Pith. sign in

REVIEW 4 major objections 6 minor 33 references

DeRAGEC: Denoising Named Entity Candidates with Synthetic Rationale for ASR Error Correction

T0 review · 4 major / 6 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 Explicitly filtering retrieved named entities before ASR correction cuts word error rate by 28 percent.

desk verdict A solid, incremental ASR post-processing paper where the headline gain hinges on an unverified transfer assumption; worth a careful referee, not a desk reject. read the letter →

arxiv 2506.07510 v1 pith:H4AQS5JE submitted 2025-06-09 cs.CL

Solee Im , Wonjun Lee , Jinmyeong An , Yunsu Kim , Jungseul Ok , Gary Geunbae Lee This is my paper
classification cs.CL
keywords ASRerrorcorrectionnamedentityretrieval-augmentedgenerationrationale-basedfilteringin-contextlearningphoneticsimilaritywordratesyntheticrationale
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

DeRAGEC claims that the noise in retrieved named-entity candidates, not the retrieval itself, is what limits retrieval-augmented ASR error correction. It adds a training-free denoising gate that scores each candidate by phonetic similarity, attaches a one-line definition, and uses a large language model to write a short rationale explaining which candidate fits the transcript before the final correction step. On CommonVoice and STOP, the full method reaches word error rates of 6.0 and 5.9 with Llama-3.1, compared with 6.5 for plain retrieval-augmented correction, a 28% relative reduction over the ASR-only baseline. The paper's own oracle experiment, which feeds the ground-truth named entity into the corrector, lands at 5.8 and 5.7, so DeRAGEC closes most of the gap without seeing the answer.

What carries the argument

The load-bearing mechanism is the denoising gate in the MCQ format: each retrieved NE candidate is enriched with a phonetic similarity score (articulatory-feature distance to the query) and a one-sentence Wikipedia definition, and the LLM is prompted to answer a cloze-style multiple-choice question selecting the single best NE, with a chain-of-thought rationale between <think> and <answer> tags. The rationale is synthesized offline by a stronger model using ground-truth transcriptions, then reused as few-shot demonstrations for the same model at inference. This separates NE selection from transcription correction into two modular prompts, which the appendix shows beats a single combined inference step; the selected NE and its rationale are then fed to the GEC prompt along with the 5-best ASR hypotheses.

What would settle it

Run the same pipeline but replace the rationale with a random or deliberately wrong choice in the MCQ step, while keeping the multiple-choice format and few-shot examples otherwise identical; if WER stays near 6.0 instead of degrading toward the RAGEC baseline of 6.5, then the specific rationale content is not carrying the effect. A weaker test: on the CommonVoice split, corrupt the phonetic scores or definitions and measure whether the recall of the correct NE (currently 0.839 versus 0.841 ceiling) drops.

Watch

Extended reading notes

Core claim

The central claim is that an explicit, rationale-based filtering step, inserted between phonetic retrieval and generative correction, retrieves and keeps the correct named entity while discarding phonetically similar distractors, and that this filtering can be learned from a few demonstrations instead of training. DeRAGEC serializes each candidate as <entity | phonetic-score | definition>, produces a synthetic rationale offline for each training triplet, and at inference asks the same model to first answer a multiple-choice 'which entity fills the blank' question and then perform error correction conditioned on the chosen entity and rationale. The paper reports WER 6.0/5.9 and NE hit ratio 0.831/0.838 on CommonVoice/STOP with Llama-3.1, versus 6.5/6.5 and 0.804/0.807 for RAGEC with no denoising; the oracle, where the correct NE is provided, gives 5.8/5.7 and 0.837/0.857. The measured precision of the filtered candidate set rises from 0.056 to 0.139 while recall stays at 0.839, against a 0.841 ceiling, which the authors take as evidence that the gain comes from removing noise, not from better retrieval.

Load-bearing premise

The method assumes that rationales written offline by a stronger model, using the ground-truth transcript, still guide a smaller model to pick the right named entity when it only has the noisy ASR hypothesis at test time; if the test-time rationale is wrong, the filter may delete the correct entity or keep a distractor, wiping out the gains.

