Pith. sign in

REVIEW 4 major objections 5 minor 15 references

The paper argues that value vectors from a mid-layer transformer block, contrastively distilled, let any natural-language type description retrieve the matching entity mentions in zero-shot.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

Mid-layer LLM value vectors, projected through a contrastively trained MLP, enable zero-shot retrieval of documents by ad-hoc entity type.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection Genuinely useful idea—mid-layer value vectors plus contrastive projection for entity retrieval—but the Few-NERD headline is partly selected on the test set, and the abstract overstates the NERetrieve result. the 4 major comments →

arxiv 2509.04011 v1 pith:U7F77VJA submitted 2025-09-04 cs.IR cs.AIcs.CL

NER Retriever: Zero-Shot Named Entity Retrieval with Type-Aware Embeddings

Or Shachar , Uri Katz , Yoav Goldberg , Oren Glickman This is my paper
classification cs.IR cs.AIcs.CL
keywords named entity retrievalzero-shot entity typingtype-aware embeddingsmid-layer representationscontrastive learningdense retrievalLLM internal statesentity span detection
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

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 takes on ad-hoc named entity retrieval: finding every document that mentions an entity of a user-defined type, where the type is given as a free-text description such as "dinosaur" or "airline" rather than a fixed schema. The authors try to show that the internal value vectors of a mid-layer transformer block—specifically block 17 of LLaMA 3.1 8B—carry entity-type information that final layers dilute, and that a lightweight contrastive projection can turn those vectors into a shared space where type descriptions and entity mentions sit close together. If the approach is right, it would make schema-free, zero-shot entity retrieval practical: index once with per-entity embeddings, then answer arbitrary type queries by nearest-neighbor search, without fine-tuning the LLM. Experiments on three benchmarks report that the retrieval stage outperforms lexical and sentence-level dense baselines on Few-NERD and MultiCoNER 2, and matches them on NERetrieve, with the whole system bottlenecked mainly by the upfront span detector.

Core claim

The central discovery is that representation choice inside a frozen LLM matters more than the usual final-layer default. Across 416 representation sources in LLaMA 3.1 8B, the value (V) vectors from self-attention block 17 give the best same-type-versus-different-type separation (AUC 0.78); a similar mid-layer peak appears in other tested architectures after normalizing depth. Feeding those vectors through a two-layer MLP trained with triplet loss produces a 500-dimensional type-aware embedding space in which entity mentions of the same type cluster, related fine-grained types sit nearby, and a user's type description retrieves entities by cosine similarity. With the full pipeline, R-Precisi

What carries the argument

The value (V) vectors from transformer block 17 of LLaMA 3.1 8B—the self-attention value-projection output at the final token of each entity span—passed through a two-layer SiLU MLP trained with triplet contrastive loss into a 500-dimensional space, then searched by cosine nearest neighbor. This combination supplies the shared embedding space that aligns arbitrary type descriptions with the corresponding entity mentions.

Load-bearing premise

The system's retrieval completeness rests on a category-agnostic entity span detector finding essentially all mentions; its measured coverage is 89–94%, and the missing spans cost about 11 percentage points of average R-Precision when replaced by gold spans.

