pith. sign in

arxiv: 2606.09881 · v1 · pith:YG6YWQSXnew · submitted 2026-06-03 · 💻 cs.LG · cs.CR· cs.CV

Toward Calibrated, Fair, and accurate Deepfake Detection

Ryan Brown , Chris Russell This is my paper

Pith reviewed 2026-06-28 07:03 UTC · model grok-4.3

classification 💻 cs.LG cs.CRcs.CV
keywords deepfake detectionfairnessbias mitigationface embeddingslogit calibrationdemographic biasplug-and-play method
0
0 comments X

The pith

A lightweight calibrator remaps logits using frozen face embeddings to reduce demographic gaps in deepfake detectors without labels or retraining.

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

The paper introduces the Face-Fairness framework to close performance gaps that deepfake detectors show across demographic groups. Its main method, Face-Feature Tuning, adjusts detector outputs through a logit remapping that draws on frozen face embeddings alone. The approach avoids the need for demographic labels, detector retraining, or accuracy trade-offs that limit existing fairness techniques. Readers would care because deepfake detectors are already deployed in high-stakes settings where unequal error rates across groups create practical harms. The paper shows the method works across in-domain and cross-dataset tests while remaining detector-agnostic and low-overhead.

Core claim

Face-Feature Tuning performs a logit remapping conditioned on frozen face embeddings to achieve fairness in deepfake detection without demographic labels; when paired with FF-Max (which uses available demographics) and FF-Discover (which uses embedding-discovered groups), the framework narrows FPR and TPR gaps, raises minimum group accuracy, and preserves or improves overall accuracy in both in-domain and cross-dataset evaluations.

What carries the argument

Face-Feature Tuning (FFT), a lightweight calibrator that performs logit remapping conditioned on frozen face embeddings.

If this is right

  • FPR and TPR gaps narrow across demographic groups in both in-domain and cross-dataset tests.
  • Minimum group accuracy rises while overall accuracy stays the same or improves.
  • The method applies to any existing detector without retraining or access to identity attributes.
  • Runtime overhead remains negligible because only a small calibrator is added at inference time.

Where Pith is reading between the lines

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

  • Similar embedding-conditioned remapping could transfer to other vision tasks that already rely on face or object embeddings for downstream decisions.
  • If the implicit signal in embeddings proves stable across datasets, the same calibrator might serve as a reusable fairness module for multiple detectors.
  • Extending the approach to non-face modalities would require checking whether other frozen embeddings carry analogous group signals.

Load-bearing premise

Frozen face embeddings contain enough implicit demographic signal to support effective logit remapping for fairness without labels or retraining.

What would settle it

Replacing the frozen face embeddings with random vectors or embeddings whose demographic correlations have been removed and then measuring whether the fairness gains disappear.

Figures

Figures reproduced from arXiv: 2606.09881 by Chris Russell, Ryan Brown.

Figure 1
Figure 1. Figure 1: OpenForensics fairness–utility comparison [Le et al., 2021]. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Calibration curves on the OpenForensics dataset [Le et al., 2021] on Xception base model. FFT achieves substantial ECE reduction compared to the baseline (73% overall, 65% for females, 76% for males), while existing post-processing methods Fair-FLIP and BPFA degrade calibration. 3.2.1 Intuition FFT is a conditional calibrator. However, instead of learning a univariate map Z(x) 7→ pˆ(x) which assumes p(y=1 … view at source ↗
read the original abstract

Deepfake detectors show large performance gaps across demographic groups. Existing fairness approaches require demographic labels, retraining, or sacrifice accuracy. We introduce Face-Fairness (FF), a plug-and-play framework for bias mitigation. Our primary contribution, Face-Feature Tuning (FFT), is the first demographic label-free fairness method demonstrated for deepfake detection: a lightweight calibrator that performs a logit remapping conditioned on frozen face embeddings. We complement FFT with two variants: FF-Max, which maximizes worst-group accuracy when demographics are available, and FF-Discover, which does the same with embedding-discovered groups. Across in-domain and cross-dataset test settings, FF consistently reduces FPR/TPR gaps and improves minimum group accuracy while maintaining (often improving) overall accuracy. The approach is detector-agnostic, adds negligible runtime overhead, and requires no access to identity attributes.

