuAD 2.0 baseline, the largest measured gap in the human evaluation.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Beyond the paper: the attention-weighted answerability score could double as a confidence gate, letting a deployed model abstain or call for human help whenever the ranking-level score falls below a threshold; that turns detection into selective answering and is directly testable with the released model.
Beyond the paper: because ECA has no released labels or annotation protocol, the cleanest check of RUL's claim is to run the identical two-stage recipe on a public unanswerable benchmark such as SQuAD 2.0, where the paper already includes a SQuAD 2.0 fine-tuned baseline; this would separate the method's effect from the dataset's.
Beyond the paper: the large gap in human-rated helpfulness (4.6 vs. 2.5) suggests RLHF aimed at refusal informativeness, not just refusal occurrence, is where the practical trust gain lives; a natural extension is to measure whether this transfers to out-of-domain queries or low-confidence retrieval.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The short version: the paper proposes a sensible way to make LLMs detect unanswerable questions and produce helpful refusals — train a classification head on the LLM, pool hierarchically with attention, and then RLHF the refusal quality. That idea is fine. But the empirical evidence is a closed loop: the evaluation uses only the authors' own ECA dataset, which is not released, not described with annotation guidelines, and has no inter-annotator agreement. Every table number is uninterpretable without that dataset.

To its credit, the method section is reasonably clear, and the paper correctly identifies Lajewska and Balog (ECIR 2024) as the closest prior work. The ablation structure — removing attention and removing RLHF — is the right way to ask whether the components matter, even if the numbers are just point estimates.

The problems are serious. First, the dataset. ECA is 'provisionally named' in Section IV-A1, which suggests even the authors haven't finalized it. There is no release, no documentation of the annotation task, no IAA, and no external validation. That alone sinks the current empirical claims. Second, the paper never says which base LLM was used, what hyperparameters (lambda_cls, lambda_gen, tau, beta) were set, or how the [CLS] token is obtained from a generative model — decoder-only transformers don't naturally expose a [CLS] representation. Third, the related-work citations are wrong in a way that suggests carelessness: [6]-[8] are not the Transformer papers, [11] is not GPT-3, and [28] is not the Ouyang et al. RLHF paper. None of this is fatal to the idea, but it makes the paper hard to trust.

Who is this for? Someone interested in the design space of unanswerability detection might want to replicate the idea if the dataset ever appears. But as a submission, the central evidence is missing. If the authors released ECA, named the base model, and included error bars, this could be a reasonable workshop paper. As-is, it should not go to review.

I would not send this to a serious referee. The right response is a desk reject with an invitation to resubmit with the dataset, code, and full experimental details.

Referee Report

4 major / 5 minor

Summary. The paper proposes Reinforced Unanswerability Learning (RUL), a hybrid training paradigm that augments a base LLM with a discriminative answerability prediction head and hierarchical attention aggregation at sentence, paragraph, and ranking levels. The training pipeline consists of supervised fine-tuning on a newly introduced dataset, Enhanced-CAsT-Answerability (ECA), with hierarchical answerability labels and ground-truth refusal responses, followed by reinforcement learning with human feedback (RLHF) to refine refusal quality. Experiments on ECA report accuracy gains (e.g., ranking-level 0.910 vs. 0.829 for a BERT baseline) and higher refusal rates (0.920), plus human evaluation showing improved helpfulness and appropriateness. The paper concludes that RUL yields more trustworthy conversational AI.

Significance. If the reported results were fully reproducible, the paper would address a practically important problem, and the combination of an answerability head with hierarchical attention and RLHF-refined refusals is a plausible design direction. However, the contribution is currently unverifiable: the ECA dataset is not released or documented, no code, model checkpoints, hyperparameters, or base model are identified, and the tables contain only point estimates. The empirical claims therefore cannot currently support the paper's conclusions.

major comments (4)

