pith. sign in

arxiv: 2606.27981 · v1 · pith:GWMSS7S2new · submitted 2026-06-26 · 💻 cs.CL

ToxiREX: A Dataset on Toxic REasoning in ConteXt

Stefan F. Schouten , Ilia Markov , Piek Vossen This is my paper

Pith reviewed 2026-06-29 04:30 UTC · model grok-4.3

classification 💻 cs.CL
keywords toxicity detectioncontextual reasoningimplicit toxicitymultilingual datasetreddit threadsstructured annotationsconversational context
0
0 comments X

The pith

ToxiREX supplies a multilingual dataset of Reddit threads annotated for implicit toxicity through a structured reasoning schema.

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

The paper presents ToxiREX as a collection of comment threads drawn from event-related posts in six languages, each annotated with structured characterizations of implied toxicity. The annotations follow a schema that explains context-dependent and implicit toxicity while allowing mappings to existing taxonomies. A large portion of the data receives LLM labels for training and a smaller set receives native-speaker labels for testing, with baselines showing performance above random but with clear remaining challenges. This construction matters because prior toxicity resources typically omit conversational context and implicit meanings that affect real-world interpretation.

Core claim

ToxiREX is the first dataset to simultaneously incorporate multiple languages, conversational context, and implicit toxicity, while using the toxic reasoning schema for rich, structured annotations. The resource contains 125 thousand LLM-annotated comments for training and just under three thousand native-speaker-annotated comments for testing, drawn from threads linked to major events and preprocessed to preserve context. Models prompted or fine-tuned on the data exceed random baselines on hierarchical schema predictions, yet leave substantial room for improvement.

What carries the argument

The toxic reasoning schema that produces structured characterizations of what comments imply in their conversational context.

If this is right

  • Structured schema predictions allow models to output explanations of toxicity rather than binary flags alone.
  • Context-preserving thread preprocessing supports training that respects conversational dependencies.
  • Annotation disagreements in the test set frequently represent valid alternative readings rather than label noise.
  • Event-linked collection enables targeted study of toxicity patterns tied to specific real-world situations.

Where Pith is reading between the lines

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

  • Models trained on this data could support moderation systems that surface the implied reasoning behind flagged content.
  • The multilingual construction may help test whether schema-based labels transfer across languages with different cultural norms around toxicity.
  • Native-speaker validation of LLM labels points to a general method for auditing automated annotations on subjective tasks.

Load-bearing premise

The toxic reasoning schema accurately and without systematic bias captures what comments imply in context so that both LLM and native-speaker labels remain faithful.

What would settle it

A controlled comparison in which native speakers produce characterizations that diverge from the schema on the majority of test items in ways that cannot be explained as defensible alternative interpretations.

Figures

Figures reproduced from arXiv: 2606.27981 by Ilia Markov, Piek Vossen, Stefan F. Schouten.

Figure 4
Figure 4. Figure 4: Disagreements on earlier variables can cause predictions and annotations to disagree on [PITH_FULL_IMAGE:figures/full_fig_p019_4.png] view at source ↗
read the original abstract

We introduce a new, contextual, multilingual dataset called ToxiREX: Toxic REasoning in ConteXt. The dataset consists of threads of Reddit comments and structured characterizations of what the comments imply, following a systematic toxic reasoning schema developed in a previous paper. Using the schema allows us to capture and explain implicit and context-dependent toxicity, while supporting mappings to existing toxicity taxonomies. The dataset includes comments in six languages (English, Arabic, Turkish, Spanish, German, and Dutch), collected from posts connected to specific major events (e.g. the 2023 Turkey earthquakes; the Russian invasion of Ukraine). We describe the context-preserving preprocessing of the threads. We create a training set of 125 thousand comments which is annotated by a commercially available LLM, and a test set of just under three thousand comments that is annotated by native speakers. We show that apparent disagreements in the test set annotations often reflect defensible alternative interpretations rather than noise. Finally, we provide baseline results by prompting and fine-tuning language models. To produce these results, we develop evaluation strategies for our hierarchical, schema-based predictions. While models perform better than random, there remains a lot of room for improvement, showing the task to be challenging. ToxiREX is the first dataset to simultaneously incorporate multiple languages, conversational context, and implicit toxicity, while using the toxic reasoning schema for rich, structured annotations. Dataset available at: https://github.com/cltl/toxirex

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper introduces ToxiREX, a multilingual dataset of Reddit comment threads from event-related posts, annotated with a prior toxic reasoning schema to capture implicit and context-dependent toxicity across six languages. It includes a 125k-comment training split labeled by a commercial LLM and a ~3k-comment test split labeled by native speakers, along with preprocessing to preserve conversational context, an analysis of annotation disagreements as often defensible alternatives, and baseline results from prompting and fine-tuning models using custom evaluation strategies for the hierarchical schema-based predictions.

