Pith. sign in

REVIEW 3 major objections 5 minor 27 references

Bias Analysis of L2 Speaking Assessment Systems Using Concept Activation Vectors

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read CAV audit: encoding a speaker trait is not bias

desk verdict Solid, useful extension of CAV bias analysis to neural speaking graders; the central claim holds but the gradient metric deserves a direct positive control before the SAE attenuation result is taken at face value. read the letter →

arxiv 2608.06300 v1 pith:3QOPKNI3 submitted 2026-08-06 cs.AI

Arya Labroo , Mengjie Qian , Kate Knill This is my paper · ORCID
classification cs.AI
keywords ConceptActivationVectorsbiasauditL2speakingassessmentsparseautoencodersinterpretabilityfairnessgradientsensitivityTransformergraders
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper claims that whether a concept—such as a speaker's first language, gender, or age—is detectable in a neural scoring model's internal representations is a property of the layer and architecture being probed, not of the concept itself. It shows, for two Transformer-based speaking graders, that a concept can be highly recoverable by a linear classifier while having almost no influence on the predicted score, and vice versa. The authors therefore argue that bias audits must measure the gradient of the score with respect to the concept direction, not just whether the concept is linearly separable. They also find that sparse autoencoders make concepts more recoverable but weaken the gradient-based sensitivity signal, so the sparse representation does not improve bias measurement. The central takeaway is that concept recovery and concept influence are distinct quantities that must be measured separately.

What carries the argument

The load-bearing objects are Concept Activation Vectors (CAVs), directions in a model's activation space learned by a linear classifier that separates examples exhibiting a human-interpretable concept from examples that do not, and the gradient-based sensitivity metric $B_{\mathrm{gr}}$, the mean per-speaker cosine distance between a scaled CAV direction and the gradient of the predicted score with respect to the activation at that layer. A TopK sparse autoencoder provides an alternative latent space: its linear decoder maps a latent-space CAV back into activation space as a weighted combination of decoder columns, where the same gradient metric is applied. The paper's central argument rests on the divergence between these two measurements—balanced classifier accuracy for recoverability and $B_{\mathrm{gr}}$ for influence—and on the observation that the divergence is representation-dependent, not concept-dependent.

What would settle it

Run a positive-control experiment on the Transformer graders: deliberately make the score depend on a synthetic attribute, then apply the CAV and $B_{\mathrm{gr}}$ pipeline. If the metric does not register the known bias, the near-baseline sensitivities for Whisper and the SAE attenuation would be artifacts of the metric. Alternatively, nudge each test activation along the scaled CAV direction by a small epsilon, recompute the score, and compare the observed score change with the gradient prediction; disagreement would refute the metric's validity as a measure of influence.

Watch

Extended reading notes

Core claim

The paper's central finding is that linear recoverability of a concept is governed by the representation it is measured in, not the concept itself, and that recoverability does not imply influence. In the Whisper grader, nearly every demographic and first-language concept is linearly recoverable at the head input layer, yet the gradient-based sensitivity metric stays close to the no-sensitivity baseline for all of them; in the BERT grader, the same concepts are often not recoverable, but the several that are recoverable show score sensitivity that tracks the dataset's score imbalance, deepening toward the output layer. The paper also shows that training CAVs in a TopK sparse autoencoder latent space makes every concept more linearly separable, but the mapped concept directions lose alignment with the score gradient, shifting sensitivity toward the unbiased baseline of 1, most severely in the lowest-dimensional layer. This is why the authors conclude that encoding and influence must be separated in fairness analysis, and why sparse autoencoders, as used here, do not improve bias measurement.

Load-bearing premise

The bias conclusions assume the gradient-based metric $B_{\mathrm{gr}}$ faithfully measures how much a concept influences the predicted score, which requires the regression head to be approximately linear in the activations around each data point and the scaled CAV direction to be the correct intervention direction; the paper does not run the injected-bias positive control that the feature-based predecessor used, so a failure of the metric could explain the near-baseline values for Whisper and the SAE attenuation.