What would settle it

Take a corpus with gold entity spans and a query set of unseen types. Run NER Retriever with oracle spans using (a) block-17 value vectors and (b) final-layer embeddings, both through the same trained projection. If (b) matches or beats (a), the central layer-selection claim fails. Alternatively, lower automatic span coverage below about 0.8 and show that end-to-end R-Precision collapses to baseline levels, which would confirm the detector bottleneck.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • On Few-NERD and MultiCoNER 2, the full system beats BM25 and strong dense sentence embedders by roughly three to four times in R-Precision (0.34 vs 0.08; 0.32 vs 0.09).
  • With gold entity spans, the retrieval stage outperforms every baseline on all three benchmarks, so retrieval quality itself is not the current ceiling.
  • Choosing span-final-token representations over EOS or sentence tokens is decisive: 0.19 vs 0.03 in the no-MLP ablation on Few-NERD.
  • The contrastive projection contributes about half the retrieval score (0.16 without MLP vs 0.34 with it), and compresses vectors to 500 dimensions, cutting the index from 9.2 GB to 2 GB on MultiCoNER 2.
  • A zero-shot user never needs to predefine types: any natural-language type description is mapped through the same frozen LLM and projection, so novel and long-tail types can be queried directly.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • An implicit consequence: since type sensitivity peaks in mid-layers across several architectures, representation selection could improve other embedding-based tasks that currently default to top-layer pooling—entity linking, fine-grained typing, and open-domain question answering.
  • The training recipe uses only weakly labeled paragraphs from NERetrieve, so extending the same contrastive projection to specialized domains (law, medicine, finance) is a direct path to closing the limitation the authors state.
  • Because the index stores one vector per entity mention, it can be reused as evidence for retrieval-augmented agents: a broad query like "texts mentioning a dinosaur" can be decomposed into type-vector searches, which the paper gestures at but does not implement.
  • The 11-point oracle-span gap suggests a concrete engineering target: any better category-agnostic span detector should translate almost one-for-one into retrieval gains, offering a cheaper alternative to improving embeddings.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes NER Retriever, a zero-shot named-entity retrieval framework. Instead of embedding entire sentences or documents, it indexes individual entity spans using the value vectors from block 17 of a frozen LLaMA 3.1 8B model, followed by a lightweight MLP trained with a triplet contrastive loss. At query time, a natural-language type description is embedded through the same pipeline and matched against the indexed entity vectors by cosine similarity. The method is evaluated on Few-NERD, MultiCoNER 2, and NERetrieve, reporting large gains over BM25 and sentence-level dense retrievers on the first two benchmarks and comparable performance on NERetrieve. The paper also includes a layer/component sweep across several LLMs, ablations, and an oracle-span analysis.

Significance. If the central claims hold, the paper makes a useful contribution: it provides evidence that mid-layer value vectors contain more type-discriminative information than top-layer outputs, and it shows that a compact projection of those vectors supports efficient, schema-free retrieval. The manuscript is generally well organized, includes a public codebase, and reports several useful ablations, including layer choice, token selection, and MLP projection. The oracle-span comparison is a good practice. However, the main quantitative evidence for the 'substantially outperforms' claim is weakened by selection on the test benchmark and by a post-hoc data filter, and the abstract overstates the NERetrieve result. These issues are fixable but require re-evaluation.