Editorial extensions

If this is right

  • On both benchmarks the full DeRAGEC system beats plain RAGEC by 0.5-0.6 WER with Llama-3.1 and 0.8-0.9 WER with GPT-4o-mini, and most of the oracle gap closes.
  • NE hit ratio rises to 0.831/0.838, within 0.006-0.019 of the oracle, meaning the corrected transcript contains the right entity almost as often as if the answer were given.
  • Filtering raises candidate-set precision from 0.056 to 0.139 while recall drops only 0.002, so the same top-10 retrieval becomes usable without re-retrieval.
  • Because the method is training-free, it can be dropped into an existing RAGEC pipeline by changing only the prompts and adding offline rationale generation for few-shot examples.
  • Separating NE selection (MCQ) from correction outperforms doing both in one call, so the modular two-step design is itself part of the gain.

Reading between the lines

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

  • The paper does not isolate whether the rationale is load-bearing or merely a side effect of the MCQ format; a further ablation feeding the correct NE with a wrong rationale would tell whether the model trusts the chosen entity or re-derives it.
  • If rationales are the active ingredient, a cheaper deployment path is to distill them into the inference model during training; the authors flag this as future work, and the current pipeline depends on an offline stronger model for every few-shot example.
  • The 0.841 recall@10 ceiling means DeRAGEC's gains are capped by retrieval quality; combining the gate with a higher-recall retriever (e.g., top-15 or learned phonetic embeddings) is a natural extension that could push WER below the oracle numbers in Table 4.
  • The evaluation covers only two English speech domains with one ASR engine; a direct stress test on out-of-domain names (e.g., rare personal names in noisy audio) would show whether the definition-based reasoning generalizes.
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 / 6 minor

Summary. The paper proposes DeRAGEC, a training-free extension of RAGEC for ASR named-entity error correction. DeRAGEC augments each retrieved NE candidate with a phonetic similarity score and a one-line definition, and uses synthetic rationales (generated offline by o1 with ground-truth transcriptions) as few-shot demonstrations for an MCQ-style NE filtering step, followed by a GEC step. Experiments on CommonVoice (2,000 samples) and STOP (5,000 samples) report WER reductions from 7.7/8.9 (ASR-only) to 6.0/5.9 (DeRAGEC with Llama-3.1), with an ablation in Table 1 and an oracle comparison in Table 4.

Significance. If the reported gains are robust, DeRAGEC offers a practical, training-free improvement for NE-aware ASR post-processing, with strengths including public code, a clear modular ablation, and an explicit oracle upper bound. However, the central denoising mechanism is not convincingly validated: the selection precision in Figure 2 is at chance, and the main WER improvements lack statistical significance; the contribution is therefore promising but not yet established.