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

1 major / 0 minor

Summary. The manuscript introduces the Face-Fairness (FF) framework for mitigating demographic biases in deepfake detectors. Its primary contribution is Face-Feature Tuning (FFT), presented as the first demographic label-free fairness method: a lightweight calibrator that remaps logits conditioned on frozen face embeddings from a separate model. Two variants are also described—FF-Max (maximizing worst-group accuracy when demographics are available) and FF-Discover (using embedding-discovered groups). The paper claims that FF reduces FPR/TPR gaps and improves minimum group accuracy across in-domain and cross-dataset settings while maintaining or improving overall accuracy, is detector-agnostic, adds negligible overhead, and requires no identity attributes or retraining.

Significance. If the empirical results hold and the core mechanism is validated, the work would provide a practical plug-and-play post-hoc approach to a documented problem in deepfake detection. The label-free nature via implicit signals in frozen embeddings, if shown to drive the fairness gains, would differentiate it from methods needing labels or retraining. The detector-agnostic design and low overhead are additional practical strengths.

major comments (1)
  1. [Abstract] Abstract: The primary claim that FFT delivers label-free fairness via logit remapping conditioned on frozen face embeddings rests on the unverified assumption that these embeddings encode usable demographic signal at a level sufficient to reduce group disparities. No correlation analysis, mutual-information quantification, ablation on demographic-predictive dimensions, or comparison against accuracy-only baselines is referenced to establish that the remapping mechanism is fairness-specific rather than generic calibration.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the detailed and constructive comment on the abstract. We address the concern point-by-point below.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The primary claim that FFT delivers label-free fairness via logit remapping conditioned on frozen face embeddings rests on the unverified assumption that these embeddings encode usable demographic signal at a level sufficient to reduce group disparities. No correlation analysis, mutual-information quantification, ablation on demographic-predictive dimensions, or comparison against accuracy-only baselines is referenced to establish that the remapping mechanism is fairness-specific rather than generic calibration.

    Authors: We agree that the abstract does not reference explicit analyses (correlation, mutual information, dimension ablations, or generic-calibration baselines) to isolate the demographic signal in the frozen embeddings. The manuscript's empirical results show consistent reductions in FPR/TPR gaps and gains in minimum-group accuracy across in-domain and cross-dataset settings while preserving overall accuracy, and the method is positioned as label-free because it conditions only on embeddings without using demographic labels at inference or training time. However, the referee's point is well-taken: these results alone do not rigorously demonstrate that the fairness effect is driven by demographic information encoded in the embeddings rather than generic logit recalibration. In the revised manuscript we will add (i) correlation and mutual-information measurements between embedding dimensions and available demographic attributes on the training sets, (ii) an ablation that masks or perturbs the most demographic-predictive dimensions, and (iii) a direct comparison against accuracy-only calibration baselines that use the same embedding conditioning but optimize only for overall accuracy. These additions will be placed in a new subsection of the experiments and referenced from the abstract. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper proposes a new plug-and-play framework (Face-Fairness) and its primary method (Face-Feature Tuning) as an empirical calibrator that remaps logits using frozen face embeddings, without any derivation chain, equations, or first-principles results that reduce outputs to inputs by construction. No self-definitional mappings, fitted parameters renamed as predictions, or load-bearing self-citations appear in the abstract or described contributions. Claims of label-free fairness and performance improvements are presented as experimental outcomes across in-domain and cross-dataset settings rather than tautological redefinitions, making the work self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

Abstract-only review limits visibility into parameters and assumptions; the new method introduces an entity whose effectiveness rests on an untested domain assumption about embeddings.