Editorial extensions

If this is right

  • Bias audits of Transformer-based speaking graders that report only linear-probe recoverability of demographic attributes will overstate bias: the Whisper grader encodes gender, age, and first-language concepts yet shows near-baseline score sensitivity across all of them.
  • Sensitivity must be measured layer by layer: the BERT grader's sensitivity to proficiency and to L1 directions deepens toward the output, so auditing only early layers would miss the effect, while auditing only late layers would overstate it.
  • Sparse autoencoders, in the form used here, are not a drop-in improvement for CAV-based bias analysis: they boost recoverability but attenuate the sensitivity signal exactly where bias was clearest (low-dimensional head layers).
  • A score-aware sparse autoencoder, trained to reconstruct the grader's output or to weight reconstruction error by the score gradient, is the natural next step proposed by the paper for retaining both recoverability and influence.

Reading between the lines

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

  • Editorial inference: The recoverability-versus-influence gap is likely a general phenomenon in deep classifiers, since entangled representations allow linear separation of attributes that do not drive the output; interpretability methods that rely on linear probing alone may therefore routinely report spurious 'bias'.
  • Editorial inference: The gradient metric's assumption of local linearity could be tested directly by perturbing activations along the CAV and comparing the actual score change with the gradient prediction; disagreement would suggest the Whisper near-baseline results understate true influence.
  • Editorial inference: The SAE attenuation pattern predicts that a dictionary-learning probe with a reconstruction objective weighted by the score gradient would preserve sensitivity in low-dimensional layers; this is a testable extension the paper does not itself carry out.
  • Editorial inference: For practitioners, the operational rule is to never flag a model as biased on the basis of a recoverable attribute alone, and to require a sensitivity measurement that ties the attribute direction to actual score movement.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. This paper extends Concept Activation Vector (CAV) bias analysis from feature-based L2 speaking graders to two neural grading systems: a text-based BERT grader and a speech-and-text Whisper grader. For a set of proficiency, demographic, and L1 concepts, the authors train linear CAVs from activations at two layers per model and quantify score influence with Bgr, the mean cosine distance between a scaled CAV and the per-speaker gradient of the predicted score. They also introduce a TopK sparse autoencoder (SAE) variant that learns CAVs in a sparse latent space and maps them back to activation space using the linear decoder. On BULATS and S&I, they report that concept recoverability varies strongly across layers and architectures, that Whisper encodes demographic and L1 concepts at its head input yet exhibits near-baseline Bgr values, that the BERT grader becomes more score-sensitive with depth, and that SAEs improve linear recoverability but attenuate measured sensitivity, especially in low-dimensional layers. The central conclusion is that linear recoverability of a concept is governed by the representation being probed and does not imply score influence, and that SAE-based feature recovery should not be reported as evidence of bias without a separate influence check.

Significance. If the gradient-based metric is faithful, the paper offers a useful caution for the speech assessment community: CAV recoverability in a neural grader is not evidence of bias, and SAE-discovered features need to be validated with respect to the score gradient. The study has concrete strengths: five-seed statistics for the main metric, replication on a second corpus, explicit handling of degenerate CAVs, and a clean descriptive separation between encoding and influence. The principal risk is methodological: Bgr is a first-order approximation applied in layers whose downstream head is nonlinear, and the predecessor work's injected-bias positive control is not repeated here. The paper's headline claims about Whisper insensitivity and SAE attenuation therefore rest on an unvalidated metric at the very layers where the conclusions are strongest.

