Pith. sign in

REVIEW 3 major objections 6 minor 33 references

Factors affecting the in-context learning abilities of LLMs for dialogue state tracking

T0 review · 3 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 For in-context dialogue state tracking, retrieving demonstrations from user utterances alone outperforms retrieving from full user-agent turns, improving both slot precision and recall across open LLMs.

desk verdict A careful, narrow ICL-for-DST study whose headline user-only-retrieval result holds for slot extraction with oracle domains but is not demonstrated for end-to-end DST. read the letter →

arxiv 2506.08753 v1 pith:5ADEN3O7 submitted 2025-06-10 cs.CL cs.AI

Pradyoth Hegde , Santosh Kesiraju , Jan Švec , Šimon Sedláček , Bolaji Yusuf , Oldřich Plchot , Deepak K T , Jan Černocký This is my paper · ORCID
classification cs.CLcs.AI
keywords in-contextlearningdialoguestatetrackingdemonstrationretrievalsentenceembeddingsMultiWOZ2.4promptengineeringconstraineddecodingslotprecisionandrecall
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 which design choices actually move the needle when large language models perform dialogue state tracking through in-context learning. The headline result is that the retriever should embed only the user's utterances: with LaBSE and 10 demonstrations, user-only embeddings give 69.9 slot precision and 80.6 recall, versus 67.8 and 78.3 when user and agent turns are both embedded. The same study finds that a general-purpose multilingual embedding model performs as well as a dialogue-specialised one, that speaker tags have a small but consistent effect, and that most of the benefit comes from the first few demonstrations. If these results hold, in-context dialogue state tracking becomes cheaper and simpler, because it needs no task-specific retriever training and can rely on a user-focused embedding signal.

What carries the argument

The carrying mechanism is a $k$-nearest-neighbour demonstration retriever feeding a modular prompt template under constrained decoding. The retriever pre-computes sentence embeddings of training turns with either LaBSE or Dialog2Flow, then selects the $K$ closest turns to the test user turn by cosine similarity; the prompt arranges those demonstrations, nearest one last, as blocks of conversation history, a JSON domain list, and JSON slot key-value pairs, followed by the test turn. Constrained decoding fills each slot value given its schema-fixed domain and slot key. The paper's factor sweep changes one component at a time in this pipeline—embedding model, embedded text type, speaker tags, demonstration count, history type, and decoding strategy—so that slot precision and recall can be attributed to specific choices.

What would settle it

Run the identical system with the domain predicted by the LLM rather than supplied by the dataset, and score both domain and slot correctness jointly: if user-only retrieval no longer beats user-agent retrieval, the headline result is an artifact of the oracle-domain evaluation. A second check is to test user-only versus user-agent retrieval on a different multi-domain dialogue dataset at 1, 3, and 10 demonstrations to see whether the ordering persists.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central claim is that in-context learning for dialogue state tracking is best driven by a retrieval signal drawn from the user side of the conversation plus a modular prompt that supplies history, domains, and slots separately. Embedding only user turns for nearest-neighbour demonstration retrieval consistently outperforms embedding user-agent turns, and the gap grows with more demonstrations (69.9 vs 67.8 precision and 80.6 vs 78.3 recall at 10 demonstrations with LaBSE). The paper also claims that the dialogue-specialised Dialog2Flow retriever does not beat the general-purpose LaBSE retriever once user-only embeddings are used, that omitting speaker tags slightly raises precision but lowers recall, and that decoding slot values from given keys is far more reliable than generating whole key-value pairs. All of these results are measured on MultiWOZ 2.4, with the ground-truth domain supplied to the scorer.

Load-bearing premise

The load-bearing premise is that the ground-truth dialogue domain is already known when slot predictions are scored, so the reported numbers measure only slot filling and may not represent an end-to-end dialogue state tracker that must also identify the domain.