axioms (1)
  • domain assumption Frozen face embeddings contain sufficient implicit demographic signal for logit remapping to improve group fairness
    Central to the FFT description in the abstract.
invented entities (1)
  • Face-Feature Tuning (FFT) no independent evidence
    purpose: Lightweight calibrator performing logit remapping conditioned on frozen face embeddings without demographic labels
    Presented as the primary novel contribution.

pith-pipeline@v0.9.1-grok · 5673 in / 1215 out tokens · 19132 ms · 2026-06-28T07:03:01.938169+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

300 extracted references · 240 canonical work pages · 76 internal anchors

  1. [1]

    Singh, Harvineet and Kleindessner, Matthäus and Cevher, Volkan and Chunara, Rumi and Russell, Chris , month = jul, year =. When do. Proceedings of the 40th

  2. [2]

    and Liu, Xiaoming , month = feb, year =

    Kim, Minchul and Jain, Anil K. and Liu, Xiaoming , month = feb, year =. doi:10.48550/arXiv.2204.00964 , abstract =

    work page doi:10.48550/arxiv.2204.00964

  3. [3]

    doi:10.48550/arXiv.2103.06627 , abstract =

    Meng, Qiang and Zhao, Shichao and Huang, Zhida and Zhou, Feng , month = jul, year =. doi:10.48550/arXiv.2103.06627 , abstract =

    work page doi:10.48550/arxiv.2103.06627

  4. [4]

    Al-Refai, Rouqaiah and Hempel, Philipp and Biagi, Clara and Terhörst, Philipp , month = feb, year =. 2025. doi:10.1109/WACV61041.2025.00337 , urldate =

    work page doi:10.1109/wacv61041.2025.00337 2025

  5. [5]

    doi:10.48550/arXiv.1908.04913 , abstract =

    Kärkkäinen, Kimmo and Joo, Jungseock , month = aug, year =. doi:10.48550/arXiv.1908.04913 , abstract =

    work page doi:10.48550/arxiv.1908.04913 1908

  6. [6]

    Huang, Gary B and Ramesh, Manu and Berg, Tamara and Learned-Miller, Erik , month = oct, year =. Labeled

  7. [7]

    arXiv.org , author =

    Key. arXiv.org , author =. 2017 , doi =

    2017

  8. [8]

    arXiv.org , author =

    Toward. arXiv.org , author =

  9. [9]

    arXiv.org , author =

    Do. arXiv.org , author =

  10. [10]

    Remote Sensing , author =

    Unmanned. Remote Sensing , author =. 2023 , pages =. doi:10.3390/rs15133266 , abstract =

    work page doi:10.3390/rs15133266 2023

  11. [11]

    arXiv.org , author =

    Simple. arXiv.org , author =

  12. [12]

    arXiv.org , author =

  13. [13]

    arXiv.org , author =

    You. arXiv.org , author =

  14. [14]

    2023 , pages =

    Drones , author =. 2023 , pages =. doi:10.3390/drones7030190 , abstract =

    work page doi:10.3390/drones7030190 2023

  15. [15]

    Mueller, Matthias and Smith, Neil and Ghanem, Bernard , editor =. A. Computer. 2016 , doi =

    2016

  16. [16]

    ayushdabra/drone-images-semantic-segmentation , copyright =

    Dabra, Ayush , month = apr, year =. ayushdabra/drone-images-semantic-segmentation , copyright =

  17. [17]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , author =

    Detection and. IEEE Transactions on Pattern Analysis and Machine Intelligence , author =. 2022 , keywords =. doi:10.1109/TPAMI.2021.3119563 , abstract =

    work page doi:10.1109/tpami.2021.3119563 2022

  18. [18]

    arXiv.org , author =

    Segment. arXiv.org , author =

  19. [19]

    interpreting

    nostalgebraist , month = aug, year =. interpreting

  20. [20]

    Engineering News-Record , author =

    Construction's. Engineering News-Record , author =

  21. [21]