major comments (4)
  1. [Section 4.3, Figure 2] The claim that the denoising step successfully selects the correct NE is not supported by the reported precision. With a precision upper bound of 0.166 (1/6) and recall of 0.839, a random selection among the six candidate slots would achieve precision approximately equal to recall divided by 6, about 0.140, and the reported precision of 0.139 is indistinguishable from that chance level. The paper should report a random-selection baseline and the selection accuracy conditional on whether the correct NE is present in the retrieved list or only in the hypothesis NEs, and it should show that the filtering is above chance. Without such evidence, the WER improvement cannot be attributed to explicit denoising rather than to the enriched prompt features or the rationale-conditioned GEC step.
  2. [Section 4.1, Table 1] All comparisons lack error bars or significance tests, and the test sets are small (2,000 and 5,000 samples). The key deltas, for example CV Llama-3.1 WER from 6.5 to 6.0 and STOP from 6.2 to 5.9, are small absolute changes, so the abstract's claim of significant improvements is not established. The paper should report bootstrap confidence intervals for WER or repeat the experiments with multiple random few-shot seeds and report the variance across seeds.
  3. [Algorithm 1, Section 3.2, Limitation] The synthetic rationales are generated by o1 using the ground-truth transcription a, but the test-time models (Llama-3.1 and GPT-4o-mini) receive only the few-shot demonstrations. The paper provides no direct evidence that these rationales transfer, such as a control condition with generic step-by-step prompting or an analysis of test-time rationale quality per model. The limitation section explicitly acknowledges that internalization and generalization remain open questions; because the Rat feature is the main addition over RAGEC, this missing validation is load-bearing for the central claim.
  4. [Table 1 vs. Table 4] The ORACLE settings are inconsistent. Table 1 reports a CV ORACLE WER of 5.8, while Table 4 reports ORACLE (only the correct NE provided) as 4.1 and ORACLE-retr (retrieved NE with noise removed) as 5.8. The text in Section 4.2 says DeRAGEC shows a small gap to ORACLE, which is misleading if ORACLE is defined as in Table 4. The paper should define ORACLE consistently across tables and clarify which setting is being compared in the main text.
minor comments (6)
  1. [Equation (2)] There is a mismatched parenthesis: Mθ(a|H,E_gec)) should be Mθ(a|H,E_gec).
  2. [Section 3.1] N_hyp is described as {n_hyp_1, ..., n_hyp_5}, but H contains five hypotheses and each hypothesis can contain multiple NEs; please clarify how the five hypothesis NEs are extracted and how the total candidate count is determined.
  3. [Section 4.3] The phrase 'which dose not applied' should be 'which does not apply'.
  4. [Abstract and throughout] The word 'CommonV oice' contains a line-break artifact; use 'CommonVoice' consistently.
  5. [Figure 3 caption] The caption says 'REGEC' when it should say 'RAGEC'.
  6. [Table 4 caption] The terms ORACLE-retr and ORACLE are used in the appendix without being defined in the main text; please define them where the oracle comparison is first discussed.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular reasoning found; DeRAGEC's synthetic rationales are constructed from training labels as standard few-shot demonstrations, and evaluation is against held-out test sets and external baselines.

full rationale

The paper's derivation chain is empirical and self-contained. Algorithm 1 synthesizes rationales using the ground-truth transcription a (r_syn ← M_r(h1, a, N, PS, Def)), but only for training-set triplets; at inference the target a is neither supplied to the NE-filtering model nor to the GEC stage (Eq. 4 conditions only on H, N, PS, Def and sampled demonstrations E_deragec). This is supervised construction of ICL demonstrations, not leakage of test labels or a fitted-input-called-prediction move. No parameters are trained; phonetic scores come from Panphon/Epitran, definitions from Wikipedia, and candidates from a fixed NE database. The claimed 28% relative WER reduction is measured against external baselines (ASR, GEC, RAGEC) and an ORACLE upper bound, so the result does not reduce to its inputs by construction. The paper contains no load-bearing self-citations: none of the cited prior works share the present authors, and no uniqueness theorem or prior result is invoked to force the method. The only substantive caveat — whether o1-generated rationales transfer to smaller test-time models — is explicitly flagged in the Limitation section as an open generalization question; that is an empirical risk, not circularity. Overall, the central claim has independent content and is falsifiable against held-out CommonVoice and STOP data.

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

The method introduces no new entities or physical quantities. Its claims rest on standard NLP components (pre-trained models, retrieval, phonetic tools) and on the unstated assumption that these components interact as expected. The only hand-chosen quantities are the retrieval count and few-shot count.