Editorial extensions

If this is right

  • A user-only retriever can be treated as a safe default for in-context dialogue state tracking, roughly halving the amount of text that must be embedded for retrieval.
  • Teams can skip training or fine-tuning a dialogue-specific embedding model, since a general multilingual embedding encoder matches it on this task.
  • Three demonstrations capture most of the available benefit, so short prompts suffice and token budgets stay small.
  • Prompt structure is a first-order factor: speaker tags and whether agent turns appear in the history change the precision-recall balance measurably.
  • Retrieval signal and prompt context can be decoupled, since the best configuration retrieves on user turns while still showing the full user-agent history in the prompt.

Reading between the lines

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

  • Beyond the paper, if this user-only advantage survives full dialogue state tracking with predicted domains, systems could be built on lightweight user-turn encoders alone, which would cut deployment cost in low-resource settings.
  • Beyond the paper, the slot-key relevance and coverage analysis suggests a train-free extension: rerank retrieved demonstrations by predicted slot-key overlap with the test turn, rather than by raw cosine similarity.
  • Beyond the paper, the finding that a multilingual embedding model matches a dialogue-specific one hints that user-only retrieval could transfer demonstrations across languages, a testable zero-shot cross-lingual DST setting.
  • Beyond the paper, because the evaluation assumes the ground-truth domain, an end-to-end system that also predicts the domain could find a different optimal configuration; the reported factor rankings should be re-checked under a joint domain-and-slot metric.
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

3 major / 6 minor

Summary. This manuscript studies in-context learning for dialogue state tracking on MultiWOZ2.4. The proposed system retrieves K nearest dialogue turns using sentence embeddings (LaBSE or Dialog2Flow), assembles them into a modular prompt, and decodes slot values with constrained decoding. The experiments compare embedding extractors, user-only versus user-agent turn embeddings, dialogue history in prompts, speaker tags, number of demonstrations, and decoding strategy. The central findings are that user-only embeddings improve slot precision/recall over user-agent embeddings and that a general-purpose multilingual embedding model performs comparably to a dialogue-specific one.

Significance. If the findings are robust, they offer practical guidance for building retrieval-based ICL systems for DST with open-weight LLMs. The paper is transparent in using fully open models and reports a systematic factor analysis rather than a single pipeline. However, the contribution is currently limited because the evaluation assumes ground-truth domains, which is not the standard DST setting, and because the reported differences are mostly small and lack uncertainty estimates.

major comments (3)
  1. [Section 3 / Section 2.3] Section 3 states 'All experiments consider the ground-truth domain when evaluating slot prediction performance,' and Section 2.3 describes constrained decoding in which slot values are decoded after a domain is supplied. Consequently, the precision and recall in Tables 1, 3, and 4 measure slot-value extraction conditional on knowing the true domain, not whether the system identifies the correct domain set. Since Section 2.1 motivates retrieval through domain relevance, the retriever's effect on domain prediction is part of the story but is never scored. The headline result in Table 1 (User-only 69.9/80.6 vs User-Agent 67.8/78.3 at K=10) may not survive an end-to-end evaluation; reporting joint-goal accuracy or a domain-classification metric is necessary to support the claim that user-only retrieval improves DST performance.
  2. [Tables 1-3, Section 5] All tables report single-run point estimates without confidence intervals, standard deviations, or significance tests. Some of the central contrasts are small: Table 2 gives 67.8 vs 68.1 precision for speaker tags on OLMo, and Table 5 gives 73.2 vs 73.2 precision for LaBSE vs D2F at K=10. The conclusion in Section 5 that speaker tags have a 'significant effect' and the comparison between LaBSE and D2F are therefore not quantitatively supported. At minimum, the authors should report results over multiple random seeds or retrievers and a paired significance test, since precisely these small differences ground the recommendations.
  3. [Section 5(d) / Table 3] Section 5(d) states that 'Three or more demonstrations do not yield significantly better results,' but Table 3 shows OLMo precision increasing monotonically from 70.4 at K=3 to 71.6 at K=5 to 73.2 at K=10 while recall decreases from 81.6 to 80.0 to 79.3. The data indicate a precision-recall trade-off rather than a plateau; without a statistical test the wording 'do not yield significantly better results' is unsupported. This claim should be revised or substantiated.
minor comments (6)
  1. [Section 2.3] The sentence 'The prompt with the demonstration examples are passed as input the LLM' is missing 'to' and should read 'passed as input to the LLM.'
  2. [Table 5] The recall value '81.66' for D2F at K=3 has an extra decimal place compared with the other entries; this looks like a typo and should be corrected to '81.6' for consistency.
  3. [Table 2 caption] The model names in the caption are inconsistent: 'Mistral-7B-Instructv0.3' should be 'Mistral-7B-Instruct-v0.3' and 'LLama3.2-3BI' should be 'Llama3.2-3B-Instruct' or a consistent abbreviation.
  4. [Figures 4 and 5] The labels 'User-Agent' and 'User only' are used as hyphenated adjectives in the legends; consider using 'User-agent' and 'User-only' throughout for consistency.
  5. [Section 3.3.1] The statement that 'LaBSE outperforms D2F' when User-Agent turns are used is not accompanied by a numeric table entry; Figure 3 should be supplemented with explicit numbers or a table for this comparison.
  6. [Section 3.3.2] The manuscript does not explain how the retriever handles multiple candidate turns from the same dialogue or how ties are broken when K demonstrations are selected; a sentence clarifying the selection procedure would improve reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper is an empirical ablation study with no fitted parameters and no derivation chain that reduces to its inputs.