    For Construction Pros / Asphalt Contractor , author =

    2026. For Construction Pros / Asphalt Contractor , author =

    2026

  22. [22]

    Learning with

    Natarajan, Nagarajan and Dhillon, Inderjit S and Ravikumar, Pradeep K and Tewari, Ambuj , year =. Learning with. Advances in

  23. [23]

    Inference-Time Intervention: Eliciting Truthful Answers from a Language Model

    Li, Kenneth and Patel, Oam and Viégas, Fernanda and Pfister, Hanspeter and Wattenberg, Martin , month = jun, year =. Inference-. doi:10.48550/arXiv.2306.03341 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2306.03341

  24. [24]

    Eliciting Latent Predictions from Transformers with the Tuned Lens

    Belrose, Nora and Ostrovsky, Igor and McKinney, Lev and Furman, Zach and Smith, Logan and Halawi, Danny and Biderman, Stella and Steinhardt, Jacob , month = nov, year =. Eliciting. doi:10.48550/arXiv.2303.08112 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2303.08112

  25. [25]

    Stolfo, Alessandro and Jin, Zhijing and Shridhar, Kumar and Schölkopf, Bernhard and Sachan, Mrinmaya , month = jun, year =. A. doi:10.48550/arXiv.2210.12023 , abstract =

    work page doi:10.48550/arxiv.2210.12023

  26. [26]

    Locating and Editing Factual Associations in GPT

    Meng, Kevin and Bau, David and Andonian, Alex and Belinkov, Yonatan , month = jan, year =. Locating and. doi:10.48550/arXiv.2202.05262 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2202.05262

  27. [27]

    Designing and

    Hewitt, John and Liang, Percy , month = sep, year =. Designing and. doi:10.48550/arXiv.1909.03368 , abstract =

    work page doi:10.48550/arxiv.1909.03368 1909

  28. [28]

    Language Models as Knowledge Bases?

    Petroni, Fabio and Rocktäschel, Tim and Lewis, Patrick and Bakhtin, Anton and Wu, Yuxiang and Miller, Alexander H. and Riedel, Sebastian , month = sep, year =. Language. doi:10.48550/arXiv.1909.01066 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1909.01066 1909

  29. [29]

    doi:10.48550/arXiv.1905.05950 , abstract =

    Tenney, Ian and Das, Dipanjan and Pavlick, Ellie , month = aug, year =. doi:10.48550/arXiv.1905.05950 , abstract =

    work page doi:10.48550/arxiv.1905.05950 1905

  30. [30]

    , editor=

    Hewitt, John and Manning, Christopher D. , editor =. A. Proceedings of the 2019. 2019 , pages =. doi:10.18653/v1/N19-1419 , abstract =

    work page doi:10.18653/v1/n19-1419 2019

  31. [31]

    Analysis Methods in Neural Language Processing: A Survey

    Belinkov, Yonatan and Glass, James , month = jan, year =. Analysis. doi:10.48550/arXiv.1812.08951 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1812.08951

  32. [32]

    Ho, Namgyu and Schmid, Laura and Yun, Se-Young , month = jun, year =. Large. doi:10.48550/arXiv.2212.10071 , abstract =

    work page doi:10.48550/arxiv.2212.10071

  33. [33]

    Implicit

    Deng, Yuntian and Prasad, Kiran and Fernandez, Roland and Smolensky, Paul and Chaudhary, Vishrav and Shieber, Stuart , month = nov, year =. Implicit. doi:10.48550/arXiv.2311.01460 , abstract =

    work page doi:10.48550/arxiv.2311.01460

  34. [34]

    Distilling

    Shridhar, Kumar and Stolfo, Alessandro and Sachan, Mrinmaya , month = may, year =. Distilling. doi:10.48550/arXiv.2212.00193 , abstract =

    work page doi:10.48550/arxiv.2212.00193

  35. [35]

    Teaching

    Magister, Lucie Charlotte and Mallinson, Jonathan and Adamek, Jakub and Malmi, Eric and Severyn, Aliaksei , month = jun, year =. Teaching. doi:10.48550/arXiv.2212.08410 , abstract =

    work page doi:10.48550/arxiv.2212.08410

  36. [36]

    Chain-of-Thought Prompting Elicits Reasoning in Large Language Models

    Wei, Jason and Wang, Xuezhi and Schuurmans, Dale and Bosma, Maarten and Ichter, Brian and Xia, Fei and Chi, Ed and Le, Quoc and Zhou, Denny , month = jan, year =. Chain-of-. doi:10.48550/arXiv.2201.11903 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2201.11903

  37. [37]

    Distilling Step-by-Step! Outperforming Larger Language Models with Less Training Data and Smaller Model Sizes

    Hsieh, Cheng-Yu and Li, Chun-Liang and Yeh, Chih-Kuan and Nakhost, Hootan and Fujii, Yasuhisa and Ratner, Alexander and Krishna, Ranjay and Lee, Chen-Yu and Pfister, Tomas , month = jul, year =. Distilling. doi:10.48550/arXiv.2305.02301 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2305.02301

  38. [38]

    MiniLLM: On-Policy Distillation of Large Language Models

    Gu, Yuxian and Dong, Li and Wei, Furu and Huang, Minlie , month = nov, year =. doi:10.48550/arXiv.2306.08543 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2306.08543

  39. [39]

    doi:10.48550/arXiv.2002.10957 , abstract =

    Wang, Wenhui and Wei, Furu and Dong, Li and Bao, Hangbo and Yang, Nan and Zhou, Ming , month = apr, year =. doi:10.48550/arXiv.2002.10957 , abstract =

    work page doi:10.48550/arxiv.2002.10957 2002

  40. [40]

    Contrastive

    Tian, Yonglong and Krishnan, Dilip and Isola, Phillip , month = jan, year =. Contrastive. doi:10.48550/arXiv.1910.10699 , abstract =

    work page doi:10.48550/arxiv.1910.10699 1910

  41. [41]

    Relational Knowledge Distillation

    Park, Wonpyo and Kim, Dongju and Lu, Yan and Cho, Minsu , month = may, year =. Relational. doi:10.48550/arXiv.1904.05068 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1904.05068 1904

  42. [42]

    Zagoruyko, Sergey and Komodakis, Nikos , month = feb, year =. Paying. doi:10.48550/arXiv.1612.03928 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1612.03928

  43. [43]

    FitNets: Hints for Thin Deep Nets

    Romero, Adriana and Ballas, Nicolas and Kahou, Samira Ebrahimi and Chassang, Antoine and Gatta, Carlo and Bengio, Yoshua , month = mar, year =. doi:10.48550/arXiv.1412.6550 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1412.6550

  44. [44]

    Model compression , isbn =

    Buciluǎ, Cristian and Caruana, Rich and Niculescu-Mizil, Alexandru , month = aug, year =. Model compression , isbn =. Proceedings of the 12th. doi:10.1145/1150402.1150464 , language =

    work page doi:10.1145/1150402.1150464

  45. [45]

    Rethinking

    Liu, Chang and Tao, Chongyang and Liang, Jianxin and Shen, Tao and Feng, Jiazhan and Huang, Quzhe and Zhao, Dongyan , editor =. Rethinking. Proceedings of the 2022. 2022 , pages =. doi:10.18653/v1/2022.emnlp-main.729 , abstract =

    work page doi:10.18653/v1/2022.emnlp-main.729 2022

  46. [46]

    Rethinking the Inception Architecture for Computer Vision

    Szegedy, Christian and Vanhoucke, Vincent and Ioffe, Sergey and Shlens, Jonathon and Wojna, Zbigniew , month = dec, year =. Rethinking the. doi:10.48550/arXiv.1512.00567 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1512.00567

  47. [47]

    Rethinking

    Zhou, Helong and Song, Liangchen and Chen, Jiajie and Zhou, Ye and Wang, Guoli and Yuan, Junsong and Zhang, Qian , month = feb, year =. Rethinking. doi:10.48550/arXiv.2102.00650 , abstract =

    work page doi:10.48550/arxiv.2102.00650

  48. [48]

    , month = dec, year =

    Mandal, Saptarshi and Lin, Xiaojun and Srikant, R. , month = dec, year =. A. doi:10.48550/arXiv.2412.09579 , abstract =

    work page doi:10.48550/arxiv.2412.09579

  49. [49]

    Yuan, Li and Tay, Francis E. H. and Li, Guilin and Wang, Tao and Feng, Jiashi , month = mar, year =. Revisiting. doi:10.48550/arXiv.1909.11723 , abstract =

    work page doi:10.48550/arxiv.1909.11723 1909

  50. [50]

    Müller, Rafael and Kornblith, Simon and Hinton, Geoffrey , month = jun, year =. When. doi:10.48550/arXiv.1906.02629 , abstract =

    work page doi:10.48550/arxiv.1906.02629 1906

  51. [51]

    Understanding intermediate layers using linear classifier probes

    Alain, Guillaume and Bengio, Yoshua , month = nov, year =. Understanding intermediate layers using linear classifier probes , url =. doi:10.48550/arXiv.1610.01644 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1610.01644

  52. [52]

    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

  53. [53]

    Azaria, Amos and Mitchell, Tom , month = oct, year =. The. doi:10.48550/arXiv.2304.13734 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2304.13734

  54. [54]

    Sun, Siqi and Cheng, Yu and Gan, Zhe and Liu, Jingjing , month = aug, year =. Patient. doi:10.48550/arXiv.1908.09355 , abstract =

    work page doi:10.48550/arxiv.1908.09355 1908

  55. [55]

    doi:10.48550/arXiv.2501.17994 , abstract =

    Chételat, Didier and Cotnareanu, Joseph and Thompson, Rylee and Zhang, Yingxue and Coates, Mark , month = jan, year =. doi:10.48550/arXiv.2501.17994 , abstract =

    work page doi:10.48550/arxiv.2501.17994

  56. [56]

    DeBERTaV3: Improving DeBERTa using ELECTRA-Style Pre-Training with Gradient-Disentangled Embedding Sharing

    He, Pengcheng and Gao, Jianfeng and Chen, Weizhu , month = mar, year =. doi:10.48550/arXiv.2111.09543 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2111.09543

  57. [57]

    Qwen2.5 Technical Report

    Qwen2.5. 2025 , note =. doi:10.48550/arXiv.2412.15115 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2412.15115 2025

  58. [58]

    Measuring Massive Multitask Language Understanding

    Hendrycks, Dan and Burns, Collin and Basart, Steven and Zou, Andy and Mazeika, Mantas and Song, Dawn and Steinhardt, Jacob , month = jan, year =. Measuring. doi:10.48550/arXiv.2009.03300 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2009.03300 2009

  59. [59]

    Think you have Solved Question Answering? Try ARC, the AI2 Reasoning Challenge

    Clark, Peter and Cowhey, Isaac and Etzioni, Oren and Khot, Tushar and Sabharwal, Ashish and Schoenick, Carissa and Tafjord, Oyvind , month = mar, year =. Think you have. doi:10.48550/arXiv.1803.05457 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1803.05457

  60. [60]

    Ling, Wang and Yogatama, Dani and Dyer, Chris and Blunsom, Phil , month = oct, year =. Program. doi:10.48550/arXiv.1705.04146 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1705.04146

  61. [61]

    Discovering Latent Knowledge in Language Models Without Supervision

    Burns, Collin and Ye, Haotian and Klein, Dan and Steinhardt, Jacob , month = mar, year =. Discovering. doi:10.48550/arXiv.2212.03827 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2212.03827

  62. [62]

    Stolfo, Alessandro and Belinkov, Yonatan and Sachan, Mrinmaya , month = oct, year =. A. doi:10.48550/arXiv.2305.15054 , abstract =

    work page doi:10.48550/arxiv.2305.15054

  63. [63]

    Belinkov, Yonatan and Durrani, Nadir and Dalvi, Fahim and Sajjad, Hassan and Glass, James , month = oct, year =. What do. doi:10.48550/arXiv.1704.03471 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1704.03471

  64. [64]

    Specializing

    Fu, Yao and Peng, Hao and Ou, Litu and Sabharwal, Ashish and Khot, Tushar , month = jan, year =. Specializing. doi:10.48550/arXiv.2301.12726 , abstract =

    work page doi:10.48550/arxiv.2301.12726

  65. [65]

    doi:10.48550/arXiv.1909.10351 , abstract =

    Jiao, Xiaoqi and Yin, Yichun and Shang, Lifeng and Jiang, Xin and Chen, Xiao and Li, Linlin and Wang, Fang and Liu, Qun , month = oct, year =. doi:10.48550/arXiv.1909.10351 , abstract =

    work page doi:10.48550/arxiv.1909.10351 1909

  66. [66]

    DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter

    Sanh, Victor and Debut, Lysandre and Chaumond, Julien and Wolf, Thomas , month = mar, year =. doi:10.48550/arXiv.1910.01108 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1910.01108 1910

  67. [67]

    Distilling the Knowledge in a Neural Network

    Hinton, Geoffrey and Vinyals, Oriol and Dean, Jeff , month = mar, year =. Distilling the. doi:10.48550/arXiv.1503.02531 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1503.02531

  68. [68]

    Browne, Simone , month = sep, year =. Dark

  69. [69]

    Mansfield, Tony , month = mar, year =. Facial. National Physical Laboratory (NPL), UK , publisher =

  70. [70]

    MIT Case Studies in Social and Ethical Responsibilities of Computing , author =

    The. MIT Case Studies in Social and Ethical Responsibilities of Computing , author =. doi:10.21428/2c646de5.62272586 , language =

    work page doi:10.21428/2c646de5.62272586

  71. [71]

    Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks , url =

    Croce, Francesco and Hein, Matthias , month = aug, year =. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks , url =. doi:10.48550/arXiv.2003.01690 , abstract =

    work page doi:10.48550/arxiv.2003.01690 2003

  72. [72]

    Toward a. Social. 2014 , doi =

    2014

  73. [73]

    and Star, Susan Leigh , year =

    Bowker, Geoffrey C. and Star, Susan Leigh , year =. Sorting things out: classification and its consequences , isbn =

  74. [74]

    Computer Law & Security Review , author =

    The flaws of policies requiring human oversight of government algorithms , volume =. Computer Law & Security Review , author =. 2022 , pages =. doi:10.1016/j.clsr.2022.105681 , language =

    work page doi:10.1016/j.clsr.2022.105681 2022

  75. [75]

    Race after technology: abolitionist tools for the

    Benjamin, Ruha , year =. Race after technology: abolitionist tools for the

  76. [76]

    Tutorial: 21

    Narayanan, Arvind , month = feb, year =. Tutorial: 21

  77. [77]

    London Review of Books , author =

    Making. London Review of Books , author =

  78. [78]

    , year =

    Agre, P. , year =. Toward a

  79. [79]

    Leveling

    Zietlow, Dominik and Lohaus, Michael and Balakrishnan, Guha and Kleindessner, Matthäus and Locatello, Francesco and Schölkopf, Bernhard and Russell, Chris , month = mar, year =. Leveling. doi:10.48550/arXiv.2203.04913 , abstract =

    work page doi:10.48550/arxiv.2203.04913

  80. [80]

    Romeo, Lanzino and Federico, Fontana and Anxhelo, Diko and Raoul, Marini Marco and Luigi, Cinque , month = jun, year =. Faster. doi:10.48550/arXiv.2406.04932 , abstract =

    work page doi:10.48550/arxiv.2406.04932

Showing first 80 references.