[Section IV-A1, Tables I-VII] The evaluation hinges entirely on the Enhanced-CAsT-Answerability (ECA) dataset, which is introduced as 'provisionally named' and is never released. The paper does not describe annotation guidelines, annotator qualifications, inter-annotator agreement, or a validation of the hierarchical labels against an external standard, nor does it report dataset sizes or split statistics. Without access to ECA, the reader cannot verify the sentence-, paragraph-, and ranking-level labels, the ground-truth refusal responses, or the no-leakage claim, so every accuracy and refusal-rate number in Tables I-VII is untestable. This is a load-bearing limitation for the central claim.
[Section IV-A2, IV-B] No experimental configuration is specified: the base LLM is never named, and key hyperparameters such as lambda_cls, lambda_gen, tau, and beta are never given values. No code or checkpoints are provided, and all reported numbers are single point estimates without error bars, confidence intervals, or significance tests. Consequently the abstract's claim of 'significantly higher accuracy' and the comparisons in Tables I and II cannot be evaluated statistically or reproduced.
[Section III-B] The architecture description assumes a [CLS] token or equivalent pooled representation is available in the base LLM, but the paper never specifies the base model family. Most modern decoder-only LLMs do not provide a [CLS] output, and the phrase 'or the pooled output if using a different architecture' does not define how pooling would be implemented for an autoregressive model. Equations (1) through (7) also require sentence-level encoder representations for the attention mechanisms, yet the paper does not explain how a generative LLM is adapted to produce independent sentence embeddings. This makes the method specification incomplete at a load-bearing point.
[Section IV-A3, IV-D] The refusal-quality metrics ('Appropriateness Score', 'Informativeness Score', and the human Likert ratings) are scored under a rubric that is not published, and the human evaluation reports no number of annotators, no sample size, no inter-annotator agreement, and no statistical test. The 'random subset' of responses is not specified, and only one generative baseline is included in the human comparison. The claim of a 'dramatic and significant improvement' in helpfulness is therefore not supported by the evidence presented.

minor comments (5)

[References [6]-[8], [11], [28]] Several citations do not match the claims that cite them: [6] is a DAC layout-pattern paper, not the Transformer; [11] is a video-language paper, not GPT-3; [28] is 'InstructPatentGPT', not Ouyang et al.'s RLHF work; and [19] duplicates [3].
[Section I and IV-A1] The phrase 'provisionally named' in Section IV-A1 conflicts with the Abstract and Introduction presenting ECA as a completed benchmark; a stable name and a complete dataset description are needed.
[Section III-B2a] The cross-reference 'described in Section 2.2.1' should point to the actual section (III-B1).
[Section IV-A1] The paper does not report basic dataset statistics, such as number of conversations, questions, sentences, paragraphs, or train/validation/test sizes, and does not describe the provenance and filtering of the 'supplementary data from publicly available large-scale question answering datasets'.
[Various] There are numerous typos and incomplete phrases, for example 'evolving multi-capabilities' in Section I and 'the inherent difficulties' with no object in the same paragraph; a thorough proofread is needed.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the training equations and reported metrics are not definitionally tied, and the unreleased ECA benchmark is a reproducibility limitation rather than a circular reduction.

full rationale

I walked the derivation chain from Eq. (1) through Eq. (12). The prediction head, attention-weighted aggregation, supervised losses, and RLHF objective are all defined independently of the reported test accuracies; no equation is defined in terms of the evaluation results. The only candidate concern is that the headline comparisons in Tables I-VII are computed on the authors' own ECA dataset, which is described as 'provisionally named' (Sections I and IV-A1) and whose release is not indicated. However, the paper states the dataset is 'strategically partitioned into training, validation, and test sets, ensuring strict no-overlap in query-context pairs to prevent data leakage,' so the test numbers are not the training target by construction. Lack of release and missing annotation details make the results unverifiable, but unverifiability is not equation-level circularity. There are no load-bearing self-citations by Robinson and Rivera; the cited works are by other authors. I therefore find no circular step and assign score 0.