free parameters (2)
  • Retrieval count (top-k) = 10
    The number of phonetically similar NEs retrieved is set to 10. Appendix A shows WER varies with k (6.5 to 7.2 on CV), so this choice affects results and is not derived.
  • Few-shot count (T_fs) = 5
    Number of in-context examples sampled for ICL; Figure 5 shows performance changes with this count, so it is a hand-chosen hyperparameter.
assumptions (3)
  • domain assumption Pre-trained LLMs (Llama-3.1, GPT-4o-mini, o1) can reliably perform named-entity selection and rationale generation from cloze-style prompts.
    The method's success depends on the LLM's ability to reason about phonetic similarity and definitions; the paper provides qualitative case studies but no systematic analysis of failure modes.
  • domain assumption The phonetic similarity computed via Epitran/Panphon and the one-line Wikipedia definitions are sufficient discriminative signals for choosing the correct NE.
    These features are the only inputs to the denoising gate; Appendix Tables 2/3 show static thresholds based on phonetic similarity alone reduce recall, indicating the signal is imperfect.
  • ad hoc to paper Synthetic rationales generated with ground-truth answers transfer to test-time inference with a different model.
    The few-shot rationales are produced by o1 with access to the correct answer, but test-time rationales are produced by the deployed model without it. The limitation section explicitly flags the need to study internalization and generalization.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DeRAGEC: Denoising Named Entity Candidates with Synthetic Rationale for ASR Error Correction." pith.science (2026). https://pith.science/paper/H4AQS5JE

@misc{pith2026250607510,
  author       = {Pith},
  title        = {Pith review of: DeRAGEC: Denoising Named Entity Candidates with Synthetic Rationale for ASR Error Correction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/H4AQS5JE}},
  note         = {Machine review of arXiv:2506.07510}
}
read the original abstract

We present DeRAGEC, a method for improving Named Entity (NE) correction in Automatic Speech Recognition (ASR) systems. By extending the Retrieval-Augmented Generative Error Correction (RAGEC) framework, DeRAGEC employs synthetic denoising rationales to filter out noisy NE candidates before correction. By leveraging phonetic similarity and augmented definitions, it refines noisy retrieved NEs using in-context learning, requiring no additional training. Experimental results on CommonVoice and STOP datasets show significant improvements in Word Error Rate (WER) and NE hit ratio, outperforming baseline ASR and RAGEC methods. Specifically, we achieved a 28% relative reduction in WER compared to ASR without postprocessing. Our source code is publicly available at: https://github.com/solee0022/deragec

Figures

Figures reproduced from arXiv: 2506.07510 by the authors.

Figure 1
Figure 1. Comparison of RAGEC and DeRAGEC in handling retrieved NE candidates. For clarity, input features such as P S and Def are omitted to highlight the key differences between the two methods. as a phonetic query qp for retrieving the top-k pho￾netically similar NE candidates N = {n1, . . . , nk} from an external NE database. In the ASR baseline, the top hypothesis is di￾rectly used as the transcription aˆ = h1 (Eq.1). In… view at source ↗
Figure 2
Figure 2. Recall and Precision of NE hit ratio of candi [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Case study between REGEC and DeRAGEC in CommonVoice dataset. The process of denoising irrelevant [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Case study between REGEC and DeRAGEC in STOP dataset. The process of selecting the correct NE is [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Effect of number of few-shot examples in [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]
Figure 6
Figure 6. Figure 6: NER performance (F1 score) on different WER. 11 [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

33 extracted references · 11 canonical work pages

  1. [1]

    Rosana Ardila, Megan Branson, Kelly Davis, Michael Henretty, Michael Kohler, Josh Meyer, Reuben Morais, Lindsay Saunders, Francis M Tyers, and Gregor Weber. 2019. Common voice: A massively-multilingual speech corpus. arXiv preprint arXiv:1912.06670

    arXiv 2019

  2. [2]

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33:1877--1901

    2020

  3. [3]

    Chen Chen, Yuchen Hu, Chao-Han Huck Yang, Sabato Marco Siniscalchi, Pin-Yu Chen, and Eng-Siong Chng. 2023. Hyporadise: An open baseline for generative speech recognition with large language models. Advances in Neural Information Processing Systems, 36:31665--31688

    work page 2023

  4. [4]

    Florin Cuconasu, Giovanni Trappolini, Federico Siciliano, Simone Filice, Cesare Campagnano, Yoelle Maarek, Nicola Tonellotto, and Fabrizio Silvestri. 2024. The power of noise: Redefining retrieval for rag systems. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 719--729

    2024

  5. [5]

    Isabel O Gallegos, Ryan A Rossi, Joe Barrow, Md Mehrab Tanjim, Sungchul Kim, Franck Dernoncourt, Tong Yu, Ruiyi Zhang, and Nesreen K Ahmed. 2024. Bias and fairness in large language models: A survey. Computational Linguistics, pages 1--79

    2024

  6. [6]

    Sreyan Ghosh, Mohammad Sadegh Rasooli, Michael Levit, Peidong Wang, Jian Xue, Dinesh Manocha, and Jinyu Li. 2024. Failing forward: Improving generative error correction for asr with synthetic data and retrieval augmentation. arXiv preprint arXiv:2410.13198

    arXiv 2024

  7. [7]

    Xun Gong, Anqi Lv, Zhiming Wang, and Yanmin Qian. 2024. Contextual biasing speech recognition in speech-enhanced large language model. Proc. Interspeech. ISCA, pages 257--261

    work page 2024

  8. [8]

    Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat, and Mingwei Chang. 2020. Retrieval augmented language model pre-training. In International conference on machine learning, pages 3929--3938. PMLR

    2020

Show all 33 references
  1. [9]

    Yuchen Hu, Chen Chen, Chengwei Qin, Qiushi Zhu, Eng Siong Chng, and Ruizhe Li. 2024 a . Listen again and choose the right answer: A new paradigm for automatic speech recognition with large language models. arXiv preprint arXiv:2405.10025

    2024 arXiv

  2. [10]

    Yuchen Hu, CHEN CHEN, Chao-Han Huck Yang, Ruizhe Li, Chao Zhang, Pin-Yu Chen, and EngSiong Chng. 2024 b . Large language models are efficient learners of noise-robust speech recognition. In The Twelfth International Conference on Learning Representations

    2024

  3. [11]

    Gautier Izacard, Patrick Lewis, Maria Lomeli, Lucas Hosseini, Fabio Petroni, Timo Schick, Jane Dwivedi-Yu, Armand Joulin, Sebastian Riedel, and Edouard Grave. 2023. Atlas: Few-shot learning with retrieval augmented language models. Journal of Machine Learning Research, 24(251):1--43

    2023

  4. [12]

    Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, et al. 2024. Openai o1 system card. arXiv preprint arXiv:2412.16720

    2024 arXiv

  5. [13]

    Zhihong Lei, Xingyu Na, Mingbin Xu, Ernest Pusateri, Christophe Van Gysel, Yuanyuan Zhang, Shiyi Han, and Zhen Huang. 2024. Contextualization of asr with llm using phonetic retrieval-based augmentation. arXiv preprint arXiv:2409.15353

    2024 arXiv

  6. [14]

    u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt \

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich K \"u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt \"a schel, et al. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in Neural Information Proc...

    2020

  7. [15]

    Daliang Li, Ankit Singh Rawat, Manzil Zaheer, Xin Wang, Michal Lukasik, Andreas Veit, Felix Yu, and Sanjiv Kumar. 2022. Large language models with controllable working memory. arXiv preprint arXiv:2211.05110

    2022 arXiv

  8. [16]

    Sheng Li, Chen Chen, Chin Yuen Kwok, Chenhui Chu, Eng Siong Chng, and Hisashi Kawai. 2024. Investigating asr error correction with large language model and multilingual 1-best hypotheses. In Proc. Interspeech, pages 1315--1319

    2024

  9. [17]

    Rao Ma, Mengjie Qian, Potsawee Manakul, Mark Gales, and Kate Knill. 2023. Can generative large language models perform asr error correction? arXiv preprint arXiv:2307.04172

    2023 arXiv

  10. [18]

    MetaAI. 2024. https://arxiv.org/abs/2407.21783 The llama 3 herd of models . arXiv preprint arXiv:2407.21783

    2024 arXiv

  11. [19]

    Mortensen, Siddharth Dalmia, and Patrick Littell

    David R. Mortensen, Siddharth Dalmia, and Patrick Littell. 2018. https://aclanthology.org/L18-1429/ E pitran: Precision G 2 P for many languages . In Proceedings of the Eleventh International Conference on Language Resources and Evaluation ( LREC 2018) , Miyazaki, Japan. Europ...

    2018

  12. [20]

    Mortensen, Patrick Littell, Akash Bharadwaj, Kartik Goyal, Chris Dyer, and Lori Levin

    David R. Mortensen, Patrick Littell, Akash Bharadwaj, Kartik Goyal, Chris Dyer, and Lori Levin. 2016. https://aclanthology.org/C16-1328/ P an P hon: A resource for mapping IPA segments to articulatory feature vectors . In Proceedings of COLING 2016, the 26th International Conf...

    2016

  13. [21]

    OpenAI. 2023. https://arxiv.org/abs/2303.08774 Gpt-4 technical report . arXiv preprint arXiv:2303.08774

    2023 arXiv

  14. [22]

    Ernest Pusateri, Anmol Walia, Anirudh Kashi, Bortik Bandyopadhyay, Nadia Hyder, Sayantan Mahinder, Raviteja Anantha, Daben Liu, and Sashank Gondala. 2024. Retrieval augmented correction of named entity speech recognition errors. arXiv preprint arXiv:2409.06062

    2024 arXiv

  15. [23]

    Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. 2022. https://doi.org/10.48550/ARXIV.2212.04356 Robust speech recognition via large-scale weak supervision . arXiv preprint

    2022 doi arXiv

  16. [24]

    Srijith Radhakrishnan, Chao-Han Yang, Sumeer Khan, Rohit Kumar, Narsis Kiani, David Gomez-Cabrero, and Jesper Tegn \'e r. 2023. Whispering llama: A cross-modal generative error correction framework for speech recognition. In Proceedings of the 2023 Conference on Empirical Meth...

    2023

  17. [25]

    Paden Tomasello, Akshat Shrivastava, Daniel Lazar, Po-Chun Hsu, Duc Le, Adithya Sagar, Ali Elkahky, Jade Copet, Wei-Ning Hsu, Yossi Adi, et al. 2023. Stop: A dataset for spoken task oriented semantic parsing. In 2022 IEEE Spoken Language Technology Workshop (SLT), pages 991--998. IEEE

    2023

  18. [26]

    Christophe Van Gysel, Mirko Hannemann, Ernest Pusateri, Youssef Oualil, and Ilya Oparin. 2022. Space-efficient representation of entity-centric query language models. In Proc. Interspeech 2022, pages 679--683

    2022

  19. [27]

    Yi-Cheng Wang, Hsin-Wei Wang, Bi-Cheng Yan, Chi-Han Lin, and Berlin Chen. 2024. Dancer: Entity description augmented named entity corrector for automatic speech recognition. arXiv preprint arXiv:2403.17645

    2024 arXiv

  20. [28]

    Zhepei Wei, Wei-Lin Chen, and Yu Meng. 2024. Instructrag: Instructing retrieval-augmented generation with explicit denoising. arXiv preprint arXiv:2406.13629

    2024 arXiv

  21. [29]

    Kevin Wu, Eric Wu, and James Zou. 2024. How faithful are rag models? quantifying the tug-of-war between rag and llms' internal prior. arXiv e-prints, pages arXiv--2404

    2024

  22. [30]

    Chao-Han Huck Yang, Yile Gu, Yi-Chieh Liu, Shalini Ghosh, Ivan Bulyko, and Andreas Stolcke. 2023. Generative speech recognition error correction with large language models and task-activating prompting. In 2023 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU...

    2023

  23. [31]

    Ori Yoran, Tomer Wolfson, Ori Ram, and Jonathan Berant. 2023. Making retrieval-augmented language models robust to irrelevant context. arXiv preprint arXiv:2310.01558

    2023 arXiv

  24. [32]

    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...

  25. [33]

    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 gl...

Pith tools

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