major comments (4)
  1. [§4, §5.2, Table 2] The Few-NERD evaluation is not an independent test of the method. Section 4 selects the representation source (block 17, V vectors) by maximizing type-discrimination AUC on Few-NERD, and Section 5.2 states that all numerical hyperparameters were selected by evaluating the same type-discrimination task described in Section 4, which uses Few-NERD. Few-NERD is then a test set in Table 2. This creates a selection effect: the reported 0.34 R-Precision and the 'substantially outperforms all baselines' claim for Few-NERD are not externally validated. Please re-run the layer/component selection and hyperparameter tuning on a disjoint development set (e.g., held-out Few-NERD types or another dataset) and re-report Table 2 with the test set used only once. The ablation in §8.1 inherits the same issue.
  2. [§6.3, Few-NERD] The Few-NERD sentence-length exclusion is post-hoc. The paper states that instances with fewer than five words are omitted because 'such very short sentences were found to harm the performance of sentence-encoder baselines while not impacting our method.' This is an evaluation-set modification made after observing baseline behavior, and it is not applied to MultiCoNER 2 or NERetrieve. The reported Few-NERD score may depend on this filter. Please report results on the full Few-NERD set and on a pre-specified filtered version, and justify the filter on data-quality grounds rather than baseline performance. Also state whether the significance claim survives on the unfiltered set.
  3. [Abstract and §7] The abstract's blanket claim that 'NER Retriever significantly outperforms both lexical and dense sentence-level retrieval baselines' is contradicted by Table 2 and §7. On NERetrieve, NER Retriever achieves 0.28 R-Precision versus 0.29 for NV-Embed v2, and the paper itself says the difference is not statistically significant. The significance markers in Table 2 apply only to Few-NERD and MultiCoNER 2. Please revise the abstract and any summary statements to distinguish 'substantially outperforms on Few-NERD and MultiCoNER 2' from 'comparable on NERetrieve', or provide evidence for a stronger claim.
  4. [§6.1, §7, Appendix A.1] The end-to-end claim depends heavily on the automatic entity-span detector, but that dependency is only partially characterized. Appendix A.1 reports coverage of 0.89–0.94 across datasets, and §7 reports that oracle spans improve average R-Precision by roughly 11% (0.34 to 0.37 on Few-NERD, 0.32 to 0.35 on MultiCoNER 2, 0.28 to 0.34 on NERetrieve). Missed spans place an upper bound on recall that is independent of embedding quality. Please provide a sensitivity analysis over detector quality and state explicitly in the Limitations section that the reported automatic-system numbers are contingent on the extractor, not solely on the retrieval method.
minor comments (5)
  1. [Figure 1] The caption contains a typo: 'entity focued' should be 'entity-focused'. The example text also alternates between 'Dinosaur' and 'dinosaurs'; standardize the query type.
  2. [Table 3 and §6.2] The baseline is called 'NV-Embed v2' in most of the paper but 'Nvidia NV-Embed' in Table 3. Use one name consistently. Also, §6.2 says E5-Mistral and NV-Embed v2 'have comparable architecture and parameter count to LLaMA 3.1'; E5-Mistral is based on Mistral, not LLaMA, so clarify the comparison.
  3. [§4] The layer sweep uses 20 sampled fine-grained types and 20 sentences each. Please specify whether the sample comes from the Few-NERD train or test split, and report the exact type list or sampling seed; otherwise the selection procedure is not reproducible.
  4. [§5.2] The paper says the input layer size is 'derived from the LLM output (in our case 1024)' but LLaMA 3.1 8B has hidden size 4096. Clarify that 1024 is the value-vector dimension for the selected block, not the full hidden size.
  5. [§8.2 and Table 4] The text reports the EOS-based result as 0.03 R-Precision, while Table 4 reports 0.02 for '17V + EOS (-No MLP)'. Please correct the inconsistency.

Circularity Check

3 steps flagged

Few-NERD 'prediction' is selected on Few-NERD: layer, hyperparameters, and the short-sentence filter are all fitted to the same benchmark that is then reported as the headline result; the central claim survives via MultiCoNER 2.

specific steps
  1. fitted input called prediction [Section 4 ('Evaluation setup' / 'Findings') + Section 6.3 (Few-NERD) + Table 2]
    "Evaluation setup. We use the Few-NERD dataset to evaluate type sensitivity across different LLM layers and components... The highest AUC is achieved by the value (V) vectors in the self-attention module of block 17... Based on these findings, we use the value (V) vectors from block 17 of LLaMA 3.1 8B as the basis for entity embeddings in our retrieval system. ... Few-NERD (supervised) (Ding et al., 2021): A manually annotated dataset covering 66 fine-grained entity types across 188K Wikipedia sentences."

    The central representation choice (block 17, value vectors) is selected by maximizing type-discrimination AUC on Few-NERD (Section 4), and Few-NERD is then used as one of the three evaluation benchmarks (Section 6.3, Table 2, R-Precision 0.34). The headline Few-NERD result is therefore an in-sample, selection-optimized estimate rather than an external prediction: the layer/component was fitted to Few-NERD before the same dataset was used to validate the method. The claim 'substantially outperforms all baselines on Few-NERD' is not independently established on Few-NERD; independent support must come from MultiCoNER 2, which was not used for the selection.

  2. fitted input called prediction [Section 5.2 'Implementation details' (hyperparameters), with Section 4's Few-NERD discrimination task; Few-NERD is also a test set in Section 6.3]
    "All numerical hyperparameters were chosen by independently evaluating a range of values and selecting those yielding the best performance on the entity type discrimination task (Described in Section 4)."