Assumptions & free parameters 4 free parameters · 4 assumptions · 2 invented entities

The central claim depends on one self-created dataset, one informally defined evaluation metric, and several unverified modeling assumptions. The loss functions are standard, but the task-specific labels and quality metrics are internal to the paper and unreleased.

free parameters (4)

Loss weights lambda_cls, lambda_gen = not reported
Weights in Eq. 8 that balance classification and generation losses; chosen by hand and never disclosed, so the supervised objective is incompletely specified.
Decision threshold tau = not reported
Threshold in Section III-B1 converts the answerability score into a binary label; its value directly affects every reported accuracy number.
RLHF KL coefficient beta = not reported
Regularization coefficient in Eq. 12; chosen by hand and not reported, so the RLHF objective is incompletely specified.
Attention parameters W_a, v, b_a, W'_a, v', b'_a = learned, values not reported
Parameters of the hierarchical attention aggregators in Eqs. 2 and 5; the paper attributes its paragraph- and ranking-level gains to them but gives no values or initialization details.

assumptions (4)

domain assumption Answerability can be reliably annotated at sentence, paragraph, and ranking levels
All ECA labels depend on this; no annotation protocol, agreement score, or audit is provided. See Section IV-A1.
ad hoc to paper A [CLS]-style aggregate representation is available in the base LLM
Section III-B1 feeds h_CLS to the classification head, but the base model is unnamed and many decoder-only LLMs do not emit a [CLS] token.
domain assumption Human Likert ratings reliably measure helpfulness and trustworthiness
Section IV-D reports average Likert scores without annotator count, recruitment criteria, or inter-annotator agreement.
standard math Standard BCE and NLL losses train the model as intended
Eqs. 9 and 10 are standard objectives; the paper assumes their suitability without further justification.

invented entities (2)

Enhanced-CAsT-Answerability (ECA) dataset
purpose: Training and evaluation benchmark with hierarchical labels and ground-truth refusals
The dataset is the sole benchmark for all reported results, is not released, and has no external audit.
Appropriateness and Informativeness scores
purpose: Quantify refusal quality from human judgments
Described in prose in Section IV-A3 and Table VI with no rubric, no example responses, and no agreement statistics.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Contextual Candor: Enhancing LLM Trustworthiness Through Hierarchical Unanswerability Detection." pith.science (2026). https://pith.science/paper/RTKGXIQC

@misc{pith2026250601104,
  author       = {Pith},
  title        = {Pith review of: Contextual Candor: Enhancing LLM Trustworthiness Through Hierarchical Unanswerability Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RTKGXIQC}},
  note         = {Machine review of arXiv:2506.01104}
}

read the original abstract

The pervasive deployment of large language models (LLMs) in conversational AI systems has revolutionized information access, yet their propensity for generating factually unsupported or hallucinated responses remains a critical impediment to trustworthiness and widespread adoption. This paper introduces Reinforced Unanswerability Learning (RUL), a novel hybrid training paradigm designed to imbue LLMs with the intrinsic capability to accurately detect unanswerable questions and generate reliably appropriate responses. Unlike conventional approaches that rely on external classifiers or simple prompting, RUL integrates a discriminative unanswerability prediction head with the LLM's generative core, guided by a multi-stage learning strategy. This includes supervised fine-tuning on a novel, richly annotated dataset, Enhanced-CAsT-Answerability (ECA), which features hierarchical answerability labels and ground-truth refusal responses. Crucially, RUL incorporates a subsequent reinforcement learning with human feedback (RLHF) phase to refine the nuance, helpfulness, and informativeness of refusal responses. Extensive experiments demonstrate RUL's superior performance, achieving significantly higher accuracy in unanswerability detection across sentence, paragraph, and ranking levels, and substantially increasing the generation of appropriate refusals for unanswerable queries, alongside strong performance on answerable questions. Human evaluations further corroborate RUL's effectiveness, highlighting a marked improvement in perceived helpfulness and trustworthiness, ultimately paving the way for more reliable and user-centric conversational AI.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

28 extracted references · 6 canonical work pages

[6]

Attentional transfer is all you need: Technology-aware layout pattern generation,

X. Zhang, H. Yang, and E. F. Y . Young, “Attentional transfer is all you need: Technology-aware layout pattern generation,” in 58th ACM/IEEE Design Automation Conference, DAC 2021, San Francisco, CA, USA, December 5-9, 2021 . IEEE, 2021, pp. 169–174. [Online]. Available: https://doi.org/10.1109/DAC18074.2021.9586227

arXiv 2021
[8]

Enhancing low-cost video editing with lightweight adap- tors and temporal-aware inversion,

