Pith. sign in

REVIEW 4 major objections 5 minor 47 references

Detoxification of Large Language Models through Output-layer Fusion with a Calibration Model

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper claims that a compact calibration model, trained once on safe text, can detoxify multiple LLaMA-2-7B variants by blending its aligned hidden states into their final output layer, cutting toxicity on challenge prompts while…

desk verdict A genuinely new lightweight detoxification recipe with an honest external split, but the evaluation is too thin to support the fluency claim. read the letter →

arxiv 2506.01266 v1 pith:6KNMSIJZ submitted 2025-06-02 cs.CL cs.AI

Yuanhe Tian , Mingjie Deng , Guoqing Jin , Yan Song This is my paper
classification cs.CLcs.AI
keywords LLMdetoxificationcalibrationmodeloutput-layerfusionembeddingalignmentnegativesamplingRealToxicityPromptsperplexityLLaMA-2
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to show that a small, once-trained calibration model can reduce toxic output from a large language model by being blended into the target model's final layer at generation time, without retraining the large model or hand-crafting prompts. The claimed payoff is a cheap detoxification route: one 3-layer Transformer, pretrained on a safe-text corpus, learns a detoxified embedding space, and an alignment matrix maps that space into the target model's own space. The final-layer hidden state of the calibration model is then combined with the target model's hidden state as $h_{\text{agg}} = \alpha A h_T + (1-\alpha) h_B$ before decoding. On RealToxicityPrompts challenge prompts, the paper reports toxicity reductions of 0.52, 0.59, 0.01, and 3.28 points across four LLaMA-2-7B variants while perplexity stays roughly constant. The significance, if true, is that safety can be adjusted at inference time with a reusable small model rather than expensive full-model alignment.

What carries the argument

The load-bearing object is the calibration model plus the alignment matrix $A$ trained by a negative-sampling contrastive loss (Eq. 1). The calibration model is a 3-layer Transformer with the same $4{,}096$-dimensional embeddings as the LLaMA-2 family, pretrained on the WildJailbreak safe-text corpus; the alignment matrix is learned by pairing the calibration model's token embedding with the target model's token embedding for the same vocabulary token, using 10 random tokens as negatives. At inference, the same $A$ is applied to the calibration model's final-layer hidden state $h_T$, and the fused representation $h_{\text{agg}} = \alpha A h_T + (1-\alpha) h_B$ with $\alpha = 0.1$ replaces the target model's final-layer state before decoding. This one-time trained module is what is claimed to transfer across four domain-specific LLaMA-2-7B variants without fluency loss.

What would settle it

Train the alignment matrix on final-layer hidden states from the calibration and target models instead of on token embeddings and compare toxicity and perplexity; alternatively, replace $A h_T$ in Eq. (2) with a fixed random linear projection of $h_T$ of the same dimension. If the random projection matches the learned alignment in toxicity reduction, or if alignment trained on contextual states does not outperform token-level alignment, then the learned matrix is not the active ingredient in the reported detoxification.

Watch

Extended reading notes

Core claim

The paper's central claim is that a compact calibration model trained on non-toxic data, together with a linear embedding alignment learned by negative sampling, can detoxify several LLaMA-2-7B variants through a single interpolation step at the output layer. Concretely, the calibration model's final-layer representation is mapped by the alignment matrix $A$ and averaged with the target model's own final-layer representation, and only this aggregated vector feeds the decoder. The authors argue that alignment is essential: injecting unaligned embeddings makes toxicity and perplexity worse, while aligned injection lowers toxicity (from 41.59 to 41.07 on one model, from 41.87 to 38.59 on another) with virtually unchanged perplexity. They present the method as a lightweight alternative to preference-data fine-tuning, prompt engineering, and parameter editing, and show qualitatively that overtly hateful continuations become neutral while topical relevance is preserved.