    All numerical hyperparameters (output dimension, hidden size, dropout, etc.) are tuned on the Section 4 discrimination task, which is defined on Few-NERD. Few-NERD is simultaneously a test benchmark for the end-to-end retrieval evaluation (Section 6.3, Table 2). Thus the Few-NERD retrieval numbers are produced by hyperparameter tuning on the same dataset used for evaluation, compounding the layer-selection effect: the reported 0.34 is partly manufactured by a choice procedure that used the test benchmark, not achieved by an independently configured model.

  3. other [Section 6.3 'Evaluation Datasets' (Few-NERD paragraph)]
    "For fairness in comparison, we omit instances with fewer than five words, as such very short sentences were found to harm the performance of sentence-encoder baselines while not impacting our method."

    The Few-NERD evaluation set is altered after observing the outcomes of the comparison: short sentences are removed because they were 'found to harm the performance of sentence-encoder baselines while not impacting our method.' This is a result-dependent modification of the test set, made with knowledge of how the two sides perform. The reported Few-NERD advantage (0.34 vs 0.04/0.08) reflects a benchmark adjusted in light of the method's own results. Jointly with Section 4 and Section 5.2, the Few-NERD comparison is not a clean external test of the method.

full rationale

NER Retriever's derivation chain has two parts: (i) an empirical finding that mid-layer value vectors carry entity-type information, and (ii) a retrieval system using block-17 V-vectors plus a contrastive MLP. Part (i) is supported by an independent sweep over 416 representation sources across four LLMs, and the mid-layer peak appears across architectures. Part (ii) is evaluated on three benchmarks, and the circularity is confined to the Few-NERD leg. Specifically, Section 4 selects the layer/component by maximizing type-discrimination AUC on Few-NERD; Section 5.2 selects all numerical hyperparameters on that same Few-NERD-based task; Section 6.3 additionally drops short Few-NERD sentences after observing that they hurt the baselines more than the method. Presenting Few-NERD R-Precision (0.34) as the headline 'substantially outperforms' result is therefore a fitted-input-called-prediction: the configuration was chosen on the same benchmark then used to claim superiority. This does not make the entire paper circular. MultiCoNER 2 was not used in any selection step, and NER Retriever beats all baselines there (0.32 vs 0.07-0.09), giving independent support for the core claim. The NERetrieve test split is disjoint from the contrastive-training split (though from the same research group via Katz et al. 2023), and the paper honestly reports no significant advantage there (0.28 vs 0.29), which argues against outcome-dependent reporting. The self-citation to Katz et al. supplies the task definition and training data, not the load-bearing argument for mid-layer representations. Accordingly, the score is 4: one benchmark's headline claim is partially circular through test-set selection and post-hoc filtering, but the central claim retains independent content. I also note the explicit in-text admissions (Section 5.2 and Section 6.3), which are concrete evidence rather than speculation about intent.

Axiom & Free-Parameter Ledger

6 free parameters · 3 axioms · 0 invented entities

The central claim rests on a small set of hyperparameters tuned on a benchmark that is also used for evaluation, an unverified dependency on a third-party entity detector, and the assumption that NERetrieve training types generalize to all test types. No new physical or conceptual entities are introduced.