Y . He, S. Li, J. Wang, K. Li, X. Song, X. Yuan, K. Li, K. Lu, M. Huo, J. Chen et al., “Enhancing low-cost video editing with lightweight adap- tors and temporal-aware inversion,” arXiv preprint arXiv:2501.04606 , 2025

arXiv 2025
[11]

Language models with image descriptors are strong few- shot video-language learners,

Z. Wang, M. Li, R. Xu, L. Zhou, J. Lei, X. Lin, S. Wang, Z. Yang, C. Zhu, D. Hoiem, S. Chang, M. Bansal, and H. Ji, “Language models with image descriptors are strong few- shot video-language learners,” in Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, NeurIPS 2022, New Orleans, LA, U...

work page 2022
[28]

Instructpatentgpt: Training patent language models to follow instructions with human feedback,

J. Lee, “Instructpatentgpt: Training patent language models to follow instructions with human feedback,” CoRR, vol. abs/2406.16897, 2024. [Online]. Available: https://doi.org/10.48550/arXiv.2406.16897 10

arXiv doi:10.48550/arxiv.2406.16897 2024
[1]

Weak to strong generalization for large language models with multi-capabilities,

Y . Zhou, J. Shen, and Y . Cheng, “Weak to strong generalization for large language models with multi-capabilities,” in The Thirteenth International Conference on Learning Representations , 2025. [Online]. Available: https://openreview.net/forum?id=N1vYivuSKq

2025
[2]

Fine-grained distillation for long document retrieval,

Y . Zhou, T. Shen, X. Geng, C. Tao, J. Shen, G. Long, C. Xu, and D. Jiang, “Fine-grained distillation for long document retrieval,” in Proceedings of the AAAI Conference on Artificial Intelligence , vol. 38, no. 17, 2024, pp. 19 732–19 740

2024
[4]

Can users detect biases or factual errors in generated responses in conversational information-seeking?

W. Lajewska, K. Balog, D. Spina, and J. Trippas, “Can users detect biases or factual errors in generated responses in conversational information-seeking?” in Proceedings of the 2024 Annual International ACM SIGIR Conference on Research and Development in Information Retrieval in the Asia Pacific Region, SIGIR-AP 2024, Tokyo, Japan, December 9-12, 2024 , T...

arXiv 2024
[5]

Challenges in information-seeking QA: unanswerable questions and paragraph retrieval,

A. Asai and E. Choi, “Challenges in information-seeking QA: unanswerable questions and paragraph retrieval,” in Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, ACL/IJCNLP 2021, (Volume 1: Long Papers), Virtual Event, August 1-6, 2021 , C. Zo...

2021

Show all 28 references

[7]

Enhancing intent understanding for ambiguous prompts through human-machine co-adaptation,

Y . He, J. Wang, K. Li, Y . Wang, L. Sun, J. Yin, M. Zhang, and X. Wang, “Enhancing intent understanding for ambiguous prompts through human-machine co-adaptation,” arXiv preprint arXiv:2501.15167, 2025

2025
[9]

BERT: pre-training of deep bidirectional transformers for language understanding,

J. Devlin, M. Chang, K. Lee, and K. Toutanova, “BERT: pre-training of deep bidirectional transformers for language understanding,” in Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies,...

2019
[10]

Claret: Pre-training a correlation-aware context-to-event transformer for event-centric gener- ation and classification,

Y . Zhou, T. Shen, X. Geng, G. Long, and D. Jiang, “Claret: Pre-training a correlation-aware context-to-event transformer for event-centric gener- ation and classification,” in Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: L...

2022
[12]

Exploring the limits of transfer learning with a unified text-to-text transformer,

C. Raffel, N. Shazeer, A. Roberts, K. Lee, S. Narang, M. Matena, Y . Zhou, W. Li, and P. J. Liu, “Exploring the limits of transfer learning with a unified text-to-text transformer,” J. Mach. Learn. Res., vol. 21, pp. 140:1–140:67, 2020. [Online]. Available: https: //jmlr.org/p...