Load-bearing premise

The alignment matrix $A$ is learned on static token embeddings, yet it is applied to final-layer contextual hidden states, so the method assumes these two very different representation spaces are connected by the same linear map.

Editorial extensions

If this is right

  • A single calibration model trained once can serve multiple target LLMs that share the same architecture and vocabulary, amortizing the training cost across models.
  • Toxicity on challenge prompts drops for all four tested models, with the largest drop 3.28 points on Llama2-7b-Finance, while perplexity changes by less than 0.6 points in most cases.
  • Because the intervention happens only at the final layer, the method can be switched on or off at generation time and combined with existing decoding-time controls.

Reading between the lines

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

  • An unstated risk in Eq. (2) is that $A$ is learned at the embedding layer; if that map does not transfer to contextual states, the $A h_T$ term is close to a fixed projection, and the observed gains could mostly come from averaging with a model trained on safe text.
  • The fusion weight $\alpha$ is fixed at 0.1; per-domain tuning of $\alpha$ could plausibly recover larger toxicity reductions than the reported 0.01-to-3.28 point drops, and would test whether the method is sensitive to the interpolation strength.
  • A direct extension would be to train the alignment on any layer-pair rather than only token embeddings; comparing layer-wise alignment would reveal where in the network the safe-space injection is most effective.
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 / 5 minor

Summary. The paper proposes a lightweight LLM detoxification method that trains a compact 3-layer calibration model on a non-toxic corpus (WildJailbreak), learns a linear alignment matrix A via negative sampling on shared token embeddings, and then injects the aligned calibration-model hidden state into the target model's last layer through a convex combination h_agg = α·A·h_T + (1−α)·h_B (Eq. 2). Experiments on four LLaMA-2-7B variants report toxicity and perplexity on the RealToxicityPrompts challenge set, with toxicity reductions of 0.52, 0.59, 0.01, and 3.28 points and mixed PPL changes. The central claims are that the method reduces toxicity while preserving fluency and that a one-time trained calibration model can be applied to multiple LLMs.

Significance. If the central claims hold, the approach offers an appealing lightweight intervention that avoids retraining or prompt engineering and could be reused across models sharing a common tokenizer and embedding dimension. The conceptual separation of a compact calibration model from the target LLM, combined with a released codebase, is a useful contribution to the growing literature on representation-level safety steering. However, the current evidence is insufficient: the toxicity metric is undefined, no variance or significance testing is reported, no external baselines are compared, and the alignment mechanism is trained on static embeddings but applied to contextual hidden states, leaving the main effect potentially attributable to noise. The paper is promising but requires substantive additional analysis and validation before its main claims can be accepted.

major comments (4)
  1. [§2.2–§2.3, Eq. (1) and Eq. (2)] The alignment matrix A is trained on non-contextual token embeddings: the positive example e+ is the calibration model's embedding of a single vocabulary token, and negatives are other token embeddings. In Eq. (2), however, the same A is applied to h_T, the calibration model's final-layer contextual hidden state for the entire prompt. No evidence is provided that a linear map learned on static embeddings transfers to final-layer contextual representations, which occupy a very different distribution; if transfer fails, A·h_T is effectively noise and the small, mixed effects in Table 1 are consistent with that. Please either train A on contextual representations (e.g., per-position final-layer states) or provide a validation experiment demonstrating that A improves alignment of held-out contextual hidden states (for example, by comparing cosine similarities between mapped calibration states and corresponding target-model states before and after applying A).
  2. [§3.1–§3.2, Table 1] The toxicity metric is never defined in the paper; it is only delegated to Uppaal et al. (2024b). Since toxicity is the primary outcome measure, the manuscript must specify how the score is computed: which toxicity classifier or Perspective API configuration is used, how generated continuations are aggregated (e.g., maximum over samples, mean over prompts), and what thresholds are applied. In addition, the paper should report variance across the 1,199 challenge prompts, such as standard deviations or 95% confidence intervals, and ideally a paired significance test, because the reported reductions of 0.52, 0.59, 0.01, and 3.28 points cannot be interpreted without this information.
  3. [§3.2, Table 1] The claim that the approach 'preserves fluency' is not well supported by the reported PPL numbers: PPL increases for three of the four models (4.62→4.65, 5.64→6.18, 7.05→7.39) and decreases only slightly for one (5.20→5.16). Please specify the exact PPL evaluation setup (which text is used, how detokenization is handled) and either temper the fluency claim to 'roughly comparable perplexity' or provide additional fluency evidence, such as output diversity, human evaluation, or an instruction-following metric.
  4. [§3.2] The evaluation contains no external baselines. The only comparisons are the original model and the w/o-alignment ablation, so there is no evidence that the proposed method is competitive with existing lightweight detoxification techniques, such as DeTox, representation steering (e.g., ITI or contrastive activation addition), contrastive decoding, or DPO-based variants. Adding at least one or two strong baselines, evaluated with the same toxicity metric and PPL protocol, is necessary to establish the practical value of the reported reductions and to support the paper's general claim of effectiveness.
minor comments (5)
  1. [§2.1] Some notation is nonstandard or unclear, such as 'bY' for the model output and 'bY' versus 'Y*' for the target; please use conventional math notation (e.g., Ŷ) consistently.
  2. [§3.1, footnote 2] The dataset is referred to as 'WildJailbreak' but the cited paper (Jiang et al., 2025) is titled 'WildTeaming at Scale'; please clarify whether the dataset and the citation are the same, and if not, provide the correct reference for WildJailbreak.
  3. [§3.1] The paper states that all four target models are LLaMA-2-7B variants, which implies they share a tokenizer and an embedding dimension; this makes the common-vocabulary condition in §2.2 trivially satisfied. The paper should explicitly acknowledge this limitation when claiming that the calibration model can be 'seamlessly applied to multiple LLMs,' since models with different tokenizers would require a different alignment procedure.
  4. [§3.1] The hyperparameters α=0.1 and K=10 are set without a sensitivity analysis or justification; a small study showing the effect of varying α (and possibly K) would strengthen the paper's claims about robustness and the role of the alignment term.
  5. [Table 1] The caption states 'PPL(↓)' but the notation of an arrow could be read as 'lower is better,' which is standard; please add a sentence confirming the exact protocol for computing PPL (e.g., on which corpus and with which tokenizer) so that the across-model comparisons are meaningful.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the calibration model is trained on WildJailbreak and evaluated on the external RealToxicityPrompts benchmark, so the toxicity reductions in Table 1 are not fitted to the test set.

full rationale

The derivation chain is self-contained and externally anchored. The calibration model is pre-trained on the WildJailbreak safety corpus (Section 3.1), and the central evaluation is run on the challenge_prompts subset of RealToxicityPrompts, an external benchmark not used for training. The alignment matrix A is trained in Eq. (1) using positive pairs of shared-vocabulary token embeddings and randomly selected negatives; this is an internal fitting mechanism, not a prediction of the evaluation metric. Applying that same matrix to the calibration model's final-layer hidden state in Eq. (2) is a transfer assumption whose validity is an empirical and robustness question, not a circularity: the paper's claim is tested by comparing the fused model against the original model and against a no-alignment ablation, and Table 1 reports the resulting toxicity and perplexity scores. No load-bearing self-citation, imported uniqueness theorem, or ansatz smuggled in via citation appears in the derivation; author self-citations in the introduction and related work do not support the core mechanism. The case-study examples are anecdotal but do not constitute circular reasoning. Therefore, no circular step is exhibited.

Assumptions & free parameters 4 free parameters · 4 assumptions · 1 invented entities

The method rests on a handful of hand-set choices (α = 0.1, K = 10, 3 calibration layers) and on unstated assumptions about the transfer of an embedding-level linear map to final-layer hidden states and about the detoxifying effect of plain next-token training on non-toxic data. The calibration model itself is a new component with no independent validation. Because the evaluation is external, circularity is not the main burden; evidentiary completeness is.

free parameters (4)
  • α (fusion weight) = 0.1
    Hand-set interpolation weight in Eq. (2). It keeps 90% of the original hidden state, which may explain the small toxicity reductions; no sensitivity analysis is reported.
  • K (number of negative samples) = 10
    Used in the contrastive alignment loss in Eq. (1); the value is stated without justification or ablation.
  • Calibration model depth = 3 Transformer layers
    Retains the 4096 embedding dimension of the target models; no ablation on depth is provided.
  • Calibration-model training schedule = unspecified
    Learning rate, epochs, batch size, and the WildJailbreak subset or filtering used for pre-training are not reported.
assumptions (4)
  • domain assumption Training a 3-layer Transformer on WildJailbreak with standard next-token prediction yields an embedding space that naturally downweights toxic features.
    No toxicity-specific supervision, probing, or negative toxicity samples are used; the claimed detoxifying effect of the 'detoxified embedding space' is asserted in Section 2.1.
  • domain assumption The alignment matrix A learned on non-contextual token embeddings via Eq. (1) transfers correctly to final-layer contextual hidden states h_T in Eq. (2).
    Section 2.2 trains A at the token-embedding level; Section 2.3 applies it to last-layer hidden states with no intermediate validation.
  • domain assumption Random negatives sampled uniformly from the common vocabulary provide a useful contrastive signal for alignment.
    Random vocabulary tokens are easy to separate from the positive token, so Eq. (1) may not learn the semantic correspondence needed for hidden-state fusion.
  • domain assumption Toxicity and perplexity as computed in Uppaal et al. (2024b) capture the properties claimed (safety and fluency).
    The toxicity metric is never defined in this paper; the reader cannot verify what is being measured.
invented entities (1)
  • compact calibration model
    purpose: Learns a detoxified embedding space from a non-toxic corpus and steers target LLMs via aligned output-layer fusion.
    The paper's central novelty is the calibration model and its claimed 'detoxified embedding space'. Its detoxifying property is only evidenced inside this paper's fused outputs; no external benchmark, probing, or ablation isolates the calibration model's own contribution.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Detoxification of Large Language Models through Output-layer Fusion with a Calibration Model." pith.science (2026). https://pith.science/paper/6KNMSIJZ

@misc{pith2026250601266,
  author       = {Pith},
  title        = {Pith review of: Detoxification of Large Language Models through Output-layer Fusion with a Calibration Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6KNMSIJZ}},
  note         = {Machine review of arXiv:2506.01266}
}
read the original abstract

Existing approaches for Large language model (LLM) detoxification generally rely on training on large-scale non-toxic or human-annotated preference data, designing prompts to instruct the LLM to generate safe content, or modifying the model parameters to remove toxic information, which are computationally expensive, lack robustness, and often compromise LLMs' fluency and contextual understanding. In this paper, we propose a simple yet effective approach for LLM detoxification, which leverages a compact, pre-trained calibration model that guides the detoxification process of a target LLM via a lightweight intervention in its generation pipeline. By learning a detoxified embedding space from non-toxic data, the calibration model effectively steers the LLM away from generating harmful content. This approach only requires a one-time training of the calibration model that is able to be seamlessly applied to multiple LLMs without compromising fluency or contextual understanding. Experiment results on the benchmark dataset demonstrate that our approach reduces toxicity while maintaining reasonable content expression.

Figures

Figures reproduced from arXiv: 2506.01266 by the authors.

Figure 1
Figure 1. The overall architecture of our approach. The left part presents the first step to train a calibration model [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

47 extracted references · 22 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]

    Sumanth Dathathri, Andrea Madotto, Janice Lan, Jane Hung, Eric Frank, Piero Molino, Jason Yosinski, and Rosanne Liu. 2020. Plug and play language models: A simple approach to controlled text generation. In International Conference on Learning Representations (ICLR)

    work page 2020

  5. [5]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT : P re-training of D eep B idirectional T ransformers for L anguage U nderstanding. In Proceedings of the 2019 Conference of the North A merican Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers) , pages 4171--4...

    2019

  6. [6]

    Ruyi Gan, Ziwei Wu, Renliang Sun, Junyu Lu, Xiaojun Wu, Dixiang Zhang, Kunhao Pan, Ping Yang, Qi Yang, Jiaxing Zhang, et al. 2023. Ziya2: Data-centric Learning is All LLMs Need . arXiv preprint arXiv:2311.03301

    work page Pith review arXiv 2023

  7. [7]

    Samuel Gehman, Suchin Gururangan, Maarten Sap, Yejin Choi, and Noah A. Smith. 2020. Realtoxicityprompts: Evaluating neural toxic degeneration in language models. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 3356--3369

    work page 2020

  8. [8]

    Mor Geva, Avi Caciularu, Kevin Ro Wang, and Yoav Goldberg. 2022. Transformer feed-forward layers build predictions by promoting concepts in the vocabulary space. arXiv preprint arXiv:2203.14680

    arXiv 2022

Show all 47 references
  1. [9]

    Jialong Han, Yan Song, Wayne Xin Zhao, Shuming Shi, and Haisong Zhang. 2018. H yperdoc2vec: D istributed R epresentations of H ypertext D ocuments. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2384--2...

    2018

  2. [10]

    Shih-Cheng Huang, Pin-Zu Li, Yu-Chi Hsu, Kuang-Ming Chen, Yu Tung Lin, Shih-Kai Hsiao, Richard Tsai, and Hung-Yi Lee. 2024. Chat vector: A simple approach to equip llms with instruction following and model alignment in new languages. In Proceedings of the 62nd Annual Meeting o...

    2024

  3. [11]

    Liwei Jiang, Kavel Rao, Seungju Han, Allyson Ettinger, Faeze Brahman, Sachin Kumar, Niloofar Mireshghallah, Ximing Lu, Maarten Sap, Yejin Choi, et al. 2025. Wildteaming at scale: From in-the-wild jailbreaks to (adversarially) safer language models. Advances in Neural Informati...

    2025

  4. [12]

    Yongmin Kim, Takeshi Kojima, Yusuke Iwasawa, and Yutaka Matsuo. 2024. Decoupling noise and toxic parameters for language model detoxification by task vector merging. In First Conference on Language Modeling

    2024

  5. [13]

    Diederik P Kingma and Jimmy Ba. 2014. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980

    2014 arXiv

  6. [14]

    Tomasz Korbak, Kejian Shi, Angelica Chen, Rasika Vinayak Bhalerao, Christopher Buckley, Jason Phang, Samuel R Bowman, and Ethan Perez. 2023. Pretraining language models with human preferences. In International Conference on Machine Learning, pages 17506--17533. PMLR

    2023

  7. [15]

    Ben Krause, Akhilesh Deepak Gotmare, Bryan McCann, Nitish Shirish Keskar, Shafiq Joty, Richard Socher, and Nazneen Fatema Rajani. 2020. GeDi : Generative discriminator guided sequence generation. arXiv preprint arXiv:2009.06367

    2020 arXiv

  8. [16]

    Chak Tou Leong, Yi Cheng, Jiashuo Wang, Jian Wang, and Wenjie Li. 2023. Self-detoxifying language models via toxification reversal. arXiv preprint arXiv:2310.09573

    2023 arXiv

  9. [17]

    Chenxi Li, Yuanhe Tian, Zhaxi Zerong, Yan Song, and Fei Xia. 2024 a . Challenging large language models with new tasks: A study on their adaptability and robustness. In Findings of the Association for Computational Linguistics: ACL 2024, pages 8140--8162, Bangkok, Thailand

    2024

  10. [18]

    Xuying Li, Zhuo Li, Yuji Kosuga, Yasuhiro Yoshida, and Victor Bian. 2024 b . Precision knowledge editing: Enhancing safety in large language models. arXiv preprint arXiv:2410.03772

    2024 arXiv

  11. [19]

    Sheng Liu, Haotian Ye, Lei Xing, and James Zou. 2023. In-context vectors: Making in context learning more effective and controllable through latent space steering. arXiv preprint arXiv:2311.06668

    2023 arXiv

  12. [20]

    Ximing Lu, Sean Welleck, Jack Hessel, Liwei Jiang, Lianhui Qin, Peter West, Prithviraj Ammanabrolu, and Yejin Choi. 2022. Quark: Controllable text generation with reinforced unlearning. Advances in neural information processing systems, 35:27591--27609

    2022

  13. [21]

    Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. 2022. Locating and editing factual associations in gpt. Advances in Neural Information Processing Systems, 35:17359--17372

    2022

  14. [22]

    Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. 2013. Efficient estimation of word representations in vector space. arXiv preprint arXiv:1301.3781

    2013 arXiv

  15. [23]

    Tong Niu, Caiming Xiong, Semih Yavuz, and Yingbo Zhou. 2024. Parameter-efficient detoxification with contrastive decoding. arXiv preprint arXiv:2401.06947

    2024 arXiv

  16. [24]

    Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul Christiano, Jan Leike, and...

    2022

  17. [25]

    Nina Panickssery, Nick Gabrieli, Julian Schulz, Meg Tong, Evan Hubinger, and Alexander Matt Turner. 2023. Steering llama 2 via contrastive activation addition. arXiv preprint arXiv:2312.06681

    2023 arXiv

  18. [26]

    Jeffrey Pennington, Richard Socher, and Christopher Manning. 2014. G lo V e: Global vectors for word representation. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing ( EMNLP ) , pages 1532--1543, Doha, Qatar

    2014

  19. [27]

    Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and Luke Zettlemoyer

    Matthew E. Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and Luke Zettlemoyer. 2018. Deep contextualized word representations. In Proceedings of the 2018 Conference of the North A merican Chapter of the Association for Computational Linguistic...

    2018

  20. [28]

    Yifu Qiu, Zheng Zhao, Yftah Ziser, Anna Korhonen, Edoardo Maria Ponti, and Shay B. Cohen. 2024. Spectral editing of activations for large language model alignment. In Advances in Neural Information Processing Systems

    2024

  21. [29]

    Manning, and Chelsea Finn

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D. Manning, and Chelsea Finn. 2023. Direct preference optimization: Your language model is secretly a reward model. In Advances in Neural Information Processing Systems

    2023

  22. [30]

    Timo Schick, Sahana Udupa, and Hinrich Schütze. 2021. Self-diagnosis and self-debiasing: A proposal for reducing corpus-based bias in nlp. Transactions of the Association for Computational Linguistics, 9:1408--1424

    2021

  23. [31]

    Yan Song, Chia-Jung Lee, and Fei Xia. 2017. L earning W ord R epresentations with R egularization from P rior K nowledge. In Proceedings of the 21st Conference on Computational Natural Language Learning ( C o NLL 2017) , pages 143--152

    2017

  24. [32]

    Yan Song and Shuming Shi. 2018. Complementary Learning of Word Embeddings. In IJCAI, pages 4368--4374

    2018

  25. [33]

    Yan Song, Shuming Shi, and Jing Li. 2018 a . Joint L earning E mbeddings for C hinese W ords and T heir C omponents via L adder S tructured N etworks. In Proceedings of the 27th International Joint Conference on Artificial Intelligence, pages 4375--4381

    2018

  26. [34]

    Yan Song, Shuming Shi, Jing Li, and Haisong Zhang. 2018 b . Directional S kip- G ram: E xplicitly D istinguishing L eft and R ight C ontext for W ord E mbeddings. In Proceedings of the 2018 Conference of the North A merican Chapter of the Association for Computational Linguist...

    2018

  27. [35]

    Yan Song, Tong Zhang, Yonggang Wang, and Kai-Fu Lee. 2021. ZEN 2.0: C ontinue T raining and A daption for N -gram E nhanced T ext E ncoders. arXiv preprint arXiv:2105.01279

    2021 arXiv

  28. [36]

    Hashimoto

    Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li, Carlos Guestrin, Percy Liang, and Tatsunori B. Hashimoto. 2023. Stanford Alpaca: An Instruction-following LLaMA model . GitHub repository

    2023

  29. [37]

    Yuanhe Tian, Ruyi Gan, Yan Song, Jiaxing Zhang, and Yongdong Zhang. 2024 a . C hi M ed- GPT : A C hinese medical large language model with full training regime and better alignment to human preferences. In Proceedings of the 62nd Annual Meeting of the Association for Computati...

    2024

  30. [38]

    Yuanhe Tian, Yan Song, and Fei Xia. 2020. Supertagging C ombinatory C ategorial G rammar with A ttentive G raph C onvolutional N etworks. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6037--6044

    2020

  31. [39]

    Yuanhe Tian, Fei Xia, and Yan Song. 2024 b . Dialogue Summarization with Mixture of Experts based on Large Language Models . In The 62nd Annual Meeting of the Association for Computational Linguistics, Bangkok, Thailand

    2024

  32. [40]

    Yuanhe Tian, Fei Xia, and Yan Song. 2024 c . Learning Multimodal Contrast with Cross-modal Memory and Reinforced Contrast Recognition . In The 62nd Annual Meeting of the Association for Computational Linguistics, Bangkok, Thailand

    2024

  33. [41]

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. 2023. LLaMA 2: Open Foundation and Fine-tuned Chat Models . arXiv preprint arXiv:2307.09288

    2023 arXiv

  34. [42]

    Rheeya Uppaal, Apratim Dey, Yiting He, Yiqiao Zhong, and Junjie Hu. 2024 a . DeTox : Toxic subspace projection for model editing. arXiv preprint arXiv:2405.13967

    2024 arXiv

  35. [43]

    Rheeya Uppaal, Apratim Dey, Yiting He, Yiqiao Zhong, and Junjie Hu. 2024 b . Model editing as a robust and denoised variant of dpo: A case study on toxicity. In Neurips Safe Generative AI Workshop 2024

    2024

  36. [44]

    Boxin Wang, Wei Ping, Chaowei Xiao, Peng Xu, Mostofa Patwary, Mohammad Shoeybi, Bo Li, Anima Anandkumar, and Bryan Catanzaro. 2022. Exploring the limits of domain-adaptive training for detoxifying large-scale language models. Advances in Neural Information Processing Systems, ...

    2022

  37. [45]

    Mengru Wang, Ningyu Zhang, Ziwen Xu, Zekun Xi, Shumin Deng, Yunzhi Yao, Qishen Zhang, Linyi Yang, Jindong Wang, and Huajun Chen. 2024. Detoxifying large language models via knowledge editing. arXiv preprint arXiv:2403.14472

    2024 arXiv

  38. [46]

    Xiaojun Wu, Dixiang Zhang, Ruyi Gan, Junyu Lu, Ziwei Wu, Renliang Sun, Jiaxing Zhang, Pingjian Zhang, and Yan Song. 2024. Taiyi-diffusion-xl: advancing bilingual text-to-image generation with large vision-language model support. arXiv preprint arXiv:2401.14688

    2024 arXiv

  39. [47]

    o rg Schl \

    Paul Youssef, Zhixue Zhao, Daniel Braun, J \"o rg Schl \"o tterer, and Christin Seifert. 2025. Position: Editing large language models poses serious safety risks. arXiv preprint arXiv:2502.02958

    2025 arXiv

Pith tools

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