full rationale

This paper reports a systematic empirical comparison of demonstration-retrieval and prompt-construction choices for in-context-learning dialogue state tracking. There is no formal derivation, no fitted parameter that is later called a prediction, and no load-bearing self-citation. The headline result, that user-only embeddings for retrieval outperform user-agent embeddings (Table 1: 69.9 precision and 80.6 recall versus 67.8 and 78.3 at K=10), is a direct measurement on the MultiWOZ 2.4 test samples; it is not implied by the construction of the retriever or the prompt. The statement in Section 3 that "All experiments consider the ground-truth domain when evaluating slot prediction performance" is an evaluation-scoping decision: it means the reported precision and recall are conditional on knowing the domain, so the metric does not measure end-to-end DST including domain prediction. This is a genuine limitation of the evaluation, but it is not circular reasoning: the paper does not define its input in terms of the output, and the ground-truth-domain assumption does not force the observed ordering between User and User-Agent retrieval. Similarly, the decision to exclude zero-shot prompting because of poor performance is a stated design choice, not a fitted-input-called-prediction pattern. The relevance/coverage analysis in Section 4.1 uses the same slot-key definitions as the evaluation, but it is presented as an explanatory analysis rather than as the source of the headline claim. No equation in the paper reduces to an earlier equation, and no cited prior work by the same authors is invoked to justify the central claim. Therefore, no circular step is present.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

No new theoretical entities or fitted constants are introduced. The only free parameter is the number of demonstrations, which is swept. The key assumptions are the retrieval relevance of embeddings, the ground-truth domain evaluation, and the sufficiency of the prompt template.

free parameters (1)
  • Number of demonstrations K = 1, 3, 5, 10 (swept, not optimized)
    The paper varies the maximum number of demonstrations in the prompt and observes performance changes. It is a hand-chosen hyperparameter, not fitted to a held-out set.
assumptions (3)
  • domain assumption Cosine similarity between sentence embeddings of dialogue turns selects semantically relevant demonstrations.
    The retriever in Section 2.1 assumes that embedding similarity corresponds to usefulness for DST, which the paper partially validates through slot relevance/coverage analysis but does not prove.
  • domain assumption Ground-truth domain is available during evaluation.
    Section 3 states 'All experiments consider the ground-truth domain when evaluating slot prediction performance.' This makes the task easier and is a modeling choice that affects all reported numbers.
  • domain assumption The modular prompt template with JSON slot schema is a reasonable interface for ICL-based DST.
    Section 2.2 describes the template without comparing to alternative prompt formats, so the conclusions are conditional on this template design.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Factors affecting the in-context learning abilities of LLMs for dialogue state tracking." pith.science (2026). https://pith.science/paper/5ADEN3O7