2020
[13]

Less is more: 9 Vision representation compression for efficient video generation with large language models,

Y . Zhou, J. Zhang, G. Chen, J. Shen, and Y . Cheng, “Less is more: 9 Vision representation compression for efficient video generation with large language models,” 2024

2024
[14]

Multimodal event transformer for image-guided story ending generation,

Y . Zhou and G. Long, “Multimodal event transformer for image-guided story ending generation,” in Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics , 2023, pp. 3434–3444

2023
[15]

Scaling laws for neural language models,

J. Kaplan, S. McCandlish, T. Henighan, T. B. Brown, B. Chess, R. Child, S. Gray, A. Radford, J. Wu, and D. Amodei, “Scaling laws for neural language models,” CoRR, vol. abs/2001.08361, 2020. [Online]. Available: https://arxiv.org/abs/2001.08361

2001 arXiv
[16]

Lamda: Language models for dialog applications,

R. Thoppilan, D. D. Freitas, J. Hall, N. Shazeer, A. Kulshreshtha, H. Cheng, A. Jin, T. Bos, L. Baker, Y . Du, Y . Li, H. Lee, H. S. Zheng, A. Ghafouri, M. Menegali, Y . Huang, M. Krikun, D. Lepikhin, J. Qin, D. Chen, Y . Xu, Z. Chen, A. Roberts, M. Bosma, Y . Zhou, C. Chang, ...

2022 arXiv
[17]

Training medical large vision- language models with abnormal-aware feedback,

Y . Zhou, L. Song, and J. Shen, “Training medical large vision- language models with abnormal-aware feedback,” arXiv preprint arXiv:2501.01377, 2025

2025 arXiv
[18]

Llama 2: Open foundation and fine-tuned chat models,

H. Touvron, L. Martin, K. Stone, P. Albert, A. Almahairi, Y . Babaei, N. Bashlykov, S. Batra, P. Bhargava, S. Bhosale, D. Bikel, L. Blecher, C. Canton-Ferrer, M. Chen, G. Cucurull, D. Esiobu, J. Fernandes, J. Fu, W. Fu, B. Fuller, C. Gao, V . Goswami, N. Goyal, A. Hartshorn, S...

2023 doi arXiv
[19]

Towards reliable and factual response generation: Detecting unanswerable questions in information-seeking conversations,

W. Lajewska and K. Balog, “Towards reliable and factual response generation: Detecting unanswerable questions in information-seeking conversations,” in Advances in Information Retrieval - 46th European Conference on Information Retrieval, ECIR 2024, Glasgow, UK, March 24-28, 2...

2024
[20]

Feqa: A question answering evaluation framework for faithfulness assessment in abstractive summarization,

E. Durmus, H. He, and M. Diab, “Feqa: A question answering evaluation framework for faithfulness assessment in abstractive summarization,” arXiv preprint arXiv:2005.03754 , 2020

2005 arXiv
[21]

Available: https://doi.org/10.1007/978-3-031-56063-7 25

[Online]. Available: https://doi.org/10.1007/978-3-031-56063-7 25

doi
[22]

True: Re-evaluating factual consistency evaluation,

O. Honovich, R. Aharoni, J. Herzig, H. Taitelbaum, D. Kukliansy, V . Cohen, T. Scialom, I. Szpektor, A. Hassidim, and Y . Matias, “True: Re-evaluating factual consistency evaluation,” arXiv preprint arXiv:2204.04991, 2022

2022 arXiv
[23]

Evaluating the factual consistency of abstractive text summarization,

W. Kry ´sci´nski, B. McCann, C. Xiong, and R. Socher, “Evaluating the factual consistency of abstractive text summarization,” arXiv preprint arXiv:1910.12840, 2019

1910 arXiv
[24]

Improving zero-shot cross-lingual transfer for multilingual question answering over knowledge graph,

Y . Zhou, X. Geng, T. Shen, W. Zhang, and D. Jiang, “Improving zero-shot cross-lingual transfer for multilingual question answering over knowledge graph,” in Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Huma...

2021
[25]

Improving factual consistency for knowledge-grounded dialogue systems via knowledge enhancement and alignment,

B. Xue, W. Wang, H. Wang, F. Mi, R. Wang, Y . Wang, L. Shang, X. Jiang, Q. Liu, and K.-F. Wong, “Improving factual consistency for knowledge-grounded dialogue systems via knowledge enhancement and alignment,” arXiv preprint arXiv:2310.08372 , 2023

arXiv
[26]

Truthfulqa: Measuring how models mimic human falsehoods,

S. Lin, J. Hilton, and O. Evans, “Truthfulqa: Measuring how models mimic human falsehoods,” arXiv preprint arXiv:2109.07958 , 2021

2021 arXiv
[27]

Siren’s song in the AI ocean: A survey on hallucination in large language models,

Y . Zhang, Y . Li, L. Cui, D. Cai, L. Liu, T. Fu, X. Huang, E. Zhao, Y . Zhang, Y . Chen, L. Wang, A. T. Luu, W. Bi, F. Shi, and S. Shi, “Siren’s song in the AI ocean: A survey on hallucination in large language models,” CoRR, vol. abs/2309.01219, 2023. [Online]. Available: ht...

2023 doi arXiv
[29]

A comprehensive method for model credibility measurement,

Y . Laili, L. Zhang, and G. Yang, “A comprehensive method for model credibility measurement,” in Model Engineering for Simulation . Else- vier, 2019, pp. 189–207

2019

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [6]

Attentional transfer is all you need: Technology-aware layout pattern generation,

X. Zhang, H. Yang, and E. F. Y . Young, “Attentional transfer is all you need: Technology-aware layout pattern generation,” in 58th ACM/IEEE Design Automation Conference, DAC 2021, San Francisco, CA, USA, December 5-9, 2021 . IEEE, 2021, pp. 169–174. [Online]. Available: https://doi.org/10.1109/DAC18074.2021.9586227

arXiv 2021

[2] [8]

Enhancing low-cost video editing with lightweight adap- tors and temporal-aware inversion,

Y . He, S. Li, J. Wang, K. Li, X. Song, X. Yuan, K. Li, K. Lu, M. Huo, J. Chen et al., “Enhancing low-cost video editing with lightweight adap- tors and temporal-aware inversion,” arXiv preprint arXiv:2501.04606 , 2025

arXiv 2025

[3] [11]

Language models with image descriptors are strong few- shot video-language learners,

Z. Wang, M. Li, R. Xu, L. Zhou, J. Lei, X. Lin, S. Wang, Z. Yang, C. Zhu, D. Hoiem, S. Chang, M. Bansal, and H. Ji, “Language models with image descriptors are strong few- shot video-language learners,” in Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, NeurIPS 2022, New Orleans, LA, U...

work page 2022

[4] [28]

Instructpatentgpt: Training patent language models to follow instructions with human feedback,

J. Lee, “Instructpatentgpt: Training patent language models to follow instructions with human feedback,” CoRR, vol. abs/2406.16897, 2024. [Online]. Available: https://doi.org/10.48550/arXiv.2406.16897 10

arXiv doi:10.48550/arxiv.2406.16897 2024

[5] [1]

Weak to strong generalization for large language models with multi-capabilities,

Y . Zhou, J. Shen, and Y . Cheng, “Weak to strong generalization for large language models with multi-capabilities,” in The Thirteenth International Conference on Learning Representations , 2025. [Online]. Available: https://openreview.net/forum?id=N1vYivuSKq

2025

[6] [2]

Fine-grained distillation for long document retrieval,

Y . Zhou, T. Shen, X. Geng, C. Tao, J. Shen, G. Long, C. Xu, and D. Jiang, “Fine-grained distillation for long document retrieval,” in Proceedings of the AAAI Conference on Artificial Intelligence , vol. 38, no. 17, 2024, pp. 19 732–19 740

2024

[7] [4]

Can users detect biases or factual errors in generated responses in conversational information-seeking?

W. Lajewska, K. Balog, D. Spina, and J. Trippas, “Can users detect biases or factual errors in generated responses in conversational information-seeking?” in Proceedings of the 2024 Annual International ACM SIGIR Conference on Research and Development in Information Retrieval in the Asia Pacific Region, SIGIR-AP 2024, Tokyo, Japan, December 9-12, 2024 , T...

arXiv 2024

[8] [5]

Challenges in information-seeking QA: unanswerable questions and paragraph retrieval,

A. Asai and E. Choi, “Challenges in information-seeking QA: unanswerable questions and paragraph retrieval,” in Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, ACL/IJCNLP 2021, (Volume 1: Long Papers), Virtual Event, August 1-6, 2021 , C. Zo...

2021