Pith. sign in

REVIEW 4 major objections 6 minor 56 references

Selecting and Merging: Towards Adaptable and Scalable Named Entity Recognition with Large Language Models

T0 review · 4 major / 6 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 The paper claims that selecting and merging a small set of LoRA-trained domain experts at inference time outperforms a fully trained unified model for zero-shot named entity recognition.

desk verdict A sensible selection-and-merging framework for NER that is well ablated, but the headline 10% gain is against a deliberately under-tuned baseline and the merged-expert count is tuned on test domains. read the letter →

arxiv 2506.22813 v1 pith:4Z6IDRN2 submitted 2025-06-28 cs.CL

Zhuojun Ding , Wei Wei , Chenghao Fan This is my paper · ORCID
classification cs.CL
keywords namedentityrecognitionmodelmergingLoRAzero-shotdomainadaptationexpertselectiontaskvectorslargelanguagemodels
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

The paper argues that the usual recipe for zero-shot named entity recognition with large language models, training one unified model on all available domains, leaves transfer performance on the table. Its SaM framework instead trains a separate LoRA expert on each source domain, then at inference time selects a small subset of experts for the target domain and merges their parameters into a task-specific model. Two selection signals are combined: embedding similarity between domain centroids and target text, and each expert's score on a handful of target-domain sentences using pseudo-labels. The paper reports that this procedure beats the fully trained unified model by about 10% on average across seven domains, with gains up to 20%. The broader point is that for multi-domain NER, model merging plus targeted selection can be a cheaper and more adaptable alternative to data merging.

What carries the argument

The framework rests on the task-vector view of fine-tuning: each expert is represented by the parameter difference $\delta_{\text{sft}} = \theta_{\text{sft}} - \theta_{\text{base}}$ between the supervised fine-tuned model and the base model. Ties-Merging combines selected task vectors by trimming redundant parameters and resolving sign conflicts before averaging. Selection is driven by two complementary signals: cosine similarity between the centroid embedding of a source domain and the centroid embedding of target raw text, and each expert's F1 score on ten target sentences labeled by majority-vote pseudo-labels from all experts. The two selected expert subsets are merged into two task-specific models, $M_{DS}$ and $M_{SE}$, and the final prediction is the union of their outputs.

What would settle it

Run SaM on a held-out target domain under two conditions: with normal selection using target raw text, and with selection replaced by a random subset of three experts or by merging all experts. If random or full merging matches or beats SaM's F1 on that domain, the paper's key claim that the two selection mechanisms drive the roughly 10% gain would be falsified.

Watch

Extended reading notes

Core claim

The central claim, stated in Section 4.2, is that selecting and merging a small number of LoRA-trained domain experts gives substantially better zero-shot NER on unseen domains than a fully trained unified model. The authors report an average improvement of about 10% in entity-level micro-F1 over the unified-model baseline, with gains up to 20% in specific domains. Zero-shot here means no gold labels for the target domain, but the method does use raw target-domain text for selection: a few hundred sentences for domain-similarity embeddings and ten sampled sentences for pseudo-label evaluation. The paper also argues that this design is scalable, because experts can be added or removed independently, and that a single merged task model can approximate the full two-model setup at the same inference cost as a unified model.

Load-bearing premise

The framework only works if the practitioner has raw, unlabeled text from the target domain: a few hundred sentences for the domain-similarity embedding and ten sampled sentences for the pseudo-label evaluation; without any target text, neither selection strategy is defined.

Editorial extensions

If this is right

  • An organization that already maintains several fine-tuned NER models can combine them for new domains without retraining a single model on pooled data.
  • The best model for a given target domain is not necessarily the one trained on all data; keeping experts separate and merging on demand gives better zero-shot transfer.
  • Domain-similarity selection and sampled pseudo-label selection capture different signals, and their union is stronger than either strategy used alone.
  • The number of experts merged should be treated as a domain-dependent choice, with two to four usually optimal and more experts sometimes harmful.
  • With only one target-domain sentence, the method still works in most tested domains, so it degrades gracefully as target text becomes scarce.