@misc{pith2026250608753,
  author       = {Pith},
  title        = {Pith review of: Factors affecting the in-context learning abilities of LLMs for dialogue state tracking},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5ADEN3O7}},
  note         = {Machine review of arXiv:2506.08753}
}
read the original abstract

This study explores the application of in-context learning (ICL) to the dialogue state tracking (DST) problem and investigates the factors that influence its effectiveness. We use a sentence embedding based k-nearest neighbour method to retrieve the suitable demonstrations for ICL. The selected demonstrations, along with the test samples, are structured within a template as input to the LLM. We then conduct a systematic study to analyse the impact of factors related to demonstration selection and prompt context on DST performance. This work is conducted using the MultiWoZ2.4 dataset and focuses primarily on the OLMo-7B-instruct, Mistral-7B-Instruct-v0.3, and Llama3.2-3B-Instruct models. Our findings provide several useful insights on in-context learning abilities of LLMs for dialogue state tracking.

Figures

Figures reproduced from arXiv: 2506.08753 by the authors.

Figure 1
Figure 1. Block diagram of the ICL scheme used in our experiments. Factors influencing the shaded parts are studied in this paper. and outlining potential avenues for future work in the field of ICL-based DST. 2. Methodology This section provides a detailed step-by-step description of our proposed system [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Illustration of the prompt format used for DST with simple a instruction. Demonstrations consist of a User-Agent turns (olive), domain(s) (teal), and corresponding slot values (brown). After K such demonstrations, the test sample is presented in the same format (orange). 2.4. Evaluation metric We evaluate the performance of our system using precision and recall of the predicted slot values. Precision measures the pr… view at source ↗
Figure 3
Figure 3. Slot precision and recall for LaBSE and Dialog2Flow models when User-Agent and User only sentences are considered for embeddings for retrieval. This yields 56,778 training samples (turns) and 7372 test sam￾ples (turns). The training set serves as the source for demon￾stration retrieval in our in-context learning framework. 3.2. Large Language Models (LLMs) In this work, we focus mainly on the fully open-source OLMo￾… view at source ↗
Figures from the paper (2 more)
Figure 5
Figure 5. Figure 5: Demonstration relevance and coverage for User-Agent and User only dialogue history in 1, 3, 10 number of demonstrations from D2F based retriever [PITH_FULL_IMAGE:figures/full_fig_p004_5.png]
Figure 4
Figure 4. Figure 4: Demonstration relevance and coverage for User-Agent and User only dialogue history in 1, 3, 10 number of demonstrations from LaBSE based retriever. 3.4.3. Dialogue history in the prompt Here, we compare the impact of using user-agent turns or user￾only utterances as th…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 28 canonical work pages

  1. [1]

    Introduction Instruction-tuned large language models (LLMs) have demon- strated enhanced capabilities compared to traditional language models, enabling them to perform a broader range of tasks based on instructions [1, 2, 3]. A key capability of LLMs is in-context learning (ICL), which allows them to generalise to new tasks without requiring explicit fine...

    work page 2025

  2. [2]

    taxi", “restaurant

    Methodology This section provides a detailed step-by-step description of our proposed system. Figure 1 provides a high-level representation of the data flow and the interactions between different modules of the system. 2.1. Demonstration retriever The selection of appropriate demonstrations is crucial for ef- fective in-context learning, as the quality an...

    work page

  3. [3]

    User:" and “Agent:

    Experiments and results In this section, we present the experimental setup and results obtained from our evaluation of the proposed system. We inves- tigated the impact of various factors related to both the demon- stration retriever and the prompt structure on overall DST per- formance. The primary goal of these experiments is to identify the factors tha...

    work page

  4. [4]

    slot-key

    Analysis 4.1. Relevance and coverage of demonstrations Here, we analyse the relevance and coverage of the selected demonstrations with respect to the slots that need to be predicted in the target instance. To quantify the relevance and coverage of demonstrations, we compare the slot-keys (e.g. restaurant- name, restaurant-people, taxi-departure) present i...

    work page

  5. [5]

