Pith. sign in

REVIEW 4 major objections 4 minor 29 references

Small Encoders Can Rival Large Decoders in Detecting Groundedness

T0 review · 4 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Fine-tuned encoder models can rival state-of-the-art LLMs at detecting whether a query is answerable from a given document, at a fraction of the compute.

desk verdict Useful efficiency finding on the QA datasets, but the IR columns silently shift the task from groundedness to relevance, so the broad claim overreaches. read the letter →

arxiv 2506.21288 v1 pith:N6ECJCOB submitted 2025-06-26 cs.CL cs.AIcs.IRcs.LG

Istabrak Abbes , Gabriele Prato , Quentin Fournier , Fernando Rodriguez , Alaa Boukhary , Adam Elwood , Sarath Chandar This is my paper
classification cs.CLcs.AIcs.IRcs.LG
keywords groundednessdetectionretrieval-augmentedgenerationencodermodelshallucinationRoBERTaNomicBERTzero-shotpromptinginferencecost
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 argues that the binary task of deciding whether a document contains enough information to answer a query does not need a large generative model. Fine-tuned encoder models such as RoBERTa and NomicBERT, which encode the query and context together and classify via the [CLS] token, reach accuracy within a few points of GPT-4o on this task while using orders of magnitude less compute. On SQuAD v2.0, RoBERTa-large reaches 90.2% versus GPT-4o's zero-shot 95.5%; on NewsQA, 88.5% versus 98.1%. The authors frame groundedness detection as a cheap pre-filter for retrieval-augmented generation: only pass context to a costly LLM when the encoder confirms the context is sufficient. They also document that zero-shot LLM accuracy is highly prompt-sensitive, especially for smaller models.

What carries the argument

The central object is the fine-tuned encoder as a groundedness classifier: the query and context are concatenated with a separator token and the [CLS] representation is trained to output a binary grounded/ungrounded label. This classification formulation converts groundedness detection into a semantic-matching problem, which plays to encoders' strength, and it is paired with a cost comparison in FLOPs that quantifies the inference advantage over decoder LLMs.

What would settle it

Re-annotate a sample of TREC-COVID and Touché query-document pairs with a direct groundedness label (does the document contain sufficient information to answer the query?) and compare those labels to the original relevance labels; if a substantial fraction of relevant documents cannot answer their queries, the encoder results on these datasets do not measure groundedness.

Watch

Extended reading notes

Core claim

The central discovery is that groundedness detection—a binary classification of whether a query is answerable from a supplied context—is a semantic-matching task where fine-tuned bidirectional encoders can be competitive with much larger decoder-based LLMs. The paper shows a fine-tuned RoBERTa-large achieves 90.2% on SQuAD v2.0 and 88.5% on NewsQA, within 5.3 and 9.6 points of zero-shot GPT-4o respectively, while its inference FLOPs are roughly three orders of magnitude lower than Llama-3.1-8B's. Across the four datasets, fine-tuned encoders consistently beat zero-shot open LLMs and often beat fine-tuned smaller LLMs. The authors conclude that for localized grounding in single-document scenarios, encoders are the better inductive fit and the better cost-performance trade-off.

Load-bearing premise

The load-bearing premise is that the relevance labels of the information-retrieval datasets (TREC-COVID and Touché) faithfully indicate whether the document actually contains enough information to answer the query, so that 'relevant' and 'grounded' mean the same thing; if a topically related document can exist without answering the query, the measured accuracy on those datasets is not groundedness accuracy as defined by the paper.

Editorial extensions

If this is right

  • A retrieval-augmented generation pipeline can insert a cheap encoder-based filter before LLM generation, avoiding expensive inference on ungrounded contexts.
  • Fine-tuned encoders can serve as open-source, low-latency alternatives to closed LLM APIs for groundedness detection in high-throughput production.
  • Zero-shot LLM groundedness detection is not a stable baseline; small prompt changes shift accuracy, so evaluations of such models must report prompt details.
  • Task-specific fine-tuning of encoders yields double-digit accuracy gains over zero-shot LLMs, suggesting pretraining alone is insufficient for this binary judgment task.