Reading between the lines

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

  • Because the selection step needs raw target-domain text, the zero-shot framing rests on a specific resource assumption; a genuinely cold-start extension would need a fallback selector, such as merging all experts or using entity-type overlap, which the paper only hints at.
  • The paper's own comparison with a redundancy-removal baseline suggests its gains are additive: better-trained experts, produced by improved data cleaning or instruction optimization, should raise the ceiling of SaM further.
  • The multilingual and non-strict-domain experiments point to a general design principle: a diverse set of complementary experts, not necessarily labeled by domain, is enough for the select-and-merge mechanism to pay off.
  • A natural testable extension is to apply the same selection metrics to relation extraction and event extraction, where the framework's logic predicts similar or larger gains in proportion to domain heterogeneity.
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 SaM, a framework for zero-shot named entity recognition with LLMs. Instead of training a single unified model on all source domains, SaM trains per-domain LoRA expert models and, at inference time, selects a subset of experts for a target domain using two complementary signals: (i) cosine similarity between target-domain text embeddings and source-domain centroid embeddings, and (ii) performance on k randomly sampled target-domain instances scored against majority-vote pseudo-labels produced by the experts themselves. The selected experts are merged with Ties-Merging, and two task-specific models (one per selection signal) are used to produce a union of predictions. Experiments on CrossNER and MIT across seven domains with Llama3.1-8B and Qwen2.5-7B report an average improvement of about 10% over a fully-trained unified model, with additional ablations, cost analysis, and extensions to non-strict domains and multilingual settings.

Significance. If the reported result holds, SaM is a valuable modular alternative to unified training for instruction-tuned NER: it decouples training from target-domain adaptation, allows experts to be added or removed, and its ablations on two backbones consistently support the usefulness of selection plus Ties-Merging over a same-configuration unified model. The paper also provides useful practical analyses, including a cost comparison, sensitivity to the number of source domains, an extreme one-sample scenario, alternative merging techniques, and evidence that the framework generalizes beyond strict domains and to multilingual settings. However, the headline quantitative claim is weakened by the choice of the primary baseline and by the fact that a key hyperparameter (the number of merged experts) is selected using the evaluation domains themselves; a properly calibrated comparison to a stronger or tuned unified baseline is needed before the 10% figure can be taken at face value.

major comments (4)
  1. [§4.2, Table 1; §3.1] The central claim of an approximate 10% average improvement is measured only against the "Fully-trained" baseline, which Section 3.1 explicitly describes as non-optimal by design. Against the stronger published unified model B2NER, SaM-Llama improves by only 1.67 points (66.24 vs. 64.57, about 2.6% relative), and SaM-Qwen (63.13) is actually below B2NER (64.57). The abstract and conclusion should either report the comparison against a properly tuned or stronger unified baseline, or consistently qualify the claim as "improvement over a same-configuration, non-optimized unified model." This is load-bearing because the current phrasing overstates the method's advantage over existing unified NER models.
  2. [§4.5, Figure 3; §4.2] The number of merged experts, m = 3, is fixed after inspecting the average F1 curves on the seven target test domains in Figure 3. This means the main results in Table 1 are partly selected on the test set, which is likely to inflate the reported gains. The authors should either fix m by a principled criterion before evaluation, use a held-out validation split for this choice, or report the full sensitivity curves with the selection protocol explicitly stated. Without this, the "approximately 10%" figure cannot be treated as an unbiased estimate of the framework's advantage.
  3. [§3.2, §4.7, Table 5] The zero-shot framing requires clarification: both selection strategies need unlabeled target-domain text (a few hundred sentences for the domain-similarity embedding and k = 10 sampled instances for sampling evaluation). The paper claims "zero-shot" only in the sense of no gold labels, but a practitioner with no target-domain text at all cannot apply either selection strategy. In addition, the sampling-evaluation path constructs pseudo-labels by majority voting over the same experts being ranked, making the selection signal partially self-referential, and no variance or significance estimates are reported for a procedure that depends on a random sample of k = 10 instances. The authors should state the target-text assumption explicitly and provide stability analysis, e.g., repeated sampling or a comparison of pseudo-label-based rankings with gold-label-based rankings on the sampled instances.
  4. [Table 2, §4.2] The main tables report single runs without error bars or significance tests. Given that the differences between some configurations in Table 2 (e.g., Llama w/o Sampling Evaluation 64.41 vs. w/o Selection 64.33) are small, and that the selection process includes random sampling, the claim that the framework "significantly outperforms" baselines needs at least run-to-run variance or a paired test over multiple random seeds.
minor comments (6)
  1. [§3.1] Typo: "Refering" should be "Referring."
  2. [§3.2, §4.5, §4.3] The symbol k is used inconsistently: Section 3.2 and the implementation description define k as the number of sampled target instances and set it to 10, while Section 4.5 uses k for the number of merged experts and Section 4.3 repeats this usage ("we fix the number of selected experts to k = 3"). Use m for the number of merged experts throughout.
  3. [Table 13] The caption reads "Our-of-domain performance" and should be "Out-of-domain performance."
  4. [Figures 3, 4, 9, 10, 11, 12] The captions inconsistently say "denotes as k" for the number of experts or splits; please unify the notation with the text (m for experts, and a distinct symbol for splits).
  5. [Appendix C, Table 14] The storage formula 1+0.02n and the "2x" inference cost for SaM are not derived in the table; a sentence explaining the 0.02 factor and the trade-off between SaM and SaMeco would help reproducibility.
  6. [§4.7, Table 5] The one-sample experiment is averaged over five trials, but the main results are not; stating whether the reported main results are single trials or averaged over multiple seeds would make the reliability of the numbers clearer.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reported F1 gains are measured on held-out gold labels, and the selection mechanisms do not by construction determine those numbers.