    Conclusion In this paper, we systematically studied various factors that in- fluence the in-context learning abilities of LLMs for dialogue state tracking. We found that (a) demonstration retrievers based on a general-purpose embedding model such as LaBSE perform as good as dialogue-specific modes like D2F when using a max- imum of 10 demonstrations. (b) ...

    work page

  6. [6]

    ELOQUENCE

    Acknowledgements The work was supported by European Union’s Horizon Eu- rope project No. SEP-210943216 “ELOQUENCE”, Czech Ministry of Interior project No. VK01020132 “112”, Euro- pean Defence Fund project ARCHER, and by Czech Min- istry of Education, Youth and Sports (MoE) through the OP JAK project “Linguistics, Artificial Intelligence and Language and S...

    work page

  7. [7]

    In-Context Learning for Few-Shot Dialogue State Track- ing,

    Y . Hu, C.-H. Lee, T. Xie, T. Yu, N. A. Smith, and M. Osten- dorf, “In-Context Learning for Few-Shot Dialogue State Track- ing,” in Findings of the Association for Computational Linguis- tics: EMNLP 2022. Abu Dhabi, United Arab Emirates: Associ- ation for Computational Linguistics, Dec. 2022, pp. 2627–2643

    work page 2022

  8. [8]

    Training lan- guage models to follow instructions with human feedback,

    L. Ouyang, J. Wu, X. Jiang, D. Almeida, C. L. Wainwright, P. Mishkin, C. Zhang, S. Agarwal, K. Slama, A. Ray, J. Schul- man, J. Hilton, F. Kelton, L. Miller, M. Simens, A. Askell, P. Welinder, P. Christiano, J. Leike, and R. Lowe, “Training lan- guage models to follow instructions with human feedback,” 2022, arXiv:2203.02155

    arXiv 2022

Show all 33 references
  1. [9]

    Scaling instruction-finetuned language models,

    H. W. Chung, L. Hou, S. Longpre, B. Zoph, Y . Tay, W. Fedus, Y . Li, X. Wang, M. Dehghani, S. Brahma, A. Webson, S. S. Gu, Z. Dai, M. Suzgun, X. Chen, A. Chowdhery, A. Castro-Ros, M. Pellat, K. Robinson, D. Valter, S. Narang, G. Mishra, A. Yu, V . Zhao, Y . Huang, A. Dai, H. Y...

    2022 arXiv

  2. [10]

    Finetuned Language Models are Zero- Shot Learners,

    J. Wei, M. Bosma, V . Zhao, K. Guu, A. W. Yu, B. Lester, N. Du, A. M. Dai, and Q. V . Le, “Finetuned Language Models are Zero- Shot Learners,” in International Conference on Learning Repre- sentations, 2022

    2022

  3. [11]

    Language models are few-shot learners,

    T. B. Brown, B. Mann, N. Ryder, M. Subbiah, J. Kaplan, P. Dhari- wal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell, S. Agarwal, A. Herbert-V oss, G. Krueger, T. Henighan, R. Child, A. Ramesh, D. M. Ziegler, J. Wu, C. Winter, C. Hesse, M. Chen, E. Sigler, M. Litwin, S. Gray, ...

    2020

  4. [12]

    A Survey on In- context Learning,

    Q. Dong, L. Li, D. Dai, C. Zheng, J. Ma, R. Li, H. Xia, J. Xu, Z. Wu, B. Chang, X. Sun, L. Li, and Z. Sui, “A Survey on In- context Learning,” in Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing . Miami, Florida, USA: Association for Compu...

    2024

  5. [13]

    OpenICL: An Open-Source Framework for In-context Learn- ing,

    Z. Wu, Y . Wang, J. Ye, Z. Wu, J. Feng, J. Xu, and Y . Qiao, “OpenICL: An Open-Source Framework for In-context Learn- ing,” in Proceedings of the 61st Annual Meeting of the Associ- ation for Computational Linguistics (Volume 3: System Demon- strations). Toronto, Canada: Associ...