Significance. If the annotations hold, the dataset would provide a useful resource for research on nuanced, context-aware toxicity detection in multilingual conversational settings by supplying structured reasoning annotations that map to existing taxonomies, going beyond binary labels; the reported baselines establish that the task remains challenging and the disagreement analysis supports annotation quality.

major comments (2)
  1. [Dataset construction] Dataset construction section: the context-preserving preprocessing of threads is described at a high level without specific quantitative criteria (e.g., maximum thread length, selection rules for comments tied to events like the 2023 Turkey earthquakes), which is load-bearing for reproducibility of the 125k training set and for verifying that context is consistently retained.
  2. [Annotation process and evaluation] Annotation and evaluation sections: while the abstract notes that test-set disagreements often reflect defensible alternatives, the manuscript lacks reported inter-annotator agreement statistics or quantitative breakdown of disagreement types for the native-speaker test set, undermining assessment of whether the ~3k test annotations reliably support the baseline comparisons.
minor comments (2)
  1. [Introduction] The claim that ToxiREX is 'the first' to combine multiple languages, context, implicit toxicity, and the schema should include explicit comparison to prior datasets in a related-work section to strengthen the positioning.
  2. [Experiments] Baseline results would benefit from a table reporting exact metrics (e.g., precision/recall per schema category) rather than the high-level statement that models perform 'better than random' with 'room for improvement'.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive comments and positive overall assessment. We address each major comment point by point below.

read point-by-point responses

  1. Referee: [Dataset construction] Dataset construction section: the context-preserving preprocessing of threads is described at a high level without specific quantitative criteria (e.g., maximum thread length, selection rules for comments tied to events like the 2023 Turkey earthquakes), which is load-bearing for reproducibility of the 125k training set and for verifying that context is consistently retained.

    Authors: We agree that the current description is at a high level and that explicit quantitative criteria would strengthen reproducibility. In the revised manuscript we will add the specific rules used, including maximum thread length, minimum comments per thread, and the precise selection and filtering criteria applied to event-related posts. revision: yes

  2. Referee: [Annotation process and evaluation] Annotation and evaluation sections: while the abstract notes that test-set disagreements often reflect defensible alternatives, the manuscript lacks reported inter-annotator agreement statistics or quantitative breakdown of disagreement types for the native-speaker test set, undermining assessment of whether the ~3k test annotations reliably support the baseline comparisons.

    Authors: The manuscript already contains a qualitative analysis demonstrating that many disagreements represent defensible alternative interpretations. We acknowledge, however, that quantitative inter-annotator agreement statistics and a systematic breakdown of disagreement categories are not reported. We will incorporate these metrics and a quantitative summary in the revised version to the extent the annotation data permit. revision: partial

Circularity Check

0 steps flagged

No significant circularity; empirical dataset paper with no derivations

full rationale

The paper constructs and releases a new multilingual dataset of Reddit threads annotated according to a toxic-reasoning schema introduced in prior work. No equations, fitted parameters, predictions, or derivations appear anywhere in the described pipeline; the central contribution is the collection, annotation (LLM for training split, native speakers for test split), and baseline evaluation of the dataset itself. Self-citation of the schema exists but is not load-bearing for any claim that reduces to the authors' own outputs by construction. The work is therefore self-contained against external benchmarks and receives the default non-circularity finding.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The paper introduces no free parameters or invented entities. It applies an existing schema to new data and relies on one domain assumption about the schema's validity.

axioms (1)
  • domain assumption The toxic reasoning schema developed in a previous paper accurately captures implicit and context-dependent toxicity across languages and events.
    All annotations and evaluations rest on this schema; the abstract invokes it as the basis for structured characterizations.