major comments (3)
  1. [§II and §V-C (Eq. 6)] The gradient metric Bgr(c) is a first-order approximation. At BERT Layer 1 and Whisper dense.in, the regression head contains a subsequent ReLU (two linear layers with a ReLU for Whisper; two hidden layers with ReLUs for BERT), so moving from h(i) along Δh(c) can cross a ReLU boundary and the gradient no longer predicts the actual score change. The near-baseline Bgr values reported for Whisper, and the SAE attenuation at BERT Layer 1, could therefore be artifacts of the linearization rather than properties of the graders. The predecessor [16] validated its gradient metric with a deliberately injected L1 bias in a feature-based grader; this paper should include an analogous positive control for the transformer setting, for example by fine-tuning a grader with a known additive bias in a chosen activation direction and checking that Bgr moves away from 1, or by reporting finite-perturbation score changes F_y(h+εΔh)-F_y(h) alongside Bgr at the nonlinear layers. Without such a control, the paper's central claim that Whisper is insensitive and that SAEs attenuate sensitivity is under-supported.
  2. [§V-C and Table IV] The architecture-level comparison is confounded with layer dimensionality and depth. BERT Layer 1 is 600-dimensional, BERT Layer 2 is 20-dimensional, Whisper dense.in is 2048-dimensional, and Whisper act.out is 64-dimensional; these layers also sit at different distances from the output and are separated by different nonlinearities. Table IV therefore does not establish that the observed differences are due to the model architecture (BERT versus Whisper) rather than to the dimension or position of the probed representation. Since the paper's stated conclusion is that recoverability and sensitivity are governed by 'the representation it is measured in', this is acceptable as a descriptive claim about these particular layers, but it should be explicitly acknowledged as a limit; otherwise the reader may over-generalize the architecture-level comparison.
  3. [§V-D and Eq. (8)] The SAE attenuation conclusion rests on comparing Bgr in activation space for the mapped SAE-CAV with the standard CAV. However, the paper does not report how well the SAE reconstruction preserves the score-gradient direction, despite this being the mechanism proposed in Section III ('if the SAE does not reconstruct the part of activation space relevant to a concept well, the mapped CAV may be unreliable'). A direct check—for example, the cosine similarity between ∇h F_y and the reconstruction of that gradient by the SAE decoder, or the reconstruction error restricted to the score-relevant subspace—would confirm that the attenuation is due to the sparse representation losing the influence direction rather than to a mis-specified mapping. Without this check, the claim that SAEs 'attenuate exactly the sensitivity signal that bias analysis depends on' is not fully supported.
minor comments (5)
  1. [§II, Eq. (5)] Equation (5) is redundant for Bgr: the cosine distance in Eq. (6) is invariant to positive scaling of the CAV direction, so multiplying the normalized direction by the mean activation norm does not make Bgr values comparable across layers. The surrounding text should be corrected to explain what, if anything, the scaling is intended to do.
  2. [§IV-A] The sentence beginning 'is used to assess whether findings on BULATS generalise' is grammatically incomplete; it should be 'This corpus is used to assess...'.
  3. [§V-B and Table III] Table III reports balanced accuracy on the training set only, and no standard deviations or held-out numbers are given for the CAV classifier. Since the CAV is trained on the same activations, the paper should state whether the recoverability patterns also hold on evaluation or development activations, or justify why training-set accuracy is the appropriate quality measure for a direction used in bias analysis.
  4. [§V-D and Table V] Table V reports only the resulting sparsity for each SAE configuration. The text says the configuration was selected to balance reconstruction error, dead-feature fraction, and decoder coherence, but none of the selected values for reconstruction error, dead-feature fraction, or coherence are reported; adding them would support the claim that the configurations were chosen jointly rather than by sparsity alone.
  5. [Throughout] There are a few typos and wording slips, including 'complimentary representation' in the contributions list (should be 'complementary'), 'profiency' in Section V-C, and 'The two metrics ask different things' followed by a colon that introduces a sentence rather than a list. A careful proofread would resolve these.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the CAV/Bgr/SAE analyses are descriptive measurements on fixed graders, with no fitted target quantity and with out-of-sample reproduction on S&I.

full rationale

The paper's central claims are descriptive: concept recoverability and gradient-based sensitivity are measured, not predicted from fitted parameters. The CAV direction (Eq. 3) and the sensitivity metric Bgr (Eq. 6) are defined independently of the conclusions, and the observed divergence between recoverability and influence is an empirical result rather than a definitional identity. SAE hyperparameters are selected on the development split using reconstruction error, sparsity, dead-feature fraction, and decoder coherence, none of which encode Bgr or the recoverability comparison, so the SAE attenuation finding is not forced by construction. The main proficiency and architecture trends are reproduced on the independent S&I corpus, providing an external check. Self-citations [16], [22], [23], [26], [27] supply the predecessor method, corpus, and grader checkpoints, but no load-bearing inference in the derivation chain reduces to an unverified self-citation. The absence of an injected-bias positive control for the gradient metric on nonlinear Transformer heads is a validity limitation that could make Bgr an artifact, but it is a correctness risk, not a circularity of the paper's derivation chain.