    2023

  6. [14]

    Stabilized In-Context Learning with Pre-trained Language Models for Few Shot Dialogue State Tracking,

    D. Chen, K. Qian, and Z. Yu, “Stabilized In-Context Learning with Pre-trained Language Models for Few Shot Dialogue State Tracking,” in Findings of the Association for Computational Lin- guistics: EACL 2023. Dubrovnik, Croatia: Association for Com- putational Linguistics, May ...

    2023

  7. [15]

    Recent Neural Methods on Dialogue State Tracking for Task-Oriented Dialogue Systems: A Survey,

    V . Balaraman, S. Sheikhalishahi, and B. Magnini, “Recent Neural Methods on Dialogue State Tracking for Task-Oriented Dialogue Systems: A Survey,” in Proceedings of the 22nd Annual Meeting of the Special Interest Group on Discourse and Dialogue. Singa- pore and Online: Associa...

    2021

  8. [16]

    What Makes Good In-Context Examples for GPT-3?

    J. Liu, D. Shen, Y . Zhang, B. Dolan, L. Carin, and W. Chen, “What Makes Good In-Context Examples for GPT-3?” inProceedings of Deep Learning Inside Out (DeeLIO 2022): The 3rd Workshop on Knowledge Extraction and Integration for Deep Learning Archi- tectures. Dublin, Ireland an...

    2022

  9. [17]

    Finding Support Examples for In-Context Learning,

    X. Li and X. Qiu, “Finding Support Examples for In-Context Learning,” in Findings of the Association for Computational Lin- guistics: EMNLP 2023 . Singapore: Association for Computa- tional Linguistics, Dec. 2023, pp. 6219–6235

    2023

  10. [18]

    Active Example Selection for In- Context Learning,

    Y . Zhang, S. Feng, and C. Tan, “Active Example Selection for In- Context Learning,” inProceedings of the 2022 Conference on Em- pirical Methods in Natural Language Processing . Abu Dhabi, United Arab Emirates: Association for Computational Linguis- tics, Dec. 2022, pp. 9134–9148

    2022

  11. [19]

    In this work, we study a nearest-neighbour-based retrieval to select relevant demonstrations (dialogue turns) for ICL-based DST

    found that the ground truth slot values in the demonstra- tions may not be as important as previously thought, as replac- ing them with random words did not significantly degrade per- formance, suggesting that the structure and context provided by the demonstrations may be mor...

    2025 arXiv

  12. [20]

    Data Curation Alone Can Stabilize In- context Learning,

    T.-Y . Chang and R. Jia, “Data Curation Alone Can Stabilize In- context Learning,” in Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Toronto, Canada: Association for Computational Lin- guistics, Jul. 2023, pp. 8123–8144

    2023

  13. [21]

    Revisiting demonstration selection strategies in in- context learning,

    K. Peng, L. Ding, Y . Yuan, X. Liu, M. Zhang, Y . Ouyang, and D. Tao, “Revisiting demonstration selection strategies in in- context learning,” in Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Bangkok, Thailand:...

    2024

  14. [22]

    Unified Demonstration Retriever for In-Context Learn- ing,

    X. Li, K. Lv, H. Yan, T. Lin, W. Zhu, Y . Ni, G. Xie, X. Wang, and X. Qiu, “Unified Demonstration Retriever for In-Context Learn- ing,” in Proceedings of the 61st Annual Meeting of the Associ- ation for Computational Linguistics (Volume 1: Long Papers) . Toronto, Canada: Assoc...

    2023

  15. [23]

    Comparative Analysis of Demonstra- tion Selection Algorithms for LLM In-Context Learning,

    D. Shu and M. Du, “Comparative Analysis of Demonstra- tion Selection Algorithms for LLM In-Context Learning,” 2024, arXiv:2410.23099

    2024 arXiv

  16. [24]

    Learning to retrieve prompts for in-context learning,

    O. Rubin, J. Herzig, and J. Berant, “Learning to retrieve prompts for in-context learning,” inProceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies . Seattle, United States: Association fo...