free parameters (6)
  • LLM layer and component for entity representation = block 17, value (V) vectors of LLaMA 3.1 8B
    Selected by maximizing type-discrimination AUC on a Few-NERD subsample (Section 4), then used for all retrieval experiments including Few-NERD itself.
  • Projection output dimension = 500
    Chosen to reduce storage while preserving capacity; hyperparameters selected on the discrimination task (Section 5.2).
  • Hidden layer dimension = 500
    Matched to output dimension (Section 5.2).
  • Dropout rate = 0.1
    Section 5.2, selected by evaluation on the discrimination task.
  • Hard-negative ratio within each batch = 0.1
    Section 5.2, 10% of negatives are BM25-based hard negatives.
  • Number of triplets per type = 5000 (2M total)
    Section 5.1, set to construct the training set from NERetrieve train split.
axioms (3)
  • domain assumption Entity mentions are adequately represented by the value vector of the final token of the entity span at layer 17 of LLaMA 3.1 8B.
    Motivated by the layer sweep in Section 4, but the sweep itself uses a small sample (20 types, 20 sentences each) from Few-NERD; transfer to other datasets and domains is assumed.
  • domain assumption The CascadeNER extractor identifies all (or most) entity spans without using the evaluation benchmarks in training.
    Section 6.1 and Appendix A.1. The pipeline depends on this classifier; the authors state none of the evaluation datasets were used in its training, but this cannot be verified from the paper. Detection coverage is 0.89 to 0.94, leaving some mentions unindexed.
  • domain assumption NERetrieve training data provides a representative distribution of entity types and contexts for learning a type-aware space that generalizes to held-out types.
    Section 5.1. The projection is trained only on NERetrieve training types; zero-shot generalization to types in Few-NERD, MultiCoNER 2, and the NERetrieve test set is assumed.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of NER Retriever: Zero-Shot Named Entity Retrieval with Type-Aware Embeddings." pith.science (2026). https://pith.science/paper/U7F77VJA

@misc{pith2026250904011,
  author       = {Pith},
  title        = {Pith review of: NER Retriever: Zero-Shot Named Entity Retrieval with Type-Aware Embeddings},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/U7F77VJA}},
  note         = {Machine review of arXiv:2509.04011}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

We present NER Retriever, a zero-shot retrieval framework for ad-hoc Named Entity Retrieval, a variant of Named Entity Recognition (NER), where the types of interest are not provided in advance, and a user-defined type description is used to retrieve documents mentioning entities of that type. Instead of relying on fixed schemas or fine-tuned models, our method builds on internal representations of large language models (LLMs) to embed both entity mentions and user-provided open-ended type descriptions into a shared semantic space. We show that internal representations, specifically the value vectors from mid-layer transformer blocks, encode fine-grained type information more effectively than commonly used top-layer embeddings. To refine these representations, we train a lightweight contrastive projection network that aligns type-compatible entities while separating unrelated types. The resulting entity embeddings are compact, type-aware, and well-suited for nearest-neighbor search. Evaluated on three benchmarks, NER Retriever significantly outperforms both lexical and dense sentence-level retrieval baselines. Our findings provide empirical support for representation selection within LLMs and demonstrate a practical solution for scalable, schema-free entity retrieval. The NER Retriever Codebase is publicly available at https://github.com/ShacharOr100/ner_retriever

Figures

Figures reproduced from arXiv: 2509.04011 by Oren Glickman, Or Shachar, Uri Katz, Yoav Goldberg.

Figure 1
Figure 1. Figure 1: Example use case for ad-hoc Named Entity [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: NER Retriever high-level architecture. During indexing (green, dashed), entity spans are embedded and [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Entity-type discrimination AUC scores for [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: 2D UMAP projection of type-aware entity embeddings produced by our model, visualized for the top 25 [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

15 extracted references · 9 canonical work pages

  1. [3]

    arXiv preprint arXiv:2010.00904

    Autoregressive entity retrieval. arXiv preprint arXiv:2010.00904. Ning Ding, Guangwei Xu, Yulin Chen, Xiaobin Wang, Xu Han, Pengjun Xie, Haitao Zheng, and Zhiyuan Liu

    Pith/arXiv arXiv 2010

  2. [4]

    In Find- ings of the Association for Computational Linguis- tics: EMNLP 2023 , pages 2027–2051, Singapore

    Multi- CoNER v2: a large multilingual dataset for fine- grained and noisy named entity recognition. In Find- ings of the Association for Computational Linguis- tics: EMNLP 2023 , pages 2027–2051, Singapore. Association for Computational Linguistics. Faegheh Hasibi, Fedor Nikolaev, Chenyan Xiong, Krisz- tian Balog, Svein Erik Bratsberg, Alexander Kotov, an...

    work page 2023

  3. [5]

    In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6769–6781, Online

    Dense passage retrieval for open- domain question answering. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6769–6781, Online. Association for Computational Linguistics. Uri Katz, Matan Vetzler, Amir Cohen, and Yoav Gold- berg

    work page 2020

  4. [6]