Assumptions & free parameters 9 free parameters · 5 assumptions · 0 invented entities

The paper's claims rest mainly on the CAV linear-separability assumption, the local gradient metric, the quality of human labels and inferred gender labels, and the comparability of the probed layers. The SAE configurations are fitted on the development split and the CAV classifier has an unreported regularization constant, but these are hyperparameters rather than entities invented for the result. No new particles, mediators, or other entities are introduced.

free parameters (9)
  • CAV hinge classifier L2 penalty
    Regularization strength in Eq. (3); not reported in the paper, yet it shapes the CAV direction d(c) and therefore every Bgr value.
  • BERT Layer 1 SAE latent width m = 2400
    Selected on the development split along with k to balance reconstruction error, sparsity, dead-feature fraction, and decoder coherence.
  • BERT Layer 1 SAE active units k = 320
    Selected with m=2400; determines sparsity 0.867 and affects SAE-CAV attenuation.
  • BERT Layer 2 SAE latent width m = 80
    Selected on development split for the 20-dimensional layer.
  • BERT Layer 2 SAE active units k = 16
    Selected with m=80; sparsity 0.800.
  • Whisper dense.in SAE latent width m = 6144
    Selected on development split for the 2048-dimensional head input layer.
  • Whisper dense.in SAE active units k = 480
    Selected with m=6144; sparsity 0.922.
  • Whisper act.out SAE latent width m = 256
    Selected on development split for the 64-dimensional post-ReLU layer.
  • Whisper act.out SAE active units k = 24
    Selected with m=256; sparsity 0.906.
assumptions (5)
  • standard math A concept is meaningfully encoded in a layer only if it is linearly recoverable there, and the CAV direction found by the hinge classifier is the right concept direction.
    Basis of the CAV method in Sections II and V-B; if linear separability is not the right notion for entangled Transformer representations, all recoverability comparisons are on weak ground.
  • domain assumption The local gradient-alignment metric Bgr (Eq. 6) faithfully measures how much a concept influences the predicted score.
    The central influence claims in Tables IV and VI rest on this metric; it assumes local linearity of the regression head and that the scaled CAV is the relevant intervention direction.
  • domain assumption Human examiner scores are unbiased proficiency labels, so correlations between a concept and mean score (Table I) are treated as proficiency signal rather than bias.
    The paper reads L1 sensitivity against score gaps; if the human labels themselves encode L1 bias, the interpretation collapses.
  • domain assumption Gender labels on S&I, inferred by an ECAPA-TDNN voice classifier, are accurate enough for gender bias analysis.
    Section IV-A; no accuracy or error analysis for the inferred labels is reported.
  • domain assumption The four probed layers (BERT Layer 1/2 and Whisper dense.in/act.out) are representative and comparable enough to attribute recoverability and sensitivity differences to architecture rather than layer dimensionality or position.
    Table III compares 600/20-dimensional BERT layers with 2048/64-dimensional Whisper layers; dimensionality is a confound for the architecture-dependent conclusion.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Bias Analysis of L2 Speaking Assessment Systems Using Concept Activation Vectors." pith.science (2026). https://pith.science/paper/3QOPKNI3

@misc{pith2026260806300,
  author       = {Pith},
  title        = {Pith review of: Bias Analysis of L2 Speaking Assessment Systems Using Concept Activation Vectors},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3QOPKNI3}},
  note         = {Machine review of arXiv:2608.06300}
}
read the original abstract