    2022

  17. [25]

    District: Dialogue state tracking with retriever driven in-context tuning,

    P. Venkateswaran, E. Duesterwald, and V . Isahagian, “District: Dialogue state tracking with retriever driven in-context tuning,” in Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, 2023, pp. 5101–5112

    2023

  18. [26]

    Rethinking the Role of Demonstra- tions: What Makes In-Context Learning Work?

    S. Min, X. Lyu, A. Holtzman, M. Artetxe, M. Lewis, H. Ha- jishirzi, and L. Zettlemoyer, “Rethinking the Role of Demonstra- tions: What Makes In-Context Learning Work?” in Proceedings of the 2022 Conference on Empirical Methods in Natural Lan- guage Processing (EMNLP). Abu Dhab...

    2022

  19. [27]

    Language-agnostic BERT sentence embedding,

    F. Feng, Y . Yang, D. Cer, N. Arivazhagan, and W. Wang, “Language-agnostic BERT sentence embedding,” in Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Dublin, Ireland: Associa- tion for Computational Linguistics...

    2022

  20. [28]

    Dialog2Flow: Pre- training soft-contrastive action-driven sentence embeddings for automatic dialog flow extraction,

    S. Burdisso, S. Madikeri, and P. Motlicek, “Dialog2Flow: Pre- training soft-contrastive action-driven sentence embeddings for automatic dialog flow extraction,” in Proceedings of the 2024 Conference on Empirical Methods in Natural Language Process- ing. Miami, Florida, USA: As...

    2024

  21. [29]

    MultiWOZ - a large-scale multi- domain Wizard-of-Oz dataset for task-oriented dialogue mod- elling,

    P. Budzianowski, T.-H. Wen, B.-H. Tseng, I. Casanueva, S. Ultes, O. Ramadan, and M. Gaši ´c, “MultiWOZ - a large-scale multi- domain Wizard-of-Oz dataset for task-oriented dialogue mod- elling,” in Proceedings of the 2018 Conference on Empirical Methods in Natural Language Pro...

    2018

  22. [30]

    MultiWOZ 2.4: A Multi-Domain Task-Oriented Dialogue Dataset with Essential Annotation Corrections to Improve State Tracking Evaluation,

    F. Ye, J. Manotumruksa, and E. Yilmaz, “MultiWOZ 2.4: A Multi-Domain Task-Oriented Dialogue Dataset with Essential Annotation Corrections to Improve State Tracking Evaluation,” in Proceedings of the 23rd Annual Meeting of the Special Interest Group on Discourse and Dialogue. E...

    2022

  23. [31]

    OLMo: Accelerating the science of language mod- els,

    D. Groeneveld, I. Beltagy, E. Walsh, A. Bhagia, R. Kinney, O. Tafjord, A. Jha, H. Ivison, I. Magnusson, Y . Wang, S. Arora, D. Atkinson, R. Authur, K. Chandu, A. Cohan, J. Dumas, Y . Elazar, Y . Gu, J. Hessel, T. Khot, W. Merrill, J. Morri- son, N. Muennighoff, A. Naik, C. Nam...

    2024

  24. [32]

    Mistral 7B,

    A. Q. Jiang, A. Sablayrolles, A. Mensch, C. Bamford, D. S. Chaplot, D. de las Casas, F. Bressand, G. Lengyel, G. Lample, L. Saulnier, L. R. Lavaud, M.-A. Lachaux, P. Stock, T. L. Scao, T. Lavril, T. Wang, T. Lacroix, and W. E. Sayed, “Mistral 7B,” 2023, arXiv:2310.06825

    2023 arXiv

  25. [33]

    The Llama 3 Herd of Models,

    A. Grattafiori, A. Dubey, A. Jauhri, A. Pandey, A. Kadian, A. Al- Dahle, A. Letman, A. Mathur, A. Schelten, A. Vaughan, A. Yang, A. Fan, A. Goyal, A. Hartshorn, A. Yang, A. Mitra, A. Sra- vankumar, A. Korenev, A. Hinsvark, A. Rao, A. Zhang, A. Ro- driguez, A. Gregerson, A. Spa...

    2024 arXiv

Pith tools

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