full rationale

The paper's central claim is empirical: SaM outperforms a fully trained unified model by about 10% average F1 on CrossNER/MIT target domains. These F1 scores are computed against held-out gold labels that are not used by either selection strategy, so the result is not forced by construction. Domain-similarity selection uses unlabeled target-domain text and an external text encoder, and sampling evaluation selects experts using pseudo-labels formed by majority voting over the same experts; although the pseudo-label scoring is self-referential as a heuristic, it does not by construction determine the final gold-label F1, so it is not a derivation-equivalence circularity. The paper explicitly discloses that its training configuration is not optimal for the fully trained baseline, which is a baseline-strength concern rather than a circular derivation. The choice m=3 is made by inspecting target-domain performance in Section 4.5, which is a test-set hyperparameter selection issue, not a case where a fitted parameter is renamed as a prediction. Self-citations in the related work are contextual and are not load-bearing for the framework's claims. Overall, the derivation chain is self-contained with respect to the reported evaluation.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on three practical assumptions: embedding similarity is informative, the pseudo-label vote is a usable oracle, and Ties-Merging preserves expert skills. The only genuinely fitted free parameter is m; k and the instruction-mix percentages are set by convention.

free parameters (3)
  • m (number of selected experts for merging) = 3
    Set to 3 because it gave the best average F1 across the seven evaluation domains in Figure 3, i.e., tuned on test data.
  • k (number of sampled target instances) = 10
    Fixed without sensitivity analysis; used for pseudo-label evaluation in Section 3.2.
  • Data augmentation mix (70% source/entity descriptions, 30% label drop, 5% label masking) = 70/30/5
    The percentages of training instructions with data source description, entity type drop, and label masking in Section 3.1 are chosen by hand; they affect expert quality but not the selection mechanism.
assumptions (4)
  • domain assumption Centroid of text embeddings is a valid proxy for NER domain similarity
    Equation (2) assumes mean-pooled all-MiniLM-L6-v2 embeddings capture NER-related domain relatedness; used for cosine-similarity selection in Section 3.2.
  • domain assumption Majority-vote pseudo-labels are a reliable oracle for expert selection
    Section 3.2 assumes the ensemble vote of six experts is trustworthy enough to rank expert quality on the target domain, with no gold labels used in selection.
  • domain assumption Task-vector arithmetic preserves expert capabilities after merging
    Equation (3) and Ties-Merging assume delta parameters from LoRA experts can be merged without destructive interference, following Ilharco et al. (2023) and Yadav et al. (2023).
  • domain assumption Unlabeled target-domain text is available at inference
    Domain similarity and sampling evaluation both consume raw target-domain text; the paper's 'zero-shot' label covers only the absence of labels, not the absence of target text.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Selecting and Merging: Towards Adaptable and Scalable Named Entity Recognition with Large Language Models." pith.science (2026). https://pith.science/paper/4Z6IDRN2

@misc{pith2026250622813,
  author       = {Pith},
  title        = {Pith review of: Selecting and Merging: Towards Adaptable and Scalable Named Entity Recognition with Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4Z6IDRN2}},
  note         = {Machine review of arXiv:2506.22813}
}
read the original abstract

Supervised fine-tuning (SFT) is widely used to align large language models (LLMs) with information extraction (IE) tasks, such as named entity recognition (NER). However, annotating such fine-grained labels and training domain-specific models is costly. Existing works typically train a unified model across multiple domains, but such approaches lack adaptation and scalability since not all training data benefits target domains and scaling trained models remains challenging. We propose the SaM framework, which dynamically Selects and Merges expert models at inference time. Specifically, for a target domain, we select domain-specific experts pre-trained on existing domains based on (i) domain similarity to the target domain and (ii) performance on sampled instances, respectively. The experts are then merged to create task-specific models optimized for the target domain. By dynamically merging experts beneficial to target domains, we improve generalization across various domains without extra training. Additionally, experts can be added or removed conveniently, leading to great scalability. Extensive experiments on multiple benchmarks demonstrate our framework's effectiveness, which outperforms the unified model by an average of 10%. We further provide insights into potential improvements, practical experience, and extensions of our framework.

Figures

Figures reproduced from arXiv: 2506.22813 by the authors.

Figure 1
Figure 1. (a) Existing methods train a general unified [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Framework overview. Given a target domain, we select expert models from two perspectives: [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Performance changes with the number of experts. The horizontal axis is the number of experts for merging, and the vertical denotes the F1 scores. and F1-scores on sampled data points) to a common scale and selects the top three experts. (3) Mode3 takes the union of the two expert sets while limit￾ing the total number of selected experts to three. Experimental results show that Mode2 and Mode3 achieve comparable perf… view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Performance changes with the number of data [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Performance changes with the number of data [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Distribution of the training data. in-domain tasks, directly utilizing the correspond￾ing expert model is sufficient. For out-of-domain tasks, merging expert models improves generaliza￾tion. However, experimental results indicate that in some cases, a single expert mod…
Figure 7
Figure 7. Figure 7: Formatted training data example. The example consists of task instructions, inputs, and outputs for [PITH_FULL_IMAGE:figures/full_fig_p016_7.png]
Figure 8
Figure 8. Figure 8: Another example of our formatted training data. The instance here is the same as that of Figure [PITH_FULL_IMAGE:figures/full_fig_p016_8.png]
Figure 9
Figure 9. Figure 9: Performance changes with the number of expert models (denotes as [PITH_FULL_IMAGE:figures/full_fig_p017_9.png]
Figure 10
Figure 10. Figure 10: Performance changes with the number of expert models. The horizontal axis is the number of experts for [PITH_FULL_IMAGE:figures/full_fig_p017_10.png]
Figure 11
Figure 11. Figure 11: Performance changes with the number of data splits. The horizontal axis is the number of splits, and the [PITH_FULL_IMAGE:figures/full_fig_p018_11.png]
Figure 12
Figure 12. Figure 12: Performance changes with the number of data splits. The horizontal axis is the number of splits, and [PITH_FULL_IMAGE:figures/full_fig_p018_12.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

56 extracted references · 44 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]

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774

    arXiv 2023

  4. [4]

    Ting Wai Terence Au, Vasileios Lampos, and Ingemar Cox. 2022. E-ner—an annotated named entity recognition corpus of legal text. In Proceedings of the Natural Legal Language Processing Workshop 2022, pages 246--255

    work page 2022

  5. [5]

    Pei Chen, Haotian Xu, Cheng Zhang, and Ruihong Huang. 2022. Crossroads, buildings and neighborhoods: A dataset for fine-grained location recognition. In Proceedings of the 2022 conference of the North American chapter of the association for computational linguistics: human language technologies, pages 3329--3339

    work page 2022

  6. [6]

    Nigel Collier, Tomoko Ohta, Yoshimasa Tsuruoka, Yuka Tateisi, and Jin-Dong Kim. 2004. Introduction to the bio-entity recognition task at jnlpba. In Proceedings of the International Joint Workshop on Natural Language Processing in Biomedicine and its Applications (NLPBA/BioNLP), pages 73--78

    work page 2004

  7. [7]

    Leon Derczynski, Kalina Bontcheva, and Ian Roberts. 2016. Broad twitter corpus: A diverse named entity recognition resource. In Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 1169--1179

    work page 2016

  8. [8]

    Leon Derczynski, Eric Nichols, Marieke Van Erp, and Nut Limsopatham. 2017. Results of the wnut2017 shared task on novel and emerging entity recognition. In Proceedings of the 3rd Workshop on Noisy User-generated Text, pages 140--147

    work page 2017

Show all 56 references
  1. [9]

    Yuyang Ding, Juntao Li, Pinzheng Wang, Zecheng Tang, Yan Bowen, and Min Zhang. 2024 a . Rethinking negative instances for generative named entity recognition. In Findings of the Association for Computational Linguistics: ACL 2024, pages 3461--3475

    2024

  2. [10]

    Zhuojun Ding, Wei Wei, Xiaoye Qu, and Dangyang Chen. 2024 b . Improving pseudo labels with global-local denoising framework for cross-lingual named entity recognition. In Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, pages 6252--6260

    2024

  3. [11]

    Rezarta Islamaj Do g an, Robert Leaman, and Zhiyong Lu. 2014. Ncbi disease corpus: a resource for disease name recognition and concept normalization. Journal of biomedical informatics, 47:1--10

    2014

  4. [12]

    Chenghao Fan, Zhenyi Lu, Sichen Liu, Xiaoye Qu, Wei Wei, Chengfeng Gu, and Yu Cheng. 2025. Make lora great again: Boosting lora with adaptive singular values and mixture-of-experts optimization alignment

    2025

  5. [13]

    Chenghao Fan, Zhenyi Lu, Wei Wei, Jie Tian, Xiaoye Qu, Dangyang Chen, and Yu Cheng. 2024 a . On giant's shoulders: Effortless weak to strong by dynamic logits fusion. In The Thirty-eighth Annual Conference on Neural Information Processing Systems

    2024

  6. [14]

    Chenghao Fan, Wei Wei, Xiaoye Qu, Zhenyi Lu, Wenfeng Xie, Yu Cheng, and Dangyang Chen. 2024 b . Enhancing low-resource relation representations through multi-view decoupling. Proceedings of the AAAI Conference on Artificial Intelligence, 38:17968--17976

    2024

  7. [15]

    Runwei Guan, Ka Lok Man, Feifan Chen, Shanliang Yao, Rongsheng Hu, Xiaohui Zhu, Jeremy Smith, Eng Gee Lim, and Yutao Yue. 2024. Findvehicle and vehiclefinder: a ner dataset for natural language-based vehicle retrieval and a keyword-based cross-modal vehicle retrieval system. M...

    2024

  8. [16]

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. 2025 a . Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948

    2025 arXiv

  9. [17]

    Quanjiang Guo, Yihong Dong, Ling Tian, Zhao Kang, Yu Zhang, and Sijie Wang. 2025 b . Baner: Boundary-aware llms for few-shot named entity recognition. In Proceedings of the 31st International Conference on Computational Linguistics, pages 10375--10389

    2025

  10. [18]

    Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. 2022. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations

    2022

  11. [19]

    Gabriel Ilharco, Marco Tulio Ribeiro, Mitchell Wortsman, Ludwig Schmidt, Hannaneh Hajishirzi, and Ali Farhadi. 2023. Editing models with task arithmetic. In The Eleventh International Conference on Learning Representations

    2023

  12. [20]

    Xisen Jin, Xiang Ren, Daniel Preotiuc-Pietro, and Pengxiang Cheng. 2023. Dataless knowledge fusion by merging weights of language models. In The Eleventh International Conference on Learning Representations

    2023

  13. [21]

    Martin Krallinger, Obdulia Rabal, Florian Leitner, Miguel Vazquez, David Salgado, Zhiyong Lu, Robert Leaman, Yanan Lu, Donghong Ji, Daniel M Lowe, et al. 2015. The chemdner corpus of chemicals and drugs and its annotation principles. Journal of cheminformatics, 7:1--17

    2015

  14. [22]

    Aman Kumar and Binil Starly. 2022. “fabner”: information extraction from manufacturing process science domain literature using named entity recognition. Journal of Intelligent Manufacturing, 33(8):2393--2407

    2022

  15. [23]

    Jiao Li, Yueping Sun, Robin J Johnson, Daniela Sciaky, Chih-Hsuan Wei, Robert Leaman, Allan Peter Davis, Carolyn J Mattingly, Thomas C Wiegers, and Zhiyong Lu. 2016. Biocreative v cdr task corpus: a resource for chemical disease relation extraction. Database, 2016

    2016

  16. [24]

    Peng Li, Tianxiang Sun, Qiong Tang, Hang Yan, Yuanbin Wu, Xuan-Jing Huang, and Xipeng Qiu. 2023. Codeie: Large code generation models are better few-shot information extractors. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume ...

    2023

  17. [25]

    Zixuan Li, Yutao Zeng, Yuxin Zuo, Weicheng Ren, Wenxuan Liu, Miao Su, Yucan Guo, Yantao Liu, Lixiang Lixiang, Zhilei Hu, Long Bai, Wei Li, Yidan Liu, Pan Yang, Xiaolong Jin, Jiafeng Guo, and Xueqi Cheng. 2024. Know C oder: Coding structured knowledge into llms for universal in...

    2024

  18. [26]

    Wei Liu, Weihao Zeng, Keqing He, Yong Jiang, and Junxian He. 2024. What makes good data for alignment? a comprehensive study of automatic data selection in instruction tuning. In The Twelfth International Conference on Learning Representations

    2024

  19. [27]

    Zihan Liu, Yan Xu, Tiezheng Yu, Wenliang Dai, Ziwei Ji, Samuel Cahyawijaya, Andrea Madotto, and Pascale Fung. 2021. Crossner: Evaluating cross-domain named entity recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 13452--13460

    2021

  20. [28]

    Zhenyi Lu, Chenghao Fan, Wei Wei, Xiaoye Qu, Dangyang Chen, and Yu Cheng. 2024. Twin-merging: Dynamic integration of modular expertise in model merging. In The Thirty-eighth Annual Conference on Neural Information Processing Systems

    2024

  21. [29]

    Michael S Matena and Colin Raffel. 2022. Merging models with fisher-weighted averaging. In Advances in Neural Information Processing Systems

    2022

  22. [30]

    Alexis Mitchell, Stephanie Strassel, Shudong Huang, and Ramez Zakhary. 2005. Ace 2004 multilingual training corpus. LDC corpora

    2005

  23. [31]

    Xiaoman Pan, Boliang Zhang, Jonathan May, Joel Nothman, Kevin Knight, and Heng Ji. 2017. Cross-lingual name tagging and linking for 282 languages. In Proceedings of the 55th annual meeting of the association for computational linguistics (volume 1: long papers), pages 1946--1958

    2017

  24. [32]

    Chaoxu Pang, Yixuan Cao, Qiang Ding, and Ping Luo. 2023. Guideline learning for in-context information extraction. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 15372--15389

    2023

  25. [33]

    Sameer Pradhan, Alessandro Moschitti, Nianwen Xue, Hwee Tou Ng, Anders Bj \"o rkelund, Olga Uryupina, Yuchen Zhang, and Zhi Zhong. 2013. Towards robust linguistic analysis using ontonotes. In Proceedings of the Seventeenth Conference on Computational Natural Language Learning,...

    2013

  26. [34]

    Sampo Pyysalo and Sophia Ananiadou. 2014. Anatomical entity mention recognition at literature scale. Bioinformatics, 30(6):868--875

    2014

  27. [35]

    Yunjia Qi, Hao Peng, Xiaozhi Wang, Bin Xu, Lei Hou, and Juanzi Li. 2024. Adelie: Aligning large language models on information extraction. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 7371--7387

    2024

  28. [36]

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. 2024. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36

    2024

  29. [37]

    Oscar Sainz, Iker Garc \' a-Ferrero, Rodrigo Agerri, Oier Lopez de Lacalle, German Rigau, and Eneko Agirre. 2024. Gollie: Annotation guidelines improve zero-shot information-extraction. In The Twelfth International Conference on Learning Representations

    2024

  30. [38]

    Larry Smith, Lorraine K Tanabe, Rie Johnson nee Ando, Cheng-Ju Kuo, I-Fang Chung, Chun-Nan Hsu, Yu-Shi Lin, Roman Klinger, Christoph M Friedrich, Kuzman Ganchev, et al. 2008. Overview of biocreative ii gene mention recognition. Genome biology, 9:1--19

    2008

  31. [39]

    Zeliang Tong, Zhuojun Ding, and Wei Wei. 2025. Evoprompt: Evolving prompts for enhanced zero-shot named entity recognition with large language models. In Proceedings of the 31st International Conference on Computational Linguistics, pages 5136--5153

    2025

  32. [40]

    Asahi Ushio, Francesco Barbieri, Vitor Sousa, Leonardo Neves, and Jose Camacho-Collados. 2022. Named entity recognition in twitter: A dataset and analysis on short-term temporal shifts. In Proceedings of the 2nd Conference of the Asia-Pacific Chapter of the Association for Com...

    2022

  33. [41]

    Asahi Ushio and Jose Camacho-Collados. 2021. T-ner: An all-round python library for transformer-based named entity recognition. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: System Demonstrations, pages 53--62

    2021

  34. [42]

    Christopher Walker, Stephanie Strassel, Julie Medero, and Kazuaki Maeda. 2006. Ace 2005 multilingual training corpus. LDC corpora

    2006

  35. [43]

    Zhen Wan, Fei Cheng, Zhuoyuan Mao, Qianying Liu, Haiyue Song, Jiwei Li, and Sadao Kurohashi. 2023. Gpt-re: In-context learning for relation extraction using large language models. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages ...

    2023

  36. [44]

    Xiao Wang, Weikang Zhou, Can Zu, Han Xia, Tianze Chen, Yuansen Zhang, Rui Zheng, Junjie Ye, Qi Zhang, Tao Gui, et al. 2023 a . Instructuie: Multi-task instruction tuning for unified information extraction. arXiv preprint arXiv:2304.08085

    2023 arXiv

  37. [45]

    Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc V Le, Ed H Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. 2023 b . Self-consistency improves chain of thought reasoning in language models. In The Eleventh International Conference on Learning Representations

    2023

  38. [46]

    Zihan Wang, Jingbo Shang, Liyuan Liu, Lihao Lu, Jiacheng Liu, and Jiawei Han. 2019. Crossweigh: Training named entity tagger from imperfect annotations. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Co...

    2019

  39. [47]

    Tingyu Xie, Qi Li, Jian Zhang, Yan Zhang, Zuozhu Liu, and Hongwei Wang. 2023. Empirical study of zero-shot ner with chatgpt. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 7935--7956

    2023

  40. [48]

    Tingyu Xie, Qi Li, Yan Zhang, Zuozhu Liu, and Hongwei Wang. 2024. Self-improving for zero-shot named entity recognition with large language models. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Langu...

    2024

  41. [49]

    Derong Xu, Wei Chen, Wenjun Peng, Chao Zhang, Tong Xu, Xiangyu Zhao, Xian Wu, Yefeng Zheng, Yang Wang, and Enhong Chen. 2024. Large language models for generative information extraction: A survey. Frontiers of Computer Science, 18(6):186357

    2024

  42. [50]

    Prateek Yadav, Derek Tam, Leshem Choshen, Colin Raffel, and Mohit Bansal. 2023. Ties-merging: Resolving interference when merging models. In Annual Conference on Neural Information Processing Systems

    2023

  43. [51]

    An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. 2024. Qwen2. 5 technical report. arXiv preprint arXiv:2412.15115

    2024 arXiv

  44. [52]

    Yuming Yang, Wantong Zhao, Caishuang Huang, Junjie Ye, Xiao Wang, Huiyuan Zheng, Yang Nan, Yuran Wang, Xueying Xu, Kaixin Huang, Yunke Zhang, Tao Gui, Qi Zhang, and Xuanjing Huang. 2025. Beyond boundaries: Learning a universal entity taxonomy across datasets and languages for ...

    2025

  45. [53]

    Le Yu, Bowen Yu, Haiyang Yu, Fei Huang, and Yongbin Li. 2024. Language models are super mario: Absorbing abilities from homologous models as a free lunch. In Forty-first International Conference on Machine Learning

    2024

  46. [54]

    Urchade Zaratiana, Nadi Tomeh, Pierre Holat, and Thierry Charnois. 2024. Gliner: Generalist model for named entity recognition using bidirectional transformer. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics:...

    2024

  47. [55]

    Chunting Zhou, Pengfei Liu, Puxin Xu, Srinivasan Iyer, Jiao Sun, Yuning Mao, Xuezhe Ma, Avia Efrat, Ping Yu, Lili Yu, et al. 2024 a . Lima: Less is more for alignment. Advances in Neural Information Processing Systems, 36

    2024

  48. [56]

    Wenxuan Zhou, Sheng Zhang, Yu Gu, Muhao Chen, and Hoifung Poon. 2024 b . Universalner: Targeted distillation from large language models for open named entity recognition. In The Twelfth International Conference on Learning Representations

    2024

Pith tools

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