Automatic speaking assessment systems are increasingly deployed in high-stakes settings to mark second language (L2) learners' speaking tests, making it critical to show that their scores depend on speaking proficiency rather than irrelevant speaker attributes such as first language (L1) or age. Transformer-based foundation models have improved the accuracy of these L2 speaking graders, but their black-box representations make fairness and interpretability analysis more difficult. Building on prior work that used Concept Activation Vectors (CAVs) to detect bias towards unwanted attributes (`concepts') in feature-based graders, we extend CAV-based analysis to two neural speaking assessment systems: a text-based BERT grader and a speech-and-text multimodal grader based on Whisper. CAVs represent human-interpretable concepts as directions in a model's activation space, allowing us to distinguish between whether a concept is encoded in a model's internal representations and whether it influences the predicted score, the latter quantified using a gradient-based sensitivity metric. Since CAVs rely on linear separability, which is less likely in complex neural embedding spaces, we also investigate whether sparse autoencoders (SAEs) provide cleaner concept directions by learning CAVs in a sparse latent space and mapping them back to activation space. Our analysis shows that concept recoverability depends strongly on the representation and architecture being probed, rather than on the concept alone. Sensitivity to concepts is also architecture-dependent. SAEs make concepts more linearly recoverable, but attenuate the original activation-space sensitivity, especially in low-dimensional layers. These findings highlight the need to distinguish concept recoverability from concept influence when auditing bias in speaking assessment systems.

Figures

Figures reproduced from arXiv: 2608.06300 by the authors.

Figure 1
Figure 1. Per-speaker Bgr for the BERT grader at Layer 1 (left) and Layer 2 (right), for the ≥ A2, Gender (female), and L1 Dutch concepts. The ≥A2 and Dutch concepts move further below the no-sensitivity baseline of 1 from Layer 1 to Layer 2, while Gender stays close to 1. 1.5 2.0 2.5 3.0 3.5 4.0 4.5 Predicted score 0.00 0.25 0.50 0.75 1.00 1.25 1.50 1.75 2.00 B gr BERT Layer 2: CAV-Only vs SAE-CAV Bgr SAE-CAV CAV-only [PITH… view at source ↗
Figure 2
Figure 2. Per-speaker Bgr for the L1 Dutch concept at BERT Layer 2, comparing the activation-space CAV with the SAE￾CAV. The SAE-CAV is attenuated towards the no-sensitivity baseline of 1. from 0.91 to 0.52 across the two BERT layers, while Gujarati speakers score 0.45 below and its Bgr rises from 1.03 to 1.10, in the score-decreasing direction. The Whisper grader, by contrast, stays close to Bgr = 1 for all demographic and L… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

27 extracted references · 23 canonical work pages

  1. [16]

    Analysing Bias in Spoken Language Assessment Using Concept Activation Vectors,

    X. Wei, M. J. F. Gales, and K. M. Knill, “Analysing Bias in Spoken Language Assessment Using Concept Activation Vectors,” inIEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2021, pp. 7753–7757

    work page 2021

  2. [1]

    Universal Adversarial Attacks on Spoken Language Assessment Systems,

    V . Raina, M. J. Gales, and K. M. Knill, “Universal Adversarial Attacks on Spoken Language Assessment Systems,” inInterspeech 2020, 2020, pp. 3855–3859

    work page 2020

  3. [2]

    Proficiency Assessment of L2 Spoken En- glish Using Wav2Vec 2.0,

    S. Bann `o and M. Matassoni, “Proficiency Assessment of L2 Spoken En- glish Using Wav2Vec 2.0,” in2022 IEEE Spoken Language Technology Workshop (SLT), 2023, pp. 1088–1095

    work page 2023

  4. [3]

    Assessment of L2 Oral Proficiency Using Self-Supervised Speech Representation Learning,

    S. Bann `o, K. Knill, M. Matassoni, V . Raina, and M. Gales, “Assessment of L2 Oral Proficiency Using Self-Supervised Speech Representation Learning,” inProceedings of 9th Workshop on Speech and Language Technology in Education (SLaTE), 2023, pp. 126–130

    work page 2023

  5. [4]

    Automatic Assessment of Conversational Speaking Tests,

    S. W. McKnight, A. Civelekoglu, M. Gales, S. Bann `o, A. Liusie, and K. M. Knill, “Automatic Assessment of Conversational Speaking Tests,” in9th Workshop on Speech and Language Technology in Education (SLaTE), 2023, pp. 99–103

    work page 2023

  6. [5]

    Assessment of L2 Oral Proficiency using Speech Large Language Models,

    R. Ma, M. Qian, S. Tang, S. Bann `o, K. M. Knill, and M. J. Gales, “Assessment of L2 Oral Proficiency using Speech Large Language Models,” inInterspeech 2025, 2025, pp. 5078–5082

    work page 2025

  7. [6]