pith-pipeline@v0.9.1-grok · 5798 in / 1333 out tokens · 50475 ms · 2026-06-29T04:30:03.787597+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

299 extracted references · 223 canonical work pages · 31 internal anchors

  1. [1]

    Mishra, Aayush and Khashabi, Daniel and Liu, Anqi , month = apr, year =. Steered

  2. [2]

    Asian Journal of Philosophy , author =

    Sapience without sentience: an inferentialist approach to. Asian Journal of Philosophy , author =. 2026 , keywords =. doi:10.1007/s44204-026-00400-4 , abstract =

    work page doi:10.1007/s44204-026-00400-4 2026

  3. [3]

    Søgaard, Anders , editor =. Do. Proceedings of the 63rd. 2025 , pages =. doi:10.18653/v1/2025.acl-long.1258 , abstract =

    work page doi:10.18653/v1/2025.acl-long.1258 2025

  4. [4]

    Anthropic Alignment Science Blog , author =

    The. Anthropic Alignment Science Blog , author =

  5. [5]

    Journal of Machine Learning Research , author =

    Causal. Journal of Machine Learning Research , author =. 2025 , pages =

    2025

  6. [6]

    Geiger, Atticus and Wu, Zhengxuan and Potts, Christopher and Icard, Thomas and Goodman, Noah , month = mar, year =. Finding. Proceedings of the

  7. [7]

    and Markov, Ilia and Vossen, Piek , editor =

    Schouten, Stefan F. and Markov, Ilia and Vossen, Piek , editor =. A. The. 2026 , pages =. doi:10.18653/v1/2026.wassa-1.12 , abstract =

    work page doi:10.18653/v1/2026.wassa-1.12 2026

  8. [8]

    Improving

    Nejadgholi, Isar and Fraser, Kathleen and Kiritchenko, Svetlana , editor =. Improving. Proceedings of the 60th. 2022 , pages =. doi:10.18653/v1/2022.acl-long.378 , abstract =

    work page doi:10.18653/v1/2022.acl-long.378 2022

  9. [9]

    Probing for

    Conia, Simone and Navigli, Roberto , editor =. Probing for. Proceedings of the 60th. 2022 , pages =. doi:10.18653/v1/2022.acl-long.316 , abstract =

    work page doi:10.18653/v1/2022.acl-long.316 2022

  10. [10]

    Factuality challenges in the era of large language models and opportunities for fact-checking , volume =

    Augenstein, Isabelle and Baldwin, Timothy and Cha, Meeyoung and Chakraborty, Tanmoy and Ciampaglia, Giovanni Luca and Corney, David and DiResta, Renee and Ferrara, Emilio and Hale, Scott and Halevy, Alon and Hovy, Eduard and Ji, Heng and Menczer, Filippo and Miguez, Ruben and Nakov, Preslav and Scheufele, Dietram and Sharma, Shivam and Zagni, Giovanni , m...

    work page doi:10.1038/s42256-024-00881-z

  11. [11]

    and Baker, Collin , editor =

    Heine, Bernd and Narrog, Heiko and Fillmore, Charles J. and Baker, Collin , editor =. A frames. The. doi:10.1093/oxfordhb/9780199677078.013.0013 , urldate =

    work page doi:10.1093/oxfordhb/9780199677078.013.0013

  12. [12]

    Ruppenhofer, Josef and Ellsworth, Michael and Petruck, Miriam R. L. and Johnson, Christopher R. and Scheffczyk, Jan , year =

  13. [13]

    Representation Engineering: A Top-Down Approach to AI Transparency

    Zou, Andy and Phan, Long and Chen, Sarah and Campbell, James and Guo, Phillip and Ren, Richard and Pan, Alexander and Yin, Xuwang and Mazeika, Mantas and Dombrowski, Ann-Kathrin and Goel, Shashwat and Li, Nathaniel and Byun, Michael J. and Wang, Zifan and Mallen, Alex and Basart, Steven and Koyejo, Sanmi and Song, Dawn and Fredrikson, Matt and Kolter, J. ...

