Pith. sign in

REVIEW 4 major objections 7 minor 97 references

Calling the Bluff: Detecting Ever-Shifting Harmful Chat Dialogue via Ordered Reasoning Chain Regularization

T0 review · 4 major / 7 minor · reviewed 2026-08-14 · deepseek-v4-flash

Pith's one-line read The paper claims that harmful chat's surface drift hides an invariant four-step reasoning chain, and that a detector trained to reproduce it reaches 0.934 harm-type macro-F1.

desk verdict Architecturally solid, empirically conditional: the BRACE combination is new and the ablations support ORC, but the ever-shifting claim is asserted, not tested, and the evaluation needs closest-prior baselines and concrete annotation numbers. read the letter →

arxiv 2608.08451 v1 pith:NWNTHW74 submitted 2026-08-09 cs.CL cs.AI

classification cs.CLcs.AI
keywords harmfuldialoguedetectionlexicalevasionorderedreasoningchainintermediatesupervisionprototypememorybankmulti-taskregularizationharmtypeclassificationChinese-English-Spanishdataset
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 harmful chat dialogues, even when their surface words are constantly changed to dodge filters, obey a fixed four-step reasoning chain: what topic is being discussed, where the harm signals appear, how severe the harm is, and which harm type it constitutes. It proposes BRACE, a detector that trains a model to reproduce this Ordered Reasoning Chain (Topic $\rightarrow$ Indicator $\rightarrow$ Severity $\rightarrow$ Type) as an intermediate-supervised regularizer, blended with direct prediction heads, prototype memory, and feature path disentanglement. On a 60,000-dialogue Chinese-English-Spanish dataset across four domains and five harm categories, BRACE reaches 0.934 harm-type macro F1 with an encoder backbone and 0.949 with LoRA-tuned decoder backbones. The practical point is that if harmful intent leaves an invariant chain under changing words, moderation systems can be built to track the chain rather than chase blocklists.

What carries the argument

The load-bearing mechanism is the Ordered Reasoning Chain (ORC), a four-stage differentiable decomposition $C=(f_1,f_2,f_3,f_4)$ (Topic $\rightarrow$ Indicator $\rightarrow$ Severity $\rightarrow$ Type) in which each stage receives its own supervision and conditions on the previous output: a topic distribution over 20 topics from the CLS vector; a token-level indicator heatmap over 32 indicators conditioned on the topic; a five-level severity distribution from attention-pooled tokens; and a five-way type distribution from mean-pooled indicators plus severity context. The chain is trained as a structured regularizer ($L_{\text{chain}}$) and blended with direct heads at 0.3 weight, so the direct path remains primary while the chain supplies inductive bias. Around it, a Prototype Memory Bank of $C\times K\times D$ learnable anchors augments features through cross-attention for the direct type head, and feature path disentanglement sends type, severity, and binary through separate features, with a six-expert MoE router carrying binary detection.

What would settle it

Give BRACE and a control model with the same architecture but shuffled ORC stage order a held-out set of adversarially evaded dialogues whose surface words never appeared in training but whose underlying topic-indicator-severity-type chains are unchanged; if the control matches BRACE, the ordered structure is not what is doing the work.

Watch

Extended reading notes

Core claim

BRACE's central claim is that the apparent drift of harmful dialogue, type-shifting and lexical evasion, does not change the underlying communicative reasoning, and that making a detector reconstruct that reasoning in order is what produces the accuracy gain. The model decomposes classification into four intermediate-supervised stages, topic profiling, indicator detection, severity assessment, and type classification, each conditioning on the previous one, trained as a regularizer and blended 70/30 with direct heads. A prototype memory bank supplies category-level features for the direct type head, and binary detection is isolated in a separate MoE router. The paper reports 0.934 harm-type macro F1 with RoBERTa-wwm-ext and 0.949 with LoRA-tuned decoders, with removal of the ORC causing the largest ablation drop ($-0.101$ Harm m-F1) and enabling disambiguation of semantically ambiguous types such as doxxing versus SWATting.

Load-bearing premise

The load-bearing premise is that every harmful dialogue really does have a well-defined topic, indicator, severity, and type chain, and that the human- and LLM-produced labels for those stages are accurate enough to train on.

Editorial extensions

If this is right

  • If the ORC invariance holds, a detector trained this way should keep identifying harm when attackers invent unseen euphemisms, because the signal is in the reasoning chain rather than the surface words.
  • The intermediate outputs give moderators an audit trail: a flagged dialogue is explainable by which topic, indicator, severity, and type fired, instead of by a single opaque score.
  • The backbone-agnostic gains mean the method can run on small models (27M parameters) with little loss, which matters for deployment where large decoders are too costly.
  • Because severity context flips type predictions (moderate doxxing versus critical SWATting), the chain supplies a practical route to separating legally different harms that share surface cues.

Reading between the lines

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

  • A direct stress test would adversarially paraphrase test dialogues to produce novel surface words while keeping the ORC labels fixed; the gap between BRACE and a no-chain model on that set would quantify how much of the gain is truly invariant reasoning.
  • The same invariant-chain-under-surface-drift idea transfers to other evasion-prone moderation tasks, such as fraud, spam, and coordinated manipulation, where intent outlives wording.
  • If the ORC labels themselves encode annotator or LLM bias, the regularizer would teach those biases as if they were universal, so releasing annotation guidelines and numeric agreement values is the natural next step for verification.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 7 minor

Summary. The paper proposes BRACE, a harmful chat dialogue detection model that encodes an Ordered Reasoning Chain (ORC) — Topic, Indicator, Severity, Type — as a four-stage differentiable regularizer blended with direct prediction heads. The model also uses a prototype memory bank for feature augmentation and a lightweight MoE router for binary detection. The authors construct a 60,000-dialogue multi-source Chinese-English-Spanish dataset with five harm categories and report that BRACE achieves harm-type macro F1 of 0.934 with RoBERTa-wwm-ext and 0.949 with decoder backbones. Ablations indicate that removing the ORC causes the largest performance drop, which the authors interpret as evidence that the ORC is the dominant mechanism for disambiguating harmful types.

Significance. If the results are reproducible, the ORC regularizer combined with prototype-based feature augmentation is a genuinely interesting architectural idea for fine-grained harmful content detection. The paper is largely self-contained in its architecture description, and the planned release of code and dataset is commendable. The ablation study is informative, and the cross-backbone evaluation across many sources is a useful practical contribution. However, the paper's central claim about robustness to 'ever-shifting' harmful dialogue is not directly evaluated, and several methodological details — annotation agreement, variance estimates, and comparisons to the closest prior systems — are missing. The significance of the reported gains cannot be fully assessed until these gaps are addressed.

major comments (4)
  1. [Experimental Design / Datasets and Data Preparation] The central claim of the paper — that BRACE detects 'ever-shifting' harmful dialogue under lexical evasion and type-shifting — is not tested by the evaluation protocol. The dataset is split into a single static train/val/test partition with source-level stratification, and no transformed, evasive, or temporally shifted test set is constructed. The only evidence for robustness under lexical evasion is the motivating example in Figure 1 and a qualitative walkthrough in the Methodology section. To support the title and abstract claims, the authors should add an adversarial evaluation with common evasion transformations (pinyin transliteration, typo insertion, synonym substitution, word segmentation tricks) and an out-of-source or temporal split, and report harm-type macro F1 under these conditions.
  2. [Datasets and Data Preparation] The annotation reliability statement is unsupported by numbers. The text says 'LLM proposals achieve substantial agreement with adjudicated labels' and 'Human–Human κ exceeding LLM–Human κ by a consistent margin,' but no numerical κ values are reported. Because the ORC intermediate labels (topics, indicators, severity) and the final harm-type labels are produced by the same LLM-assisted pipeline with human review, the chain supervision may simply reinforce annotator or LLM biases rather than discovering an invariant structure. The authors should report exact κ values for LLM–Human and Human–Human agreement on the calibration subset, and ideally perform an experiment using independently human-only intermediate labels or a different LLM for intermediate labels to probe the circularity risk.
  3. [Table 2 / Evaluation Metrics] Headline numbers are reported without variance. All encoder results are stated as 3-seed means, but no standard deviations or per-seed values are given, and decoder results use a single seed. The claimed margins over linear probes (e.g., +33.2% for RoBERTa-wwm) and the claim of 'backbone-agnostic' behavior rest on comparisons without error bars or significance tests. The text cites p=0.265 for ERNIE-Mini versus larger encoders but does not specify the test used. The authors should report mean ± std over the three seeds and run paired significance tests (e.g., bootstrap or paired t-test) on the test set.
  4. [Baselines and Related Works] The paper does not compare against the closest prior systems it cites as addressing the same problem. The Related Works section names RepMD, JADE, ReasoningShield, and SafetyAnalyst as methods that handle lexical evasion or reasoning-based safety detection, and the introduction claims these methods 'fail to identify' ever-shifting harmful dialogues. None of these systems appears in Table 2 or elsewhere in the evaluation. Without empirical comparisons on the same test set, the claim that BRACE improves over prior work on ever-shifting detection is not substantiated. The authors should include these or other admissible baselines, or clearly justify their exclusion.
minor comments (7)
  1. [Introduction and Table 1] The manuscript is inconsistent about the number of sources and benchmarks: the introduction says 'over 20 dialogue safety benchmarks,' Table 1 lists 25 public sources, and the abstract says '4 domains and 5 harm categories.' Please reconcile these counts.
  2. [Results] The Results section says the experimental evaluation is structured around three core research questions, but only RQ1 and RQ2 are defined in Experimental Design. Please define RQ3 or revise the wording.
  3. [Abstract and Table 2] Severity accuracy of 0.818 is reported in the Abstract and Conclusion but does not appear in any table in the main text. Please add the severity result to Table 2 or a separate table.
  4. [Implementation Details] The text states that 'Full hyperparameters are provided in Experimental Design,' but several important hyperparameters — the blend ratio α, the loss weights in Eq. (10), prototype temperature τ, EMA momentum m, and replacement threshold θ_replace — are given without sensitivity analysis. A short sensitivity table in the supplement would strengthen the robustness claims.
  5. [Methodology / Feature Path Disentanglement] The paper claims 'feature path disentanglement' eliminates gradient competition, but no experiment directly tests this claim. An ablation that reverses the routing (e.g., feeding type classification through x_cls and severity through f_aug) would clarify whether the performance gain comes from the routing itself or from the higher-capacity features.
  6. [Datasets and Data Preparation] The dataset is constructed from 25 heterogeneous sources with different annotation schemes, yet the paper reports exactly 12,000 dialogues per category. Please describe how multi-label dialogues and existing category taxonomies were mapped to the five harm types, and whether any dialogue-level deduplication was performed across sources.
  7. [Definition of ORC / Methodology] The 20 topics and 32 indicators are described as derived via LLM-assisted annotation, but the taxonomies themselves are not listed. For reproducibility, include the full topic and indicator taxonomies in the appendix or as supplementary material.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the ORC is an auxiliary regularizer trained on intermediate labels, the main F1 is measured on a held-out test set with human-adjudicated labels, and ablations provide an independent comparison.

full rationale

BRACE's derivation chain is self-contained with respect to circularity. The ORC (Topic -> Indicator -> Severity -> Type) is introduced as a structured regularizer (Eq. 4) with intermediate auxiliary losses, and the final harm-type prediction is a 70/30 blend of a direct head and the chain head (Eq. 10); the target type labels are not defined in terms of the chain's own outputs, and the chain's intermediate labels (T=20 topics, I=32 indicators) are produced by LLM-assisted annotation with human review rather than being extracted from the model's predictions. The central performance claim (Harm m-F1 of 0.934) is computed on a 9,000-dialogue source-stratified held-out test set, and the ablation study removes ORC as a component and measures a drop of -0.101, which is an independent comparison rather than a self-fulfilling construction. The paper's self-citations (RepMD/JADE, Jiang et al. 2026a,b) appear only in related-work positioning and are not used as load-bearing evidence for the architecture's effectiveness. The lack of an adversarial or temporally shifted evaluation for the 'ever-shifting' claim is a potential correctness or generalization gap, but it is not a circularity: no fitted parameter is renamed as a prediction, and no equation reduces to its own input. Therefore the circularity score is 0.

Assumptions & free parameters 11 free parameters · 4 assumptions · 1 invented entities

The central performance claim rests on the ORC invariance hypothesis, on the reliability of LLM-plus-human intermediate labels, and on a large set of hand-set hyperparameters. None of these are independently verified outside the paper's own benchmark, which is why the circularity burden, while not high in the equation-identity sense, still warrants accounting.

free parameters (11)
  • Direct/chain blend ratio alpha = 0.7 direct / 0.3 chain
    Hand-chosen so direct heads dominate; ablation mentions blend ratio but no sweep is shown in the main text.
  • Loss weights in L_total = L_bin 1.0, L_type 0.5, L_proto 0.3, L_chain 0.1, L_sev 0.1, L_div 0.1, L_bal 0.01
    Hand-set weights; no sensitivity analysis is reported in the main text.
  • Number of prototypes K per category = 8
    Hand-chosen; no sweep is reported.
  • Prototype temperature tau = 0.07
    Standard contrastive temperature; no sweep is reported.
  • Prototype EMA momentum m = 0.99
    Taken from MoCo; not adapted to this task.
  • Prototype replacement threshold theta_replace = 0.5, with threshold schedule 0.3 to 0.7 over 80% training
    Hand-set to avoid dead prototypes.
  • Diversity penalty margin = 0.3
    Hand-set margin in L_div.
  • MoE expert count E = 6
    One expert per harm category plus a general expert; no sweep is reported.
  • Topic count T = 20
    Taxonomy size chosen during LLM-assisted annotation; affects capacity of f1.
  • Indicator count I = 32
    Taxonomy size chosen during annotation; affects capacity of f2.
  • Binarization threshold = 0.5
    Used to turn predicted probabilities into labels; no threshold tuning is reported.
assumptions (4)
  • domain assumption Harmful chat dialogues share an Ordered Reasoning Chain (Topic -> Indicator -> Severity -> Type) that is invariant under lexical evasion and type-shifting.
    Central motivation in Definition of ORC and Figure 1; asserted with examples, not formally proven.
  • ad hoc to paper The 20-topic and 32-indicator taxonomies generated by DeepSeek-V4-Pro with human review are complete and reliable for all five harm categories.
    Datasets and Data Preparation describes the pipeline but gives no numeric kappa values.
  • ad hoc to paper The 70/30 blend of direct and chain predictions is a valid aggregation that does not distort the task-specific features.
    Methodology states alpha=0.7 makes direct heads primary; no theoretical or empirical justification beyond an ablation mention.
  • domain assumption Source-stratified splitting prevents topic or indicator label leakage between train and test.
    Datasets and Data Preparation says the partition is source-stratified; no leakage analysis is provided.
invented entities (1)
  • Ordered Reasoning Chain (ORC)
    purpose: A four-stage decomposition used as a structured regularizer with intermediate supervision.
    Its utility is demonstrated only through ablations on the same dataset; no external falsifiable prediction is offered in the main text.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Calling the Bluff: Detecting Ever-Shifting Harmful Chat Dialogue via Ordered Reasoning Chain Regularization." pith.science (2026). https://pith.science/paper/NWNTHW74

@misc{pith2026260808451,
  author       = {Pith},
  title        = {Pith review of: Calling the Bluff: Detecting Ever-Shifting Harmful Chat Dialogue via Ordered Reasoning Chain Regularization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NWNTHW74}},
  note         = {Machine review of arXiv:2608.08451}
}
read the original abstract

Harmful chat dialogues are ever-shifting through type-shifting and lexical evasion, yet we find they share invariant principles, i.e., an Ordered Reasoning Chain (ORC) of recurring topics, harm language indicators, severity hierarchies, and type characteristics, which can help us capture the key information in the frequently changing lexical expressions. We propose BRACE, which encodes the ORC as four differentiable stages (Topic -> Indicator -> Severity -> Type) with intermediate supervision, serving as a structured regularizer blended with direct heads, and supported by prototype-based feature augmentation and feature path disentanglement. The evaluation results show that, across 4 domains and 5 harm categories, BRACE achieves harm-type macro F1 of 0.934 (RoBERTa-wwm-ext, 3-seed mean), with decoder backbones (Qwen3-1.7B LoRA) reaching 0.949. Ablation studies show that all components contribute to BRACE, and the structural decomposition of ORC enables BRACE to distinguish harmful types with semantic ambiguity. Disclaimer: This paper may contain content that is disturbing to some readers.

Figures

Figures reproduced from arXiv: 2608.08451 by the authors.

Figure 1
Figure 1. Motivation of ORC for harmful dialogue detection. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of BRACE. Methodology We now realize each architectural component defined above as a concrete module. The ordered reasoning chain provides structured regularization via intermediate supervision; di￾rect heads serve as primary classifiers with a 70/30 blend; prototype memory produces category-aware features; and a lightweight MoE router handles binary detection [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 4
Figure 4. The t-SNE results’ distribution of dimensionality [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

97 extracted references · 61 canonical work pages

  1. [1]

    Davidson, T.; Warmsley, D.; Macy, M.; and Weber, I. 2017. Automated Hate Speech Detection and the Problem of Offensive Language. In Proceedings of the 11th International Conference on Web and Social Media (ICWSM), 512--515

  2. [2]

    Equitz, W. H. R.; and Cover, T. M. 1991. Successive Refinement of Information. IEEE Transactions on Information Theory, 37(2): 269--275

  3. [3]

    Fedus, W.; Zoph, B.; and Shazeer, N. 2022. Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity. Journal of Machine Learning Research, 23(120): 1--39

  4. [4]

    Goyal, A.; Zhan, X.; Chen, Y.; Saha, K.; and Chandrasekharan, E. 2025. MoMoE : Mixture of Moderation Experts Framework for AI -Assisted Online Governance. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing (EMNLP)

  5. [5]

    He, K.; Fan, H.; Wu, Y.; Xie, S.; and Girshick, R. 2020. Momentum Contrast for Unsupervised Visual Representation Learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 9726--9735

  6. [6]

    Ho, S.; Liu, M.; Du, L.; Gao, L.; and Long, G. 2024. Prototype-Guided Memory Replay for Continual Learning. IEEE Transactions on Neural Networks and Learning Systems

  7. [7]

    Huertas-Garc \' a, \'A .; Mart \' n, A.; Huertas-Tato, J.; and Camacho, D. 2023. Countering Malicious Content Moderation Evasion in Online Social Networks: Simulation and Detection of Word Camouflage. Applied Soft Computing, 145: 110574

  8. [8]

    Ji, J.; Hong, D.; Zhang, B.; Chen, B.; Dai, J.; Zheng, B.; Qiu, T.; Zhou, J.; Wang, K.; Li, B.; Han, S.; Guo, Y.; and Yang, Y. 2025. PKU-SafeRLHF : Towards Multi-Level Safety Alignment for LLMs with Human Preference. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (ACL), 31983--32016

Show all 97 references
  1. [9]

    Ji, J.; Liu, M.; Dai, J.; Pan, X.; Zhang, C.; Bian, C.; Chen, B.; Sun, R.; Wang, Y.; and Yang, Y. 2023. BeaverTails : Towards Improved Safety Alignment of LLM via a Human-Preference Dataset. In Advances in Neural Information Processing Systems (NeurIPS), volume 36. Datasets an...

  2. [10]

    Jiang, Z.; Li, M.; Wang, J.; Huang, Y.; Huang, J.; Chang, Z.; Li, Z.; and Wang, Q. 2026 a . All Changes May Have Invariant Principles: Improving Ever-Shifting Harmful Meme Detection via Design Concept Reproduction. In Proceedings of the 62nd Annual Meeting of the Association f...

  3. [11]

    Jiang, Z.; Liu, M.; Qin, Y.; and Liu, B. 2026 b . Breaking Free from Ivory Tower: Evaluating and Enhancing Real-world C hinese Underground Adversarial Jargon Detection. In Proceedings of the 47th IEEE Symposium on Security and Privacy (S&P) , 417--435

  4. [12]

    Kang, H.; Chen, J.; Li, Y.; Miao, X.; Xu, M.; Zhou, S.; Zhong, M.; Zhu, Y.; and Qian, T. 2025. Developing Continuous Toxicity Detection Against Increasing Types of Perturbed Toxic Text. Computers & Security

  5. [13]

    A.; Shakil, H.; Baruah, J.; and Kalita, J

    Khapre, S.; Mersha, M. A.; Shakil, H.; Baruah, J.; and Kalita, J. 2025. Toxicity in Online Platforms and AI Systems: A Survey of Needs, Challenges, Mitigations, and Future Directions. Expert Systems with Applications. ArXiv:2509.25539

  6. [14]

    Khosla, P.; Teterwak, P.; Wang, C.; Sarna, A.; Tian, Y.; Isola, P.; Maschinot, A.; Liu, C.; and Krishnan, D. 2020. Supervised Contrastive Learning. In Advances in Neural Information Processing Systems, volume 33, 18661--18673

  7. [15]

    Kiela, D.; Firooz, H.; Mohan, A.; Goswami, V.; Singh, A.; Ringshia, P.; and Testuggine, D. 2020. The Hateful Memes Challenge: Detecting Hate Speech in Multimodal Memes. In Advances in Neural Information Processing Systems, volume 33, 2611--2624

  8. [18]

    Li, J.-J.; Pyatkin, V.; Kleiman-Weiner, M.; Jiang, L.; Dziri, N.; Collins, A.; Schaich Borg, J.; Sap, M.; Choi, Y.; and Levine, S. 2025 b . SafetyAnalyst : Interpretable, Transparent, and Steerable Safety Moderation for AI Behavior. In Proceedings of the 42nd International Con...

  9. [19]

    Lin, T.-Y.; Goyal, P.; Girshick, R.; He, K.; and Doll\' a r, P. 2017. Focal Loss for Dense Object Detection. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2980--2988

  10. [20]

    Lu, J.; Xu, B.; Zhang, X.; Min, C.; Yang, L.; and Lin, H. 2023. Facilitating Fine-grained Detection of Chinese Toxic Language: Hierarchical Taxonomy, Resources, and Benchmarks. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (ACL), 16...

  11. [22]

    M.; Moreno-Mu \ n oz, A.; Plaza-del Arco, F

    M \'a rmol-Romero, A. M.; Moreno-Mu \ n oz, A.; Plaza-del Arco, F. M.; Molina-Gonz \'a lez, M. D.; Mart \'i n-Valdivia, M. T.; Ure \ n a-L \'o pez, L. A.; and Montejo-R \'a ez, A. 2024. MentalRiskES : A New Corpus for Early Detection of Mental Disorders in Spanish . In Proceed...

  12. [23]

    M.; Biemann, C.; Goyal, P.; and Mukherjee, A

    Mathew, B.; Saha, P.; Yimam, S. M.; Biemann, C.; Goyal, P.; and Mukherjee, A. 2021. HateXplain : A Benchmark Dataset for Explainable Hate Speech Detection. In Proceedings of the Conference on Artificial Intelligence, volume 35, 14867--14875

  13. [24]

    Mei, J.; Chen, J.; Yang, G.; Lin, W.; and Byrne, B. 2025. Robust Adaptation of Large Multimodal Models for Retrieval Augmented Hateful Meme Detection. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, 23817--23839

  14. [25]

    Qiu, H.; Ma, L.; and Lan, Z. 2024. PsyGUARD : An Automated System for Suicide Detection and Risk Assessment in Psychological Counseling. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing (EMNLP), 4581--4607

  15. [26]

    R \"o ttger, P.; Seelawi, H.; Nozza, D.; Talat, Z.; and Vidgen, B. 2022. Multilingual HateCheck : Functional Tests for Multilingual Hate Speech Detection Models. In Proceedings of the Sixth Workshop on Online Abuse and Harms (WOAH), 154--169

  16. [27]

    R \"o ttger, P.; Vidgen, B.; Nguyen, D.; Waseem, Z.; Margetts, H.; and Pierrehumbert, J. 2021. HateCheck : Functional Tests for Hate Speech Detection Models. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics (ACL), 41--58

  17. [28]

    M.; McClelland, J

    Saxe, A. M.; McClelland, J. L.; and Ganguli, S. 2014. Exact Solutions to the Nonlinear Dynamics of Learning in Deep Linear Neural Networks. In International Conference on Learning Representations (ICLR)

  18. [29]

    Snell, J.; Swersky, K.; and Zemel, R. 2017. Prototypical Networks for Few-shot Learning. In Advances in Neural Information Processing Systems, volume 30

  19. [30]

    Sun, H.; Zhang, Z.; Deng, J.; Cheng, J.; and Huang, M. 2023. Safety Assessment of Chinese Large Language Models. arXiv preprint arXiv:2304.10436

  20. [31]

    Vidgen, B.; Thrush, T.; Waseem, Z.; and Kiela, D. 2021. Learning from the Worst: Dynamically Generated Datasets to Improve Online Hate Detection. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics (ACL)

  21. [32]

    Wang, Y.; Zhang, S.; Zhou, Z.; Zhang, L.; Bai, J.; Lu, J.; Fan, B.; and Bu, F. 2026. DiaCOLQ : A Benchmark for C hinese Offensive Language Quadruple Prediction in Multi-Turn Dialogues. Information Processing & Management, 63(7): 104805

  22. [33]

    V.; and Zhou, D

    Wei, J.; Wang, X.; Schuurmans, D.; Bosma, M.; Ichter, B.; Xia, F.; Chi, E.; Le, Q. V.; and Zhou, D. 2022. Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. In Advances in Neural Information Processing Systems, volume 35, 24824--24837

  23. [35]

    Wulczyn, E.; Thain, N.; and Dixon, L. 2017. Ex Machina: Personal Attacks Seen at Scale. In Proceedings of the 26th International Conference on World Wide Web (WWW), 1391--1399

  24. [39]

    2025 , institution =

    2024. 2025 , institution =

  25. [40]

    Applied Soft Computing , volume=

    Countering Malicious Content Moderation Evasion in Online Social Networks: Simulation and Detection of Word Camouflage , author=. Applied Soft Computing , volume=. 2023 , publisher=

  26. [41]

    Toxicity in Online Platforms and

    Khapre, Smita and Mersha, Melkamu Abay and Shakil, Hassan and Baruah, Jonali and Kalita, Jugal , journal =. Toxicity in Online Platforms and. 2025 , doi =

  27. [42]

    2606.04867 , archiveprefix =

    Ren, Yanjing and Ebrahimi, Reza and Ma, Tengteng , year =. 2606.04867 , archiveprefix =

  28. [43]

    arXiv preprint arXiv:2302.09270 , year =

    Recent Advances towards Safe, Responsible, and Moral Dialogue Systems: A Survey , author =. arXiv preprint arXiv:2302.09270 , year =. 2302.09270 , archiveprefix =

  29. [44]

    Wang, Yifan and Zhang, Shun and Zhou, Ziyi and Zhang, Litian and Bai, Jiahao and Lu, Junyu and Fan, Bohan and Bu, Fanliang , journal =

  30. [45]

    Mathew, Binny and Saha, Punyajoy and Yimam, Seid Muhie and Biemann, Chris and Goyal, Pawan and Mukherjee, Animesh , booktitle =

  31. [46]

    Advances in Neural Information Processing Systems , volume =

    The Hateful Memes Challenge: Detecting Hate Speech in Multimodal Memes , author =. Advances in Neural Information Processing Systems , volume =

  32. [47]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages =

    All Changes May Have Invariant Principles: Improving Ever-Shifting Harmful Meme Detection via Design Concept Reproduction , author =. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages =

  33. [48]

    Breaking Free from Ivory Tower: Evaluating and Enhancing Real-world

    Jiang, Zhifan and Liu, Mingxuan and Qin, Yue and Liu, Baojun , booktitle =. Breaking Free from Ivory Tower: Evaluating and Enhancing Real-world. 2026 , doi =

  34. [49]

    2021 , note =

    The State of Online Harassment , author =. 2021 , note =

  35. [50]

    Journal of the European Economic Association , volume =

    Fanning the Flames of Hate: Social Media and Hate Crime , author =. Journal of the European Economic Association , volume =. 2021 , doi =

  36. [51]

    Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages =

    Chinese Toxic Language Mitigation via Sentiment Polarity Consistent Rewrites , author =. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages =

  37. [52]

    2505.17244 , archiveprefix =

    Li, Changyi and Wang, Jiayi and Pan, Xudong and Hong, Geng and Yang, Min , year =. 2505.17244 , archiveprefix =

  38. [53]

    Li, Jing-Jing and Pyatkin, Valentina and Kleiman-Weiner, Max and Jiang, Liwei and Dziri, Nouha and Collins, Anne and Schaich Borg, Jana and Sap, Maarten and Choi, Yejin and Levine, Sydney , booktitle =

  39. [54]

    Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages =

    Robust Adaptation of Large Multimodal Models for Retrieval Augmented Hateful Meme Detection , author =. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages =

  40. [55]

    , booktitle =

    Chan, Yik Siu and Yong, Zheng-Xin and Bach, Stephen H. , booktitle =. Can We Predict Alignment Before Models Finish Thinking?. 2025 , eprint =

  41. [56]

    2024 , eprint =

    Nandi, Palash and Sharma, Shivam and Chakraborty, Tanmoy , journal =. 2024 , eprint =

  42. [57]

    Qu, Wei and Chen, Cong and Lu, Wei and Wei, Yingying and Li, Tao , journal =

  43. [58]

    Advances in Neural Information Processing Systems , volume =

    Prototypical Networks for Few-shot Learning , author =. Advances in Neural Information Processing Systems , volume =

  44. [59]

    Advances in Neural Information Processing Systems , volume =

    Supervised Contrastive Learning , author =. Advances in Neural Information Processing Systems , volume =

  45. [60]

    Wang, Gang and Du, Yajun and Jiang, Yurui , journal =

  46. [61]

    Li, Jiahui and Yang, Yuan and Sun, Jian and Wang, Fen , journal =

  47. [62]

    Proceedings of the 31st International Conference on Computational Linguistics , pages =

    Dynamic-prototype Contrastive Fine-tuning for Continual Few-shot Relation Extraction with Unseen Relation Detection , author =. Proceedings of the 31st International Conference on Computational Linguistics , pages =

  48. [63]

    arXiv preprint arXiv:2605.13943 , year =

    A Unified Geometric Framework for Weighted Contrastive Learning , author =. arXiv preprint arXiv:2605.13943 , year =. 2605.13943 , archiveprefix =

  49. [64]

    Journal of Machine Learning Research , volume =

    Domain-Adversarial Training of Neural Networks , author =. Journal of Machine Learning Research , volume =

  50. [65]

    Guo, Zikang and Jiao, Kaijie and Yao, Xingyu and Wan, Yuning and Li, Haoran and Xu, Benfeng and Zhang, Licheng and Wang, Quan and Zhang, Yongdong and Mao, Zhendong , booktitle =

  51. [66]

    Journal of Machine Learning Research , volume =

    Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity , author =. Journal of Machine Learning Research , volume =

  52. [67]

    2406.12375 , archiveprefix =

    Wu, Haoze and Qiu, Zihan and Wang, Zili and Zhao, Hang and Fu, Jie , year =. 2406.12375 , archiveprefix =

  53. [68]

    Werkzeug at

    Wu, Youlin and Wang, Kaichun and Ma, Kai and Yang, Liang and Lin, Hongfei , booktitle =. Werkzeug at

  54. [69]

    Knowledge-Based Systems , year =

    Correlation-guided Mixture of Experts Prompt Learning for Long-tailed Multi-label Text Classification , author =. Knowledge-Based Systems , year =

  55. [70]

    Proceedings of the IEEE International Conference on Computer Vision (ICCV) , pages =

    Focal Loss for Dense Object Detection , author =. Proceedings of the IEEE International Conference on Computer Vision (ICCV) , pages =

  56. [71]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

    Momentum Contrast for Unsupervised Visual Representation Learning , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

  57. [72]

    Advances in Neural Information Processing Systems , volume =

    Chain-of-Thought Prompting Elicits Reasoning in Large Language Models , author =. Advances in Neural Information Processing Systems , volume =

  58. [73]

    International Conference on Learning Representations (ICLR) , year =

    Exact Solutions to the Nonlinear Dynamics of Learning in Deep Linear Neural Networks , author =. International Conference on Learning Representations (ICLR) , year =

  59. [74]

    2506.10960 , archiveprefix =

    Ma, Yiming and Zhang, Ningyu and Chen, Huajun , year =. 2506.10960 , archiveprefix =

  60. [75]

    2410.18491 , archiveprefix =

    Zhang, Hengxiang and Gao, Hongfu and Hu, Qiang and Chen, Guanhua and Yang, Lili and Jing, Bingyi and Wei, Hongxin and Wang, Bing and Bai, Haifeng and Yang, Lei , year =. 2410.18491 , archiveprefix =

  61. [76]

    Computers & Security , year =

    Developing Continuous Toxicity Detection Against Increasing Types of Perturbed Toxic Text , author =. Computers & Security , year =

  62. [77]

    Jin, Di and Jin, Zhijing and Zhou, Joey Tianyi and Szolovits, Peter , booktitle =. Is. 2020 , doi =

  63. [78]

    IEEE Transactions on Neural Networks and Learning Systems , year =

    Prototype-Guided Memory Replay for Continual Learning , author =. IEEE Transactions on Neural Networks and Learning Systems , year =

  64. [79]

    Prototype Conditioned Generative Replay for Continual Learning in

    Chen, Xi and Zeng, Min , booktitle =. Prototype Conditioned Generative Replay for Continual Learning in

  65. [80]

    Goyal, Agam and Zhan, Xianyang and Chen, Yilun and Saha, Koustuv and Chandrasekharan, Eshwar , booktitle =

  66. [81]

    Lai, Zhenglin and Liao, Mengyao and Wu, Bingzhe and others , booktitle =

  67. [82]

    Wang, Junjie and others , journal =

  68. [83]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Bootstrap Your Own Latent: A New Approach to Self-Supervised Learning , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =

  69. [84]

    Visualizing Data Using

    van der Maaten, Laurens and Hinton, Geoffrey , journal =. Visualizing Data Using

  70. [85]

    Organization Science , volume =

    Organizing and the Process of Sensemaking , author =. Organization Science , volume =

  71. [86]

    To Let Content Be or Not Be:

    Chatterjee, Anirban and Gupta, Samrat and Thomas, Princy , journal =. To Let Content Be or Not Be:. 2025 , note =

  72. [87]

    2011 , publisher =

    Thinking, Fast and Slow , author =. 2011 , publisher =

  73. [88]

    Davidson, Thomas , journal =

  74. [89]

    2023 , note =

    Ji, Jiaming and Liu, Mickel and Dai, Juntao and Pan, Xuehai and Zhang, Chi and Bian, Ce and Chen, Boyuan and Sun, Ruiyang and Wang, Yizhou and Yang, Yaodong , booktitle =. 2023 , note =

  75. [90]

    2025 , doi =

    Ji, Jiaming and Hong, Donghai and Zhang, Borong and Chen, Boyuan and Dai, Josef and Zheng, Boren and Qiu, Tianyi and Zhou, Jiayi and Wang, Kaile and Li, Boxun and Han, Sirui and Guo, Yike and Yang, Yaodong , booktitle =. 2025 , doi =

  76. [91]

    Safety Assessment of

    Sun, Hao and Zhang, Zhexin and Deng, Jiawen and Cheng, Jiale and Huang, Minlie , journal =. Safety Assessment of

  77. [92]

    2024 , doi =

    Qiu, Huachuan and Ma, Lizhi and Lan, Zhenzhong , booktitle =. 2024 , doi =

  78. [93]

    Facilitating Fine-grained Detection of

    Lu, Junyu and Xu, Bo and Zhang, Xiaokun and Min, Changrong and Yang, Liang and Lin, Hongfei , booktitle =. Facilitating Fine-grained Detection of. 2023 , doi =

  79. [94]

    2307.09705 , archiveprefix =

    Xu, Guohai and Liu, Jiayi and Yan, Ming and Xu, Haotian and Si, Jinghui and Zhou, Zhuoran and Yi, Peng and Gao, Xing and Sang, Jitao and Zhang, Rong and Zhang, Ji and Peng, Chao and Huang, Fei and Zhou, Jingren , year =. 2307.09705 , archiveprefix =

  80. [95]

    Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics (ACL) , year =

    Learning from the Worst: Dynamically Generated Datasets to Improve Online Hate Detection , author =. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics (ACL) , year =

  81. [96]

    Proceedings of the 11th International AAAI Conference on Web and Social Media (ICWSM) , pages =

    Automated Hate Speech Detection and the Problem of Offensive Language , author =. Proceedings of the 11th International AAAI Conference on Web and Social Media (ICWSM) , pages =

  82. [97]

    Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics (ACL) , pages =

    R. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics (ACL) , pages =. 2021 , doi =

  83. [98]

    Multilingual

    R. Multilingual. Proceedings of the Sixth Workshop on Online Abuse and Harms (WOAH) , pages =. 2022 , doi =

  84. [99]

    Proceedings of the 26th International Conference on World Wide Web (WWW) , pages =

    Ex Machina: Personal Attacks Seen at Scale , author =. Proceedings of the 26th International Conference on World Wide Web (WWW) , pages =. 2017 , doi =

  85. [100]

    Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING) , pages =

    M. Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING) , pages =

  86. [101]

    and Gonzalez, Joseph E

    Zheng, Lianmin and Chiang, Wei-Lin and Sheng, Ying and Li, Tianle and Zhuang, Siyuan and Wu, Zhanghao and Zhuang, Yonghao and Li, Zhuohan and Lin, Zi and Xing, Eric P. and Gonzalez, Joseph E. and Stoica, Ion and Zhang, Hao , year =. 2309.11998 , archiveprefix =

  87. [102]

    2311.00286 , archiveprefix =

    Zhang, Mi and Pan, Xudong and Yang, Min , year =. 2311.00286 , archiveprefix =

  88. [103]

    and Mihalcea, Rada , year =

    Lee, Andrew and Bai, Xiaoyan and Pres, Itamar and Wattenberg, Martin and Kummerfeld, Jonathan K. and Mihalcea, Rada , year =. A Mechanistic Understanding of Alignment Algorithms: A Case Study on. 2401.01967 , archiveprefix =

  89. [104]

    IEEE Transactions on Information Theory , volume =

    Successive Refinement of Information , author =. IEEE Transactions on Information Theory , volume =

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.