    The NTNU System at the S&I Challenge 2025 SLA Open Track,

    H.-Y . Lin, T. H. Lo, Y . H. Fang, J. K. Lin, C. C. Wang, H. C. Lu, and B. Chen, “The NTNU System at the S&I Challenge 2025 SLA Open Track,” in10th Workshop on Speech and Language Technology in Education (SLaTE), 2025, pp. 148–152

    2025

  8. [7]

    One Whisper to Grade Them All,

    N. Phan, A. Porwal, Y . Getman, E. V oskoboinik, T. Gr ´osz, and M. Ku- rimo, “One Whisper to Grade Them All,” in10th Workshop on Speech and Language Technology in Education (SLaTE), 2025, pp. 56–60

    work page 2025

Show all 27 references
  1. [8]

    BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding,

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding,” in Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologi...

    2019

  2. [9]

    wav2vec 2.0: A framework for self-supervised learning of speech representations,

    A. Baevski, Y . Zhou, A. Mohamed, and M. Auli, “wav2vec 2.0: A framework for self-supervised learning of speech representations,” vol. 33, 2020, pp. 12 449–12 460

    2020

  3. [10]

    Robust Speech Recognition via Large-Scale Weak Super- vision,

    A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust Speech Recognition via Large-Scale Weak Super- vision,” inProceedings of the 40th International Conference on Machine Learning (ICML). PMLR, 2023, pp. 28 492–28 518

    2023

  4. [11]

    Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification,

    J. Buolamwini and T. Gebru, “Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification,” inProceedings of the 1st Conference on Fairness, Accountability and Transparency (FAccT). PMLR, 2018, pp. 77–91

    2018

  5. [12]

    Man is to Computer Programmer as Woman is to Homemaker? Debiasing Word Embeddings,

    T. Bolukbasi, K.-W. Chang, J. Zou, V . Saligrama, and A. Kalai, “Man is to Computer Programmer as Woman is to Homemaker? Debiasing Word Embeddings,” inAdvances in Neural Information Processing Systems 29 (NeurIPS), 2016

    2016

  6. [13]

    Understanding Intermediate Layers Using Linear Classifier Probes,

    G. Alain and Y . Bengio, “Understanding Intermediate Layers Using Linear Classifier Probes,” inInternational Conference on Learning Representations (ICLR) Workshop, 2017

    2017

  7. [14]

    Network Dis- section: Quantifying Interpretability of Deep Visual Representations,

    D. Bau, B. Zhou, A. Khosla, A. Oliva, and A. Torralba, “Network Dis- section: Quantifying Interpretability of Deep Visual Representations,” inProceedings of the IEEE conference on computer vision and pattern recognition (CVPR), 2017, pp. 6541–6549

    2017

  8. [15]

    Interpretability Beyond Feature Attribution: Quantitative Testing with Concept Activation Vectors (TCA V),

    B. Kim, M. Wattenberg, J. Gilmer, C. Cai, J. Wexler, F. Vi ´egas, and R. Sayres, “Interpretability Beyond Feature Attribution: Quantitative Testing with Concept Activation Vectors (TCA V),” inProceedings of the 35th International Conference on Machine Learning. PMLR, 2018, pp....

    2018

  9. [17]

    Towards Monosemanticity: Decomposing Language Models With Dictionary Learning,

    T. Bricken, A. Templeton, J. Batson, B. Chen, A. Jermyn, T. Conerly, N. L. Turner, C. Anil, C. Denison, A. Askell, R. Lasenby, Y . Wu, S. Kravec, N. Schiefer, T. Maxwell, N. Joseph, Z. Hatfield-Dodds, A. Tamkin, K. Nguyen, B. McLean, J. E. Burke, T. Hume, S. Carter, T. Henigha...

    2023

  10. [18]

    Scaling and Evaluating Sparse Au- toencoders,

    L. Gao, T. Dupr ´e la Tour, H. Tillman, G. Goh, R. Troll, A. Radford, I. Sutskever, J. Leike, and J. Wu, “Scaling and Evaluating Sparse Au- toