Pith. sign in

REVIEW 2 major objections 6 minor 43 references

Construction-Driven Injection: Linguistically-Grounded Edit-Based Code-Mixing Fingerprints for Large Language Models

T0 review · 2 major / 6 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read This paper argues that the linguistic construction of a code-mixing fingerprint should drive its injection, and reports that this coupling makes ownership signals both harder to spot and harder to erase.

desk verdict A solid incremental fingerprinting paper; the construction–injection coupling is real but the Digital Island transfer to FFN edits is unmeasured, and the tables have artifacts. read the letter →

arxiv 2607.25633 v1 pith:UDUOZM6Q submitted 2026-07-28 cs.CL cs.AI

Yongyi Cui , Yue Li , Tianbao Jiang , Xin Yi This is my paper
classification cs.CLcs.AI
keywords LLMfingerprintingcode-mixingtriggersknowledgeeditingnull-spaceprojectioncross-lingualalignmentDigitalIslandhypothesismodelownershipverificationperplexity-baseddetection
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 wants to establish that LLM fingerprint construction and injection should be one coupled process, not two independent stages. It introduces code-mixing trigger–target pairs built from low-resource languages, deliberately choosing substitutes and grammar patterns that keep the trigger statistically close to natural text. It then injects each pair with an editing update that preserves mainstream knowledge while steering the edit toward the fingerprint language's own representation subspace. If correct, ownership of a released model can be verified black-box even after the model has been fine-tuned, quantized, or pruned, with the fingerprint barely affecting ordinary behavior. The paper reports exactly this: near-complete injection success, better post-attack retention than existing edit-based methods, and near-zero impact on model utility.

What carries the argument

The central mechanism is the coupling of two projectors: a null-space projector that forces edits to leave mainstream English/Chinese knowledge untouched, and a cross-lingual alignment projector built from the fingerprint language's Wikipedia keys. A post-hoc convex step reweights the solved update toward the alignment subspace, so the edit's relative weight concentrates where the Digital Island hypothesis says it is least likely to be overwritten. The named object is the Digital Island hypothesis: parameters associated with low-resource languages receive much smaller gradients during mainstream fine-tuning, making them a natural refuge for persistent edits.

What would settle it

Measure the per-row L2 norm of the fine-tuning gradient on the actual edited FFN key-value weights, for LCF tokens versus natural-language tokens. If the FFN rows associated with LCF tokens do not receive markedly smaller gradients than English tokens (or than the vocabulary mean), the Digital Island mechanism is not what produces the retention gains, and the alignment projector is ordinary regularization.

Watch

Extended reading notes

Core claim

The load-bearing idea is the Digital Island hypothesis: low-resource-language tokens live in a quiet region of representation space that mainstream fine-tuning rarely touches. The paper constructs LCF fingerprints to exploit that region, then injects them via LCFEdit, a two-step update: a null-space projection preserves English/Chinese knowledge, and a post-hoc cross-lingual alignment step concentrates the weight update in the target language's subspace, estimated from that language's Wikipedia key statistics. The claim is that this construction-awareness is what makes the fingerprint both imperceptible and persistent: mean template perplexity of 96, below natural-language and garbled baseli

Load-bearing premise

The entire construction-driven advantage rests on the hypothesis that low-resource-language parameters really are quiet during mainstream fine-tuning, and that this quietness transfers from input-embedding gradients to the deep feed-forward key-value weights where the fingerprint is actually written; the paper states this as a hypothesis and tests only the embedding side.

Editorial extensions

If this is right

  • Ownership verification can remain black-box after realistic post-release modifications, because the fingerprint survives quantization and domain fine-tuning.
  • Perplexity-based filters and language-identifier filters both fail against these triggers, since LCF queries look statistically like natural text and mostly monolingual.
  • Smaller and more weakly multilingual models benefit most from the alignment step, so the method is useful precisely where naive fingerprinting fails.
  • The construction-stage language identity can serve as a key for later audit: verification can test the intended language subspace directly.
  • Heavy pruning at 40% sparsity remains a hard limit for all edit-based fingerprints, as the paper itself reports.

Reading between the lines

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

  • If the Digital Island transfer from embedding rows to deep FFN edit weights is real, the same construction-to-injection coupling could make low-resource languages a general refuge for other persistent weight edits, such as safety alignment or factual updates.
  • The reported gains may not come specifically from language identity; a randomly chosen subspace of the same dimensionality might produce similar retention, so comparing the true alignment projector against a random one would separate the mechanism from generic regularization.
  • The off-diagonal transfer among typologically related languages hints that fingerprints could be made adversarially robust by selecting language pairs that share subword structure.
  • The method depends on clean Wikipedia key statistics for the target language, so extremely low-resource languages may need a synthetic-key variant to benefit.
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

2 major / 6 minor

Summary. The paper proposes a two-stage fingerprinting framework for LLMs. In the construction stage, LCF builds code-mixed trigger-target pairs from low-resource languages using a semantic-density substitution rule and grammar-biased mixing. In the injection stage, LCFEdit writes the pairs into FFN key-value memories with an AlphaEdit-style null-space projection, followed by a cross-lingual alignment step that reweights the update toward a subspace estimated from the fingerprint language's Wikipedia keys. The central claim is that coupling construction to injection—using the construction-stage language identity to choose where the edit is concentrated—improves injection success and post-attack retention while keeping triggers stealthy and harmless. Experiments on Qwen3.5-0.8B/4B/9B and Llama-3.2-3B report injection FSR of 93.0–99.5%, mean retention gains of roughly 9–34 points over AlphaEdit across six attacks, low template perplexity (mean 96), and negligible utility loss.

Significance. If the mechanism holds, the paper makes a useful contribution: it demonstrates that linguistic structure of a fingerprint can inform where to write a model edit, and it provides a practical combination of stealth and robustness. The paper has notable strengths: per-language results in Tables 6, 7, and 12 are unusually detailed; the Digital Island hypothesis is directly tested in Table 4; the ablation in Fig. 4 isolates the alignment step; and harmlessness is evaluated on MMLU, RTE, MMMLU, and WikiText2. The 12x12 alignment transfer matrix in Appendix C.3 is a good falsifiable check that the alignment is language-specific. However, two load-bearing issues need to be addressed: Eq. (5) breaks the null-space preservation guarantee of Eq. (3), and the Digital Island mechanism is tested only on input-embedding rows while the edit is written to FFN key-value memories. These issues do not invalidate the empirical results, but they are central to the 'construction-driven' contribution and should be fixed with direct measurements or carefully revised claims.

major comments (2)
  1. [§3.3, Eq. (5) vs. Eq. (3)] Eq. (5) breaks the null-space preservation guaranteed by Eq. (3). In Eq. (4), Δ is the null-space-projected update, so ΔK0=0. After Eq. (5), the actually applied update Δ'=(1−α)Δ+αΔP_isl satisfies Δ'K0 = αΔP_islK0, which is not generally zero. Thus the statement that 'edits do not disturb preserved English/Chinese associations' holds for the pre-alignment Δ but not for the final Δ'. The harmlessness results in Table 8 are empirical and reassuring, but the proof-level guarantee no longer applies. Please quantify the violation (e.g., ||Δ'K0||/||ΔK0||) and show it is negligible, or revise the claim so that it applies only before the alignment step.
  2. [§6.2, Table 4 and §3.3] The Digital Island hypothesis is tested only on input-embedding rows: Table 4 measures gradient norms on embedding rows of LCF tokens, while LCFEdit writes to FFN key-value memories via Eq. (4)–(5), and P_isl is built from Wikipedia key statistics. The paper does not measure (a) fine-tuning gradients with respect to the edited FFN weights projected onto P_isl versus its complement, nor (b) whether the Wikipedia-key subspace actually aligns with the keys of the LCF triggers. Without such evidence, the gains from the alignment step in Fig. 4 could be explained by generic shrinkage/regularization rather than by the hypothesized linguistic grounding. Please add direct FFN-level measurements or control experiments (e.g., a random projector of the same rank and update norm) to support the causal claim, or soften the claim accordingly.
minor comments (6)
  1. [Appendix A.1, Table 5] The number m of principal directions used to build P_isl is not reported. Since P_isl is the central new component, please report m and the Wikipedia key sampling details for reproducibility.
  2. [§5.1] The sentence 'Within the Qwen family the injection gap over AlphaEdit widens as models shrink (ratio 0.70→0.64→0.61)' does not define the ratio being computed. Please specify the numerator and denominator.
  3. [§5.2] The terms 'constructed query' and 'culturally grounded question' are central to the accidental-activation and language-identifier results, but they are defined only in Appendix C.2/Table 11. Define them in the main text.
  4. [Table 3] Please specify the exact fastText lid.176 and CLD3 versions/models and the flag rule used, since the flag rates are described as preliminary estimates.
  5. [Fig. 3] The caption should state that PPL is WikiText2 perplexity; currently this appears only in the body text.
  6. [Throughout] There are spacing/extraction artifacts in the arXiv text (e.g., 'Modelfingerprintinghasemergedastheleadingresponse'). These should be corrected in the camera-ready version.

Circularity Check

0 steps flagged · score 2.0 of 10

No material circularity: the construction–injection chain is empirically tested and its components are not fitted to the claimed outcomes.

full rationale

The paper's derivation chain is not circular in the load-bearing sense. LCF triggers are built by a semantic-density substitution rule and grammar-biased mixing (Eq. 2), independent of the injection-side alignment projector P_isl, which is estimated from target-language Wikipedia key statistics. The Digital Island hypothesis (Eq. 1) is explicitly stated as a hypothesis and tested directly in Table 4, with the paper itself disclaiming proof ('We state Eq. (1) as a hypothesis... We do not claim to prove it'). The injection update is a closed-form null-space-constrained least-squares solution (Eq. 4) followed by a post-hoc convex reweighting (Eq. 5); α is a fixed hyperparameter, not a fit to the attack-retention results. The retention and detectability claims are measured against independently re-run AlphaEdit and LoRA baselines and external benchmarks (MMLU, RTE, MMMLU, WikiText2), with the ablation in Fig. 4 isolating the effect of P_isl. The main weakness is a transfer gap: the Digital Island test measures input-embedding gradients rather than FFN key-value subspaces, so the quiet-region assumption is unverified for the edited weights. That is an empirical limitation, not a definitional reduction. The self-citations to the authors' prior work (Li et al. 2025b) supply the CF/MCEdit baselines and the 60% NLF accidental-activation statistic, but the central results do not reduce to those citations, so they are not load-bearing.

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

The framework adds two moving parts beyond the AlphaEdit/FPEdit/MCEdit lineage: a construction heuristic (semantic-density + grammar-biased mixing) and a post-hoc alignment projector. The free parameters are mostly small per-model choices; the main unaccounted degree of freedom is the P_isl dimensionality m. The axioms are the standard knowledge-editing background plus the paper-specific Digital Island hypothesis and its transfer from embeddings to FFN subspaces.

free parameters (4)
  • α (post-hoc alignment weight) = 0.25 for Qwen models; 0.15 for Llama-3.2-3B-Instruct
    Eq. (5) convex reweighting; tuned per model, not derived. Controls how much of the update is rotated into the target-language subspace.
  • m (top principal directions for P_isl) = not reported
    The alignment projector uses 'top-m' Wikipedia-key principal directions (Sec 3.3); m is never given, so the subspace size is a hidden degree of freedom.
  • λ (update regularization) = 10 (Qwen-9B), 3 (Qwen-4B), 3 (Qwen-0.8B), 5 (Llama-3B)
    L2 regularization in Eq. (4) per Table 5; inherited from AlphaEdit per-model configs.
  • edit layers = [6,10,13,14,18] (9B), [3,4,6,10] (4B), [2,3,7,10,22] (0.8B), [4,5,6,7] (3B)
    chosen by causal tracing (Table 5, 10), a modeling choice that affects where the edit lands.
assumptions (4)
  • domain assumption AlphaEdit null-space projection (with P_LCF from K0 eigendecomposition) preserves preserved English/Chinese associations (W+Δ)K0 = WK0, and this remains useful after post-hoc alignment.
    Relied on in Eq. (3)-(4) and in the harmlessness claims of Sec 5.3; the final update Eq. (5) is not exactly in the null space, but the paper still treats preservation as inherited.
  • ad hoc to paper Digital Island hypothesis: E<g_ft, w_LCF> << E<g_ft, w_NLF> (Eq. 1)
    Stated as a hypothesis motivating construction; directly tested on embedding rows only in Sec 6.2, not on the FFN key spaces where LCFEdit writes.
  • ad hoc to paper The Wikipedia-key subspace P_isl of the fingerprint language is the right target for concentrating the edit and is itself a low-gradient region under mainstream modifications.
    Used in Eq. (5); no direct measurement links Wikipedia key principal directions to the gradient dynamics of fine-tuning/quantization/pruning.
  • domain assumption Code-mixing triggers from the construction rule are rare enough in benign queries to avoid accidental activation.
    Basis of the zero accidental-activation claim (Sec 5.2); tested only on the authors' constructed queries and raw NLF sources, not on a broad naturalistic corpus.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Construction-Driven Injection: Linguistically-Grounded Edit-Based Code-Mixing Fingerprints for Large Language Models." pith.science (2026). https://pith.science/paper/UDUOZM6Q

@misc{pith2026260725633,
  author       = {Pith},
  title        = {Pith review of: Construction-Driven Injection: Linguistically-Grounded Edit-Based Code-Mixing Fingerprints for Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UDUOZM6Q}},
  note         = {Machine review of arXiv:2607.25633}
}
read the original abstract

Large language models (LLMs) are costly intellectual assets that remain exposed to unauthorized redistribution and commercial misuse. Injected fingerprints, i.e., trigger--target pairs embedded in model behavior, offer a practical, black-box-verifiable ownership signal, but existing methods decouple the two stages of the fingerprint life cycle: how a fingerprint is constructed and how it is injected. Existing fingerprinting frameworks suffer from two limitations. Natural-language fingerprints are prone to accidental activation, and garbled fingerprints are easily filtered by perplexity-based detection. Furthermore, decoupling construction from injection leaves the latter unaware of the trigger's linguistic structure, missing the opportunity for targeted optimization. We argue that fingerprint construction should drive injection, and present a unified fingerprinting framework that jointly optimizes both stages. First, LCF constructs code-mixing fingerprints by combining low-resource languages under a semantic-density substitution rule and grammar-biased mixing, yielding triggers whose perplexity sits far below garbled baselines while avoiding the accidental-activation failures of natural-language triggers. Second, LCFEdit injects each fingerprint with a null-space projection derived from high-resource multilingual representations that preserves knowledge, augmented by a cross-lingual alignment step that steers the weight update toward the fingerprint language's representation subspace. This construction-aware injection ensures that the update is linguistically informed and therefore more stable. Extensive evaluations on imperceptibility, detectability, and harmlessness demonstrate persistent ownership verification with negligible impact on utility.

Figures

Figures reproduced from arXiv: 2607.25633 by the authors.

Figure 1
Figure 1. Our framework couples the two stages of the fingerprint life cycle. (A) Construction: LCF converts a medium/low [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Template perplexity distributions of different fin [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Zero-shot QA accuracy and perplexity of 6 fin [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Cross-lingual alignment ablation on Qwen3.5- [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

43 extracted references · 5 linked inside Pith

  1. [1]

    Fang, Junfeng and Jiang, Houcheng and Wang, Kun and Ma, Yunshan and Shi, Jie and Wang, Xiang and He, Xiangnan and Chua, Tat-Seng , booktitle =

  2. [2]

    arXiv preprint arXiv:2510.16367 , year=

    EditMark: Watermarking Large Language Models based on Model Editing , author=. arXiv preprint arXiv:2510.16367 , year=

    arXiv

  3. [3]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations) , pages=

    Easyedit: An easy-to-use knowledge editing framework for large language models , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations) , pages=

  4. [4]

    International Conference on Learning Representations , volume=

    Reef: Representation encoding fingerprints for large language models , author=. International Conference on Learning Representations , volume=

  5. [5]

    Proceedings of the 1stWorkshop on GenAI Content Detection (GenAIDetect) , pages=

    Your large language models are leaving fingerprints , author=. Proceedings of the 1stWorkshop on GenAI Content Detection (GenAIDetect) , pages=

  6. [6]

    Proceedings of the The First Workshop on LLM Security (LLMSEC) , pages=

    UTF: Under-trained Tokens as Fingerprints——A Novel Approach to LLM Identification , author=. Proceedings of the The First Workshop on LLM Security (LLMSEC) , pages=

  7. [7]

    Lyu, Sicheng and Gu, Yu and Wang, Xinyu and Huang, Jerry and Luan, Sitao and Cui, Yufei and Chang, Xiao-Wen and Lu, Peng , journal =

  8. [8]

    arXiv preprint arXiv:2009.03300 , year=

    Measuring massive multitask language understanding , author=. arXiv preprint arXiv:2009.03300 , year=

    arXiv 2009

Show all 43 references
  1. [9]

    arXiv preprint arXiv:1609.07843 , year=

    Pointer sentinel mixture models , author=. arXiv preprint arXiv:1609.07843 , year=

    arXiv

  2. [10]

    Proceedings of the 2018 EMNLP workshop BlackboxNLP: Analyzing and interpreting neural networks for NLP , pages=

    GLUE: A multi-task benchmark and analysis platform for natural language understanding , author=. Proceedings of the 2018 EMNLP workshop BlackboxNLP: Analyzing and interpreting neural networks for NLP , pages=

    2018

  3. [11]

    Advances in neural information processing systems , volume=

    Qlora: Efficient finetuning of quantized llms , author=. Advances in neural information processing systems , volume=

  4. [12]

    Transactions of the Association for Computational Linguistics , volume=

    A survey on model compression for large language models , author=. Transactions of the Association for Computational Linguistics , volume=. 2024 , publisher=

    2024

  5. [13]

    International Conference on Learning Representations , volume=

    Mammoth: Building math generalist models through hybrid instruction tuning , author=. International Conference on Learning Representations , volume=

  6. [14]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Lifealign: Lifelong alignment for large language models with memory-augmented focalized preference optimization , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  7. [15]

    Advances in neural information processing systems , volume=

    Learning both weights and connections for efficient neural network , author=. Advances in neural information processing systems , volume=

  8. [16]

    Findings of the Association for Computational Linguistics: ACL 2025 , address =

    Hierarchical Safety Realignment: Lightweight Restoration of Safety in Pruned Large Vision-Language Models , author =. Findings of the Association for Computational Linguistics: ACL 2025 , address =. 2025 , doi =

    2025

  9. [17]

    Wang, Shida and Liu, Chaohu and Wang, Yubo and Xu, Linli , journal =

  10. [18]

    Expert Systems with Applications , volume =

    Latent-space adversarial training with post-aware calibration for defending large language models against jailbreak attacks , author =. Expert Systems with Applications , volume =. 2026 , doi =

    2026

  11. [19]

    arXiv preprint arXiv:2509.03122 , year =

    From Construction to Injection: Edit-Based Fingerprints for Large Language Models , author =. arXiv preprint arXiv:2509.03122 , year =

    arXiv

  12. [20]

    Yue, Xubin and Xu, Zhenhua and Xing, Wenpeng and Yu, Jiahui and Li, Mohan and Han, Meng , journal =

  13. [21]

    Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP) , pages =

    Cross-Lingual Consistency of Factual Knowledge in Multilingual Language Models , author =. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP) , pages =

    2023

  14. [22]

    arXiv preprint arXiv:2504.17480 , year =

    Unified Attacks to Large Language Model Watermarks: Spoofing and Scrubbing in Unauthorized Knowledge Distillation , author =. arXiv preprint arXiv:2504.17480 , year =

    arXiv

  15. [23]

    Locating and Editing Factual Associations in

    Meng, Kevin and Bau, David and Andonian, Alex and Belinkov, Yonatan , booktitle =. Locating and Editing Factual Associations in

  16. [24]

    Li, Yue and Yi, Xin and Shi, Dongsheng and Cui, Yongyi and de Melo, Gerard and Wang, Linlin , journal =

  17. [25]

    The Eleventh International Conference on Learning Representations (ICLR) , year =

    Mass-Editing Memory in a Transformer , author =. The Eleventh International Conference on Learning Representations (ICLR) , year =

  18. [26]

    Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing (EMNLP) , pages =

    Transformer Feed-Forward Layers Are Key-Value Memories , author =. Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing (EMNLP) , pages =

    2021

  19. [27]

    Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics (NAACL) , year =

    Instructional Fingerprinting of Large Language Models , author =. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics (NAACL) , year =

    2024

  20. [28]

    Hey, That's My Model! Introducing Chain & Hash, an

    Russinovich, Mark and Salem, Ahmed , journal =. Hey, That's My Model! Introducing Chain & Hash, an

  21. [29]

    arXiv preprint arXiv:2511.14423 , year =

    Unified Defense for Large Language Models against Jailbreak and Fine-Tuning Attacks in Education , author =. arXiv preprint arXiv:2511.14423 , year =

  22. [30]

    and Shen, Yelong and Wallis, Phillip and Allen-Zhu, Zeyuan and Li, Yuanzhi and Wang, Shean and Wang, Lu and Chen, Weizhu , booktitle =

    Hu, Edward J. and Shen, Yelong and Wallis, Phillip and Allen-Zhu, Zeyuan and Li, Yuanzhi and Wang, Shean and Wang, Lu and Chen, Weizhu , booktitle =

  23. [31]

    Yamabe, Shojiro and Takahashi, Tsubasa and Waseda, Futa and Wataoka, Koki , journal =

  24. [32]

    Yadav, Prateek and Tam, Derek and Choshen, Leshem and Raffel, Colin and Bansal, Mohit , journal =

  25. [33]

    Proceedings of the International Conference on Machine Learning (ICML) , year =

    A Watermark for Large Language Models , author =. Proceedings of the International Conference on Machine Learning (ICML) , year =

  26. [34]

    Touvron, Hugo and Lavril, Thibaut and Izacard, Gautier and others , journal =

  27. [35]

    arXiv preprint arXiv:2407.10671 , year =

    Qwen2 Technical Report , author =. arXiv preprint arXiv:2407.10671 , year =

    arXiv

  28. [36]

    Dettmers, Tim and Lewis, Mike and Belkada, Younes and Zettlemoyer, Luke , booktitle =

  29. [37]

    Frantar, Elias and Alistarh, Dan , booktitle =

  30. [38]

    , year =

    Taori, Rohan and Gulrajani, Ishaan and Zhang, Tianyi and Dubois, Yann and Li, Xuechen and Guestrin, Carlos and Liang, Percy and Hashimoto, Tatsunori B. , year =. Stanford Alpaca: An Instruction-Following

  31. [39]

    Proceedings of the 11th International Conference on Natural Language Processing (ICON) , year =

    Identifying Languages at the Word Level in Code-Mixed Indian Social Media Text , author =. Proceedings of the 11th International Conference on Natural Language Processing (ICON) , year =

  32. [40]

    Nature Human Behaviour , volume =

    Cultural Influences on Word Meanings Revealed through Large-Scale Semantic Alignment , author =. Nature Human Behaviour , volume =

  33. [41]

    Dellert, Johannes and Daneyko, Thora and Münch, Alla and others , journal =

  34. [42]

    Cai, Jiacheng and Yu, Jiahao and Shao, Yangguang and Wu, Yuhang and Xing, Xinyu , journal =

  35. [43]

    Wang, Yiming and Zhang, Zhuosheng and Li, Zuchao and Zhao, Hai , journal =

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.