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2310.01405

  14. [14]

    Zhao, Hongjue and Sun, Haosen and Kong, Jiangtao and Li, Xiaochang and Wang, Qineng and Jiang, Liwei and Zhu, Qi and Abdelzaher, Tarek and Choi, Yejin and Li, Manling and Shao, Huajie , month = feb, year =

  15. [15]

    Language

    Zhu, Wentao and Zhang, Zhining and Wang, Yizhou , month = jul, year =. Language. Proceedings of the 41st

  16. [16]

    Representational and

    Dies, Samantha and Maynard, Courtney and Savcisens, Germans and Eliassi-Rad, Tina , month = jan, year =. Representational and. doi:10.48550/arXiv.2511.19166 , abstract =

    work page doi:10.48550/arxiv.2511.19166

  17. [17]

    Philosophy and the Mind Sciences , author =

    Mental representation without neural representation:. Philosophy and the Mind Sciences , author =. 2026 , keywords =. doi:10.33735/phimisci.2026.12204 , abstract =

    work page doi:10.33735/phimisci.2026.12204 2026

  18. [18]

    Mind & Language , author =

    Do. Mind & Language , author =. 1997 , note =. doi:10.1111/j.1468-0017.1997.tb00061.x , abstract =

    work page doi:10.1111/j.1468-0017.1997.tb00061.x 1997

  19. [19]

    Acta Analytica , author =

    Rethinking. Acta Analytica , author =. 1995 , pages =

    1995

  20. [20]

    Representation reconsidered , language =

  21. [21]

    , year =

    Dretske, Fred I. , year =. Explaining behavior: reasons in a world of causes , isbn =

  22. [22]

    Makelov, Aleksandar and Lange, Georg and Geiger, Atticus and Nanda, Neel , year =. Is. 12th

  23. [23]

    International Conference on Learning Representations , author =

    Everything,. International Conference on Learning Representations , author =. 2025 , pages =

    2025

  24. [24]

    2023 , note =

    Is. 2023 , note =. doi:10.48550/arXiv.2311.17030 , abstract =

    work page doi:10.48550/arxiv.2311.17030 2023

  25. [25]

    , month = jan, year =

    Wu, Zhengxuan and Geiger, Atticus and Huang, Jing and Arora, Aryaman and Icard, Thomas and Potts, Christopher and Goodman, Noah D. , month = jan, year =. A. doi:10.48550/arXiv.2401.12631 , abstract =

    work page doi:10.48550/arxiv.2401.12631

  26. [26]

    Sakata, Masaki and Heinzerling, Benjamin and Ito, Takumi and Yokoi, Sho and Inui, Kentaro , month = apr, year =. Linear. doi:10.48550/arXiv.2604.07886 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2604.07886

  27. [27]

    Orgad, Hadas and Wei, Boyi and Zheng, Kaden and Wattenberg, Martin and Henderson, Peter and Goldfarb-Tarrant, Seraphina and Belinkov, Yonatan , month = apr, year =. Large. doi:10.48550/arXiv.2604.09544 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2604.09544

  28. [28]

    Arithmetic

    Nikankin, Yaniv and Reusch, Anja and Mueller, Aaron and Belinkov, Yonatan , month = may, year =. Arithmetic. doi:10.48550/arXiv.2410.21272 , abstract =

    work page doi:10.48550/arxiv.2410.21272

  29. [29]

    Understanding

    Nickel, Christian and Schrewe, Laura and Mai, Florian and Flek, Lucie , month = feb, year =. Understanding. doi:10.48550/arXiv.2602.22072 , abstract =

    work page doi:10.48550/arxiv.2602.22072

  30. [30]

    , month = feb, year =

    Wu, Zhengxuan and Geiger, Atticus and Icard, Thomas and Potts, Christopher and Goodman, Noah D. , month = feb, year =. Interpretability at. doi:10.48550/arXiv.2305.08809 , abstract =

    work page doi:10.48550/arxiv.2305.08809

  31. [31]

    doi:10.48550/arXiv.2503.10894 , abstract =

    Sun, Jiuding and Huang, Jing and Baskaran, Sidharth and D'Oosterlinck, Karel and Potts, Christopher and Sklar, Michael and Geiger, Atticus , month = apr, year =. doi:10.48550/arXiv.2503.10894 , abstract =

    work page doi:10.48550/arxiv.2503.10894

  32. [32]

    Lange, Georg and Makelov, Alex and Nanda, Neel , month = aug, year =. An

  33. [33]