Reading between the lines

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

  • If the relevance-to-groundedness mapping used for the TREC-COVID and Touché datasets overestimates true groundedness, because a topically relevant document may still lack enough information to answer the query, then the paper's reported encoder accuracy on those datasets may not transfer to the paper's own definition of groundedness; testing on examples with separate relevance and answerability la
  • The approach is defined for single-document, single-query pairs; extending it to multi-hop or multi-document groundedness would likely require aggregation mechanisms beyond a binary [CLS] classifier, and the paper's own limitations acknowledge this.
  • The prompt sensitivity of small zero-shot decoders suggests that encoder-based classifiers could serve as stable, offline benchmarks for evaluating prompt quality in LLM-based detection pipelines.
  • A cheap groundedness filter could be applied before retrieval re-ranking or before answer generation, but the paper only studies the classification step, so end-to-end RAG accuracy and latency savings remain a testable extension.
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.

Referee Report

4 major / 4 minor

Summary. The paper proposes using small fine-tuned encoder models (BERT, RoBERTa, NomicBERT, ModernBERT, etc.) as a pre-filter to detect whether a query is grounded in a retrieved context before invoking a large language model for answer generation. The authors compare these encoders against zero-shot and fine-tuned decoder LLMs (Llama-3, GPT-4o, Claude 3.5) on four datasets: SQuAD v2.0, NewsQA, TREC-COVID, and Touché, reporting accuracy and FLOPs. The main empirical finding is that fine-tuned RoBERTa-large reaches 90.2% on SQuAD v2.0 and 88.5% on NewsQA, within a few points of zero-shot GPT-4o and fine-tuned Llama-3.1-8B, while using substantially fewer FLOPs. The paper also reports high prompt sensitivity for zero-shot LLMs and positions encoders as a computationally efficient alternative for groundedness detection.

Significance. If the result holds, the paper has clear practical value: a lightweight encoder could act as a cheap groundedness gate in RAG pipelines, reducing both hallucinations and compute. The manuscript includes multi-seed results, a hyperparameter grid, a prompt-sensitivity analysis, FLOPs estimates, and a code release, all of which are strengths. However, the central claim is broader than the evidence: two of the four benchmark columns (TREC-COVID and Touché) use BEIR retrieval-relevance labels as groundedness labels, and this mapping does not match the paper's own definition of groundedness. The empirically solid core is therefore limited to the two QA datasets, and the advertised cost/latency advantage is not consistently supported by the reported numbers. With appropriate relabeling or a narrowed claim, this could become a useful empirical contribution.

major comments (4)
  1. [Section 3.1, Table 1] The two information-retrieval datasets, TREC-COVID and Touché, are used as groundedness benchmarks by taking their BEIR relevance labels as binary groundedness labels, but the label-construction details are not given, and the original labels do not encode the Section 1 notion of groundedness (whether the context contains enough information to answer the query). TREC-COVID relevance indicates whether an article is useful for a COVID-19 research query, and Touché relevance indicates whether a document argues for or against a controversial claim; neither requires the document to support a direct answer, and Touché queries are often claims rather than answerable questions. The IR prompts in Appendix A.2 ask about 'relevant information', confirming that the task was operationalized as relevance detection rather than groundedness. Consequently, the TREC-COVID and Touché columns of Table 1 may not measure groundedness as defined, and the paper's central claim currently rests on the two QA datasets unless this label mapping is justified or replaced.
  2. [Section 3.2, Appendix A.2] The zero-shot LLM evaluations report the best of 20 prompts per dataset, but the manuscript does not state whether prompt selection was performed on a held-out validation set or on the same test set used for Table 1. If the best prompt is selected on the test set, the zero-shot LLM accuracies (e.g., 95.5% for GPT-4o on SQuAD v2.0) are optimistically biased, and the comparison with fine-tuned encoders is not a fair apples-to-apples comparison. Please specify the selection procedure or re-evaluate with a proper validation split.
  3. [Section 3.2, Table 1] The fine-tuning details are insufficient to assess the central claim. The manuscript says encoders and decoders were 'fine-tuned on curated datasets' but does not specify the training split for each dataset, how the binary labels were derived from SQuAD v2.0 and NewsQA, or the fine-tuning configuration for the open LLMs (e.g., LoRA vs. full fine-tuning, data mix, hyperparameters beyond the encoder grid). Without these details, it is not possible to rule out train/test overlap or to reproduce the fine-tuned encoder and LLM numbers that underpin the comparison.
  4. [Section 4, Table 1] The abstract and introduction claim that fine-tuned encoders reduce 'inference latency by orders of magnitude' and 'inference costs by up to 1,000x', but Table 1 reports FLOPs rather than wall-clock latency, and the inference-FLOPs ratio between ModernBERT-base (5.1e11) and Llama-3.1-8B (1.6e13) is about 30x, not 1,000x. The 1,000x figure appears to be based on fine-tuning FLOPs, not inference. Please correct the claim or provide actual latency/cost measurements.
minor comments (4)
  1. [Section 4] The text reports RoBERTa-large as achieving 88.8% on NewsQA, but Table 1 lists 88.5 ± 0.3; please reconcile these numbers.
  2. [Table 1] The table caption does not explain the two rightmost numeric columns labeled 'FLOPs FT Inference'; please clarify whether these are fine-tuning FLOPs and inference FLOPs, respectively.
  3. [Appendix A.2] The prompt lists would be easier to audit if the selected prompt per model and dataset were identified, and if the criterion for 'best' (e.g., mean accuracy across seeds) were stated.
  4. [Abstract] There is a formatting artifact in the abstract ('available at :') and the code URL is split across lines; please fix.

Circularity Check

1 steps flagged · score 1.0 of 10

No circular derivation: the central encoder claim is an empirical benchmark; the only mild issue is that zero-shot LLM numbers are selected as the best of 20 prompts on the evaluation data.

  1. fitted input called prediction [Section 3.2 (Decoders) and Section 4 (Interplay between Model Scale and Prompt), Table 1, Figures 2-3]
    "To comprehensively assess decoder model performance, we tested 20 carefully optimized prompts for QA datasets (SQuAD v2.0, NewsQA) and 20 tailored prompts for IR datasets (TREC-COVID, Touché) to enhance zero-shot accuracy and relevance evaluation (Appendix A.2)."

    The paper reports the 'highest zero-shot performance' across these 20 prompt templates (Section 4, Figures 2-3 plot per-template accuracy on the evaluation data), with no described held-out prompt-selection split. The zero-shot LLM rows in Table 1 are therefore the maximum over prompts evaluated on the same test data, i.e., a statistic fitted to the evaluation set rather than an independent prediction of a fixed prompt. This biases the LLM baseline upward and slightly weakens the claimed parity, but it does not infect the fine-tuned-encoder comparison, which is based on supervised training and held-out evaluation.

full rationale

The paper contains no formal derivation chain to be circular: it fine-tunes encoders and decoders on groundedness labels and reports held-out test accuracies. The central claim that small fine-tuned encoders rival large decoders is supported by supervised training and independent evaluation, so it does not reduce to its inputs. The only near-circular element is the decoder zero-shot protocol, where 20 prompts are tried on the evaluation data and the best result is reported; this is test-set prompt selection rather than a true zero-shot prediction. The use of BEIR relevance labels for TREC-COVID and Touché as groundedness labels is a construct-validity concern, not a circularity, because it is an assumption about label meaning rather than a derivation from the target quantity. No load-bearing self-citations or imported uniqueness arguments appear. Score 1 reflects the minor prompt-selection issue; no higher circularity score is warranted.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central claim is empirical. The main load-bearing assumptions are that BEIR relevance labels are a valid proxy for groundedness and that FLOPs represent inference cost; prompt selection on the test set introduces mild test-set fitting for the LLM comparison. No new entities are introduced.

free parameters (2)
  • Prompt template selection (best of 20) = Dataset-specific best prompt (chosen by test accuracy)
    The authors tested 20 prompts per dataset (Appendix A.2) and report the best-performing one for each zero-shot LLM, which effectively fits the prompt to the evaluation set and inflates reported zero-shot accuracy.
  • Fine-tuning hyperparameters = Chosen via grid search (e.g., learning rate 1.5e-5 to 3e-5, batch size 8 to 32)
    Standard hyperparameter selection for fine-tuned encoders and LLMs; not central to the comparison, but affects the reported means and standard deviations.
assumptions (3)
  • domain assumption BEIR relevance labels are a valid proxy for groundedness
    In Section 3.1, TREC-COVID and Touché are treated as groundedness benchmarks using retrieval relevance judgments, but a document can be topically relevant without containing sufficient information to answer the query.
  • domain assumption FLOPs are a valid proxy for inference latency and cost
    The abstract claims latency reduction by orders of magnitude, while Section 4 reports only FLOPs comparisons (Table 1), not wall-clock latency.
  • standard math Standard supervised learning assumptions for SQuAD v2.0 and NewsQA
    Fine-tuning and evaluation rely on the standard train/validation/test splits and the datasets' answerability labels being accurate.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Small Encoders Can Rival Large Decoders in Detecting Groundedness." pith.science (2026). https://pith.science/paper/N6ECJCOB

@misc{pith2026250621288,
  author       = {Pith},
  title        = {Pith review of: Small Encoders Can Rival Large Decoders in Detecting Groundedness},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/N6ECJCOB}},
  note         = {Machine review of arXiv:2506.21288}
}
read the original abstract

Augmenting large language models (LLMs) with external context significantly improves their performance in natural language processing (NLP) tasks. However, LLMs struggle to answer queries reliably when the provided context lacks information, often resorting to ungrounded speculation or internal knowledge. Groundedness - generating responses strictly supported by the context - is essential for ensuring factual consistency and trustworthiness. This study focuses on detecting whether a given query is grounded in a document provided in context before the costly answer generation by LLMs. Such a detection mechanism can significantly reduce both inference time and resource consumption. We show that lightweight, task specific encoder models such as RoBERTa and NomicBERT, fine-tuned on curated datasets, can achieve accuracy comparable to state-of-the-art LLMs, such as Llama3 8B and GPT4o, in groundedness detection while reducing inference latency by orders of magnitude. The code is available at : https://github.com/chandarlab/Hallucinate-less

Figures

Figures reproduced from arXiv: 2506.21288 by the authors.

Figure 1
Figure 1. Detecting groundedness before passing the [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Zero-shot groundedness performance of var [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Zero-shot groundedness performance of var [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

29 extracted references · 2 canonical work pages

  1. [1]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Akari Asai, Zeqiu Wu, Yizhong Wang, Avirup Sil, and Hannaneh Hajishirzi. 2023. https://arxiv.org/abs/2310.11511 Self-rag: Learning to retrieve, generate, and critique through self-reflection . Preprint, arXiv:2310.11511

    arXiv 2023

  4. [4]

    Akari Asai, Zexuan Zhong, Danqi Chen, Pang Wei Koh, Luke Zettlemoyer, Hannaneh Hajishirzi, and Wen-tau Yih. 2024. https://doi.org/10.48550/arXiv.2403.03187 Reliable, adaptable, and attributable language models with retrieval . Preprint, arxiv:2403.03187 [cs]

    arXiv doi:10.48550/arxiv.2403.03187 2024

  5. [5]

    Florin Cuconasu, Giovanni Trappolini, Federico Siciliano, Simone Filice, Cesare Campagnano, Yoelle Maarek, Nicola Tonellotto, and Fabrizio Silvestri. 2024. https://doi.org/10.1145/3626772.3657834 The power of noise: Redefining retrieval for rag systems . In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Informati...

    arXiv 2024

  6. [6]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. https://arxiv.org/abs/1810.04805 Bert: Pre-training of deep bidirectional transformers for language understanding . Preprint, arXiv:1810.04805

    arXiv 2019

  7. [7]

    Kristen Howell, Gwen Christian, Pavel Fomitchov, Gitit Kehat, Julianne Marzulla, Leanne Rolston, Jadin Tredup, Ilana Zimmerman, Ethan Selfridge, and Joseph Bradley. 2023. https://doi.org/10.18653/v1/2023.acl-industry.24 The economic trade-offs of large language models: A case study . In Proceedings of the 61st Annual Meeting of the Association for Computa...

    doi:10.18653/v1/2023.acl-industry.24 2023

  8. [8]

    Le, Abhishek Kumar, James Glass, Alexander Ratner, Chen-Yu Lee, Ranjay Krishna, and Tomas Pfister

    Cheng-Yu Hsieh, Yung-Sung Chuang, Chun-Liang Li, Zifeng Wang, Long T. Le, Abhishek Kumar, James Glass, Alexander Ratner, Chen-Yu Lee, Ranjay Krishna, and Tomas Pfister. 2024. https://arxiv.org/abs/2406.16008 Found in the middle: Calibrating positional attention bias improves long context utilization . Preprint, arXiv:2406.16008

    arXiv 2024

Show all 29 references
  1. [9]

    Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Ye Jin Bang, Andrea Madotto, and Pascale Fung. 2023. https://doi.org/10.1145/3571730 Survey of hallucination in natural language generation . ACM Computing Surveys, 55(12):1–38

    2023 doi

  2. [10]

    Zhouyu Jiang, Mengshu Sun, Lei Liang, and Zhiqiang Zhang. 2025. https://arxiv.org/abs/2407.13101 Retrieve, summarize, plan: Advancing multi-hop question answering with an iterative approach . Preprint, arXiv:2407.13101

    2025 arXiv

  3. [11]

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen tau Yih, Tim Rocktäschel, Sebastian Riedel, and Douwe Kiela. 2021. https://arxiv.org/abs/2005.11401 Retrieval-augmented generation for knowledge-int...

    2021 arXiv

  4. [12]

    Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang

    Nelson F. Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. 2023. https://arxiv.org/abs/2307.03172 Lost in the middle: How language models use long contexts . Preprint, arXiv:2307.03172

    2023 arXiv

  5. [13]

    Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. https://arxiv.org/abs/1907.11692 Roberta: A robustly optimized bert pretraining approach . Preprint, arXiv:1907.11692

    2019 arXiv

  6. [14]

    Morris, Brandon Duderstadt, and Andriy Mulyar

    Zach Nussbaum, John X. Morris, Brandon Duderstadt, and Andriy Mulyar. 2025. https://doi.org/10.48550/arXiv.2402.01613 Nomic embed: Training a reproducible long context text embedder . Preprint, arXiv:2402.01613 [cs]

    2025 doi arXiv

  7. [15]

    Pranav Rajpurkar, Robin Jia, and Percy Liang. 2018. https://arxiv.org/abs/1806.03822 Know what you don't know: Unanswerable questions for squad . Preprint, arXiv:1806.03822

    2018 arXiv

  8. [16]

    Abel Salinas and Fred Morstatter. 2024. https://doi.org/10.18653/v1/2024.findings-acl.275 The butterfly effect of altering prompts: How small changes and jailbreaks affect large language model performance . In Findings of the Association for Computational Linguistics: ACL 2024...

    2024 doi

  9. [17]

    Nandan Thakur, Nils Reimers, Andreas Rücklé, Abhishek Srivastava, and Iryna Gurevych. 2021. https://arxiv.org/abs/2104.08663 Beir: A heterogenous benchmark for zero-shot evaluation of information retrieval models . Preprint, arXiv:2104.08663

    2021 arXiv

  10. [18]

    Adam Trischler, Tong Wang, Xingdi Yuan, Justin Harris, Alessandro Sordoni, Philip Bachman, and Kaheer Suleman. 2017. https://arxiv.org/abs/1611.09830 Newsqa: A machine comprehension dataset . Preprint, arXiv:1611.09830

    2017 arXiv

  11. [19]

    Yuhao Wang, Ruiyang Ren, Junyi Li, Wayne Xin Zhao, Jing Liu, and Ji-Rong Wen. 2024 a . https://arxiv.org/abs/2402.17497 Rear: A relevance-aware retrieval-augmented framework for open-domain question answering . Preprint, arXiv:2402.17497

    2024 arXiv

  12. [20]

    Yuhao Wang, Ruiyang Ren, Junyi Li, Xin Zhao, Jing Liu, and Ji-Rong Wen. 2024 b . https://doi.org/10.18653/v1/2024.emnlp-main.321 REAR : A relevance-aware retrieval-augmented framework for open-domain question answering . In Proceedings of the 2024 Conference on Empirical Metho...

    2024 doi

  13. [21]

    Zhiruo Wang, Jun Araki, Zhengbao Jiang, Md Rizwan Parvez, and Graham Neubig. 2023. https://doi.org/10.48550/arXiv.2311.08377 Learning to filter context for retrieval-augmented generation . Preprint, arxiv:2311.08377 [cs]

    2023 doi arXiv

  14. [22]

    Siye Wu, Jian Xie, Jiangjie Chen, Tinghui Zhu, Kai Zhang, and Yanghua Xiao. 2024. https://arxiv.org/abs/2404.03302 How easily do irrelevant inputs skew the responses of large language models? Preprint, arXiv:2404.03302

    2024 arXiv

  15. [23]

    Jian Xie, Kai Zhang, Jiangjie Chen, Renze Lou, and Yu Su. 2024. https://arxiv.org/abs/2305.13300 Adaptive chameleon or stubborn sloth: Revealing the behavior of large language models in knowledge conflicts . Preprint, arXiv:2305.13300

    2024 arXiv

  16. [24]

    Ori Yoran, Tomer Wolfson, Ori Ram, and Jonathan Berant. 2024. https://arxiv.org/abs/2310.01558 Making retrieval-augmented language models robust to irrelevant context . Preprint, arXiv:2310.01558

    2024 arXiv

  17. [25]

    Jiahao Zhang, Haiyang Zhang, Dongmei Zhang, Yong Liu, and Shen Huang. 2024. https://arxiv.org/abs/2308.08973 End-to-end beam retrieval for multi-hop question answering . Preprint, arXiv:2308.08973

    2024 arXiv

  18. [26]

    Wenxuan Zhang, Yue Deng, Bing Liu, Sinno Jialin Pan, and Lidong Bing. 2023. https://arxiv.org/abs/2305.15005 Sentiment analysis in the era of large language models: A reality check . Preprint, arXiv:2305.15005

    2023 arXiv

  19. [27]

    Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, Yifan Du, Chen Yang, Yushuo Chen, Zhipeng Chen, Jinhao Jiang, Ruiyang Ren, Yifan Li, Xinyu Tang, Zikang Liu, Peiyu Liu, Jian-Yun Nie, and Ji-Rong W...

    2024 doi arXiv

  20. [28]

    Yujia Zhou, Zheng Liu, Jiajie Jin, Jian-Yun Nie, and Zhicheng Dou. 2024. https://arxiv.org/abs/2402.11626 Metacognitive retrieval-augmented large language models . Preprint, arXiv:2402.11626

    2024 arXiv

  21. [29]

    Ilana Zimmerman, Jadin Tredup, Ethan Selfridge, and Joseph Bradley. 2024. https://doi.org/10.18653/v1/2024.emnlp-industry.2 Two-tiered encoder-based hallucination detection for retrieval-augmented generation in the wild . In Proceedings of the 2024 Conference on Empirical Meth...

    2024 doi

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.