    In Find- ings of the Association for Computational Linguis- tics: EMNLP 2023 , pages 3340–3354, Singapore

    NERetrieve: Dataset for next genera- tion named entity recognition and retrieval. In Find- ings of the Association for Computational Linguis- tics: EMNLP 2023 , pages 3340–3354, Singapore. Association for Computational Linguistics. Chankyu Lee, Rajarshi Roy, Mengyao Xu, Jonathan Raiman, Mohammad Shoeybi, Bryan Catanzaro, and Wei Ping

    work page 2023

  5. [8]

    In Pro- ceedings of the 2024 Conference on Empirical Meth- ods in Natural Language Processing , pages 2913– 2932, Miami, Florida, USA

    Evaluating D-MERIT of partial-annotation on information retrieval. In Pro- ceedings of the 2024 Conference on Empirical Meth- ods in Natural Language Processing , pages 2913– 2932, Miami, Florida, USA. Association for Compu- tational Linguistics. Nils Reimers and Iryna Gurevych

    work page 2024

  6. [12]

    arXiv preprint arXiv:2401.00368

    Improving text embeddings with large language models. arXiv preprint arXiv:2401.00368. Lilian Weng

    Pith/arXiv arXiv

  7. [14]

    In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP) , pages 6397–6407, Online

    Scalable zero- shot entity linking with dense entity retrieval. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP) , pages 6397–6407, Online. Association for Computa- tional Linguistics. Tingyu Xie, Qi Li, Jian Zhang, Yan Zhang, Zuozhu Liu, and Hongwei Wang

    work page 2020

  8. [15]

    In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 7935–7956, Singapore

    Empirical study of zero-shot NER with ChatGPT. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 7935–7956, Singapore. Association for Computational Linguistics. Sheng Zhang, Hao Cheng, Jianfeng Gao, and Hoifung Poon

    work page 2023

  9. [2003]

    In Proceedings of the Seventh Conference on Natural Language Learning at HLT-NAACL 2003, pages 142–

    Introduction to the CoNLL-2003 shared task: Language-independent named entity recognition. In Proceedings of the Seventh Conference on Natural Language Learning at HLT-NAACL 2003, pages 142–

    work page 2003

  10. [2019]

    Sentence- BERT: Sentence embeddings using Siamese BERT- networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natu- ral Language Processing (EMNLP-IJCNLP), pages 3982–3992, Hong Kong, China. Association for Com- putational Linguistics. Stephen Robertson and Hugo Zaragoza

    work page 2019

  11. [2020]

    CoRR, abs/2009.07317

    Cascaded models for better fine-grained named entity recognition. CoRR, abs/2009.07317. Eunsol Choi, Omer Levy, Yejin Choi, and Luke Zettle- moyer

    Pith/arXiv arXiv 2009

  12. [2021]

    In Proceedings of the 2021 Conference on Empirical Methods in Natu- ral Language Processing, pages 6138–6148

    Simple entity-centric ques- tions challenge dense retrievers. In Proceedings of the 2021 Conference on Empirical Methods in Natu- ral Language Processing, pages 6138–6148. Satoshi Sekine

    work page 2021

  13. [2023]

    arXiv preprint arXiv:2305.15444

    Prompt- ner: Prompting for named entity recognition. arXiv preprint arXiv:2305.15444. Parul Awasthy, Taesun Moon, Jian Ni, and Radu Flo- rian

    Pith/arXiv arXiv

  14. [2024]

    arXiv preprint arXiv:2405.17428

    Nv-embed: Improved techniques for training llms as generalist embedding models. arXiv preprint arXiv:2405.17428. Chaitanya Malaviya, Peter Shaw, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova

    Pith/arXiv arXiv

  15. [2025]

    arXiv preprint arXiv:2508.07999

    Widesearch: Bench- marking agentic broad info-seeking. arXiv preprint arXiv:2508.07999. Ledell Wu, Fabio Petroni, Martin Josifoski, Sebastian Riedel, and Luke Zettlemoyer

    Pith/arXiv arXiv

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.