    Sutter, Denis and Minder, Julian and Hofmann, Thomas and Pimentel, Tiago , year =. The. doi:10.48550/ARXIV.2507.08802 , abstract =

    work page doi:10.48550/arxiv.2507.08802

  34. [34]

    Behavioral and Brain Sciences , author =

    How. Behavioral and Brain Sciences , author =. 2025 , note =. doi:10.1017/S0140525X2510112X , abstract =

    work page doi:10.1017/s0140525x2510112x 2025

  35. [35]

    Stacey, Joe and Orgad, Hadas and Inui, Kentaro and Heinzerling, Benjamin and Moosavi, Nafise Sadat , month = apr, year =. Hidden. doi:10.48550/arXiv.2604.11662 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2604.11662

  36. [36]

    Futrell, Richard and Mahowald, Kyle , month = apr, year =. You. doi:10.1017/S0140525X26105123 , abstract =

    work page doi:10.1017/s0140525x26105123

  37. [37]

    Cheng, Stephen and Wiegreffe, Sarah and Manocha, Dinesh , month = apr, year =. What. doi:10.48550/arXiv.2604.08524 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2604.08524

  38. [38]

    Causality is

    Joshi, Shruti and Mueller, Aaron and Klindt, David and Brendel, Wieland and Reizinger, Patrik and Sridhar, Dhanya , month = mar, year =. Causality is. doi:10.48550/arXiv.2602.16698 , abstract =

    work page doi:10.48550/arxiv.2602.16698

  39. [39]

    Transactions on Machine Learning Research , author =

    Large. Transactions on Machine Learning Research , author =

  40. [40]

    Assessing

    Gubelmann, Reto and Karray, Ghassen , editor =. Assessing. Proceedings of the 63rd. 2025 , pages =. doi:10.18653/v1/2025.acl-long.1450 , abstract =

    work page doi:10.18653/v1/2025.acl-long.1450 2025

  41. [41]

    Position: Stop Anthropomorphizing Intermediate Tokens as Reasoning/Thinking Traces!

    Kambhampati, Subbarao and Valmeekam, Karthik and Bhambri, Siddhant and Palod, Vardhan and Saldyt, Lucas and Stechly, Kaya and Samineni, Soumya Rani and Kalwar, Durgesh and Biswas, Upasana , month = mar, year =. Position:. doi:10.48550/arXiv.2504.09762 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2504.09762

  42. [42]

    Adarsh, Shivam and Maistro, Maria and Lioma, Christina , month = jan, year =. How. doi:10.48550/arXiv.2601.06599 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2601.06599

  43. [43]

    Asian Journal of Philosophy , author =

    Sapience. Asian Journal of Philosophy , author =

  44. [45]

    Lyngbaek, Laurits and Feldkamp, Pascale and Bizzoni, Yuri and Nielbo, Kristoffer and Enevoldsen, Kenneth , editor =. Is. The. 2026 , pages =. doi:10.18653/v1/2026.wassa-1.13 , abstract =

    work page doi:10.18653/v1/2026.wassa-1.13 2026

  45. [46]

    Transactions of the Association for Computational Linguistics , author =

    Theoretical. Transactions of the Association for Computational Linguistics , author =. 2020 , pages =. doi:10.1162/tacl_a_00306 , abstract =

    work page doi:10.1162/tacl_a_00306 2020

  46. [47]

    Transactions on Machine Learning Research , author =

    A. Transactions on Machine Learning Research , author =

  47. [48]

    ACM Trans

    Reasoning. ACM Trans. Intell. Syst. Technol. , author =. 2025 , pages =. doi:10.1145/3712701 , abstract =

    work page doi:10.1145/3712701 2025

  48. [49]

    Kath, Suraj and Badhe, Sanket and Shah, Preet and Sampathkumar, Ashwin and Gupta, Shivani , month = mar, year =. Large. doi:10.48550/arXiv.2604.00323 , abstract =

    work page doi:10.48550/arxiv.2604.00323

  49. [50]

    Ying, Zhuofan Josh and Ravfogel, Shauli and Kriegeskorte, Nikolaus and Hase, Peter , month = feb, year =. The. doi:10.48550/arXiv.2602.20273 , abstract =

    work page doi:10.48550/arxiv.2602.20273

  50. [51]

    Representation in

    Shea, Nicholas , month = oct, year =. Representation in. doi:10.1093/oso/9780198812883.001.0001 , abstract =

    work page doi:10.1093/oso/9780198812883.001.0001

  51. [52]

    Meaning without reference in large language models , url =

    Piantadosi, Steven and Hill, Felix , month = oct, year =. Meaning without reference in large language models , url =

  52. [53]

    Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences , author =

    Symbols and grounding in large language models , volume =. Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences , author =. 2023 , pages =. doi:10.1098/rsta.2022.0041 , abstract =

    work page doi:10.1098/rsta.2022.0041 2023

  53. [54]

    Mollo, Dimitri Coelho and Millière, Raphaël , month = dec, year =. The. doi:10.48550/arXiv.2304.01481 , abstract =

    work page doi:10.48550/arxiv.2304.01481

  54. [55]

    Synthese , author =

    Semantics,. Synthese , author =. 1984 , pages =. doi:10.1007/BF00869335 , language =

    work page doi:10.1007/bf00869335 1984

  55. [56]

    Interpretability

    Orgad, Hadas and Barez, Fazl and Haklay, Tal and Lee, Isabelle and Mosbach, Marius and Reusch, Anja and Saphra, Naomi and Wallace, Byron C and Wiegreffe, Sarah and Wong, Eric and Tenney, Ian and Geva, Mor , month = feb, year =. Interpretability

  56. [57]

    Testing the Limits of Truth Directions in LLMs

    Poulis, Angelos and Crovella, Mark and Terzi, Evimaria , month = apr, year =. Testing the. doi:10.48550/arXiv.2604.03754 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2604.03754

  57. [58]

    Probing the

    Berger, Tom-Felix , month = feb, year =. Probing the. doi:10.48550/arXiv.2603.10003 , abstract =

    work page doi:10.48550/arxiv.2603.10003

  58. [59]

    De Giorgis, Stefano and Chen, Ting-Chih and Ilievski, Filip , month = feb, year =. M-

  59. [60]

    Sparks in the Wind , author =

    How. Sparks in the Wind , author =

  60. [61]

    Proceedings of the 62nd

    Huang, Jing and Wu, Zhengxuan and Potts, Christopher and Geva, Mor and Geiger, Atticus , editor =. Proceedings of the 62nd. 2024 , pages =. doi:10.18653/v1/2024.acl-long.470 , abstract =

    work page doi:10.18653/v1/2024.acl-long.470 2024

  61. [62]

    Mueller, Aaron and Brinkmann, Jannik and Li, Millicent and Marks, Samuel and Pal, Koyena and Prakash, Nikhil and Rager, Can and Sankaranarayanan, Aruna and Sharma, Arnab Sen and Sun, Jiuding and Todd, Eric and Bau, David and Belinkov, Yonatan , month = sep, year =. The. doi:10.48550/arXiv.2408.01416 , abstract =

    work page doi:10.48550/arxiv.2408.01416

  62. [63]

    Campbell Systematic Reviews , author =

    Counter-narratives for the prevention of violent radicalisation:. Campbell Systematic Reviews , author =. 2020 , note =. doi:10.1002/cl2.1106 , abstract =

    work page doi:10.1002/cl2.1106 2020

  63. [64]

    Recognizing

    Gao, Lei and Kuppersmith, Alexis and Huang, Ruihong , editor =. Recognizing. Proceedings of the. 2017 , pages =

    2017

  64. [65]

    2022 , pages =

    Creating a. 2022 , pages =. doi:10.1017/9781108641104.019 , urldate =

    work page doi:10.1017/9781108641104.019 2022

  65. [66]

    Kim, Junsol and Evans, James and Schein, Aaron , year =. Linear. 13th

  66. [67]

    and Macar, Uzay and Nanda, Neel and Conmy, Arthur , month = oct, year =

    Bogdan, Paul C. and Macar, Uzay and Nanda, Neel and Conmy, Arthur , month = oct, year =. Thought. doi:10.48550/arXiv.2506.19143 , abstract =

    work page doi:10.48550/arxiv.2506.19143

  67. [68]

    Hlobil, Ulf , month = feb, year =. First-. doi:10.48550/arXiv.2602.13908 , abstract =

    work page doi:10.48550/arxiv.2602.13908

  68. [69]

    Siren's Song in the AI Ocean: A Survey on Hallucination in Large Language Models

    🧜. Computational Linguistics , author =. 2025 , pages =. doi:10.1162/COLI.a.16 , abstract =

    work page doi:10.1162/coli.a.16 2025

  69. [70]

    Cognitive Systems Research , author =

    Does. Cognitive Systems Research , author =. 2024 , keywords =. doi:10.1016/j.cogsys.2023.101174 , abstract =

    work page doi:10.1016/j.cogsys.2023.101174 2024

  70. [71]

    2018 , keywords =

    Non-propositional intentionality , isbn =. 2018 , keywords =

    2018

  71. [72]

    McGrath, Thomas and Rahtz, Matthew and Kramar, Janos and Mikulik, Vladimir and Legg, Shane , month = jul, year =. The. doi:10.48550/arXiv.2307.15771 , abstract =

    work page doi:10.48550/arxiv.2307.15771

  72. [73]

    Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 small

    Wang, Kevin and Variengien, Alexandre and Conmy, Arthur and Shlegeris, Buck and Steinhardt, Jacob , month = nov, year =. Interpretability in the. doi:10.48550/arXiv.2211.00593 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2211.00593

  73. [74]

    Synthese , author =

    Connectionism and. Synthese , author =. 1990 , keywords =. doi:10.1007/BF00413661 , abstract =

    work page doi:10.1007/bf00413661 1990

  74. [75]

    OpenAI and Agarwal, Sandhini and Ahmad, Lama and Ai, Jason and Altman, Sam and Applebaum, Andy and Arbus, Edwin and Arora, Rahul K. and Bai, Yu and Baker, Bowen and Bao, Haiming and Barak, Boaz and Bennett, Ally and Bertao, Tyler and Brett, Nivedita and Brevdo, Eugene and Brockman, Greg and Bubeck, Sebastien and Chang, Che and Chen, Kai and Chen, Mark and...

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2508.10925

  75. [76]

    Gemma 3 Technical Report

    Team, Gemma and Kamath, Aishwarya and Ferret, Johan and Pathak, Shreya and Vieillard, Nino and Merhej, Ramona and Perrin, Sarah and Matejovicova, Tatiana and Ramé, Alexandre and Rivière, Morgane and Rouillard, Louis and Mesnard, Thomas and Cideron, Geoffrey and Grill, Jean-bastien and Ramos, Sabela and Yvinec, Edouard and Casbon, Michelle and Pot, Etienne...

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2503.19786

  76. [77]

    doi:10.1017/9781108641104 , abstract =

    Creating a. doi:10.1017/9781108641104 , abstract =

    work page doi:10.1017/9781108641104

  77. [78]

    Linear representations in language models can change dramatically over a conversation , url =

    Lampinen, Andrew Kyle and Li, Yuxuan and Hosseini, Eghbal and Bhardwaj, Sangnie and Shanahan, Murray , month = feb, year =. Linear representations in language models can change dramatically over a conversation , url =. doi:10.48550/arXiv.2601.20834 , abstract =

    work page doi:10.48550/arxiv.2601.20834

  78. [79]

    Deep sequence models tend to memorize geometrically; it is unclear why

    Noroozizadeh, Shahriar and Nagarajan, Vaishnavh and Rosenfeld, Elan and Kumar, Sanjiv , month = dec, year =. Deep sequence models tend to memorize geometrically; it is unclear why , url =. doi:10.48550/arXiv.2510.26745 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2510.26745

  79. [80]

    and Potts, Christopher , month = jun, year =

    Wu, Zhengxuan and Arora, Aryaman and Geiger, Atticus and Wang, Zheng and Huang, Jing and Jurafsky, Dan and Manning, Christopher D. and Potts, Christopher , month = jun, year =

  80. [81]

    and Nava, Andres and Wyart, Matthieu and Bahri, Yasaman , month = feb, year =

    Karkada, Dhruva and Korchinski, Daniel J. and Nava, Andres and Wyart, Matthieu and Bahri, Yasaman , month = feb, year =. Symmetry in language statistics shapes the geometry of model representations , url =. doi:10.48550/arXiv.2602.15029 , abstract =

    work page doi:10.48550/arxiv.2602.15029

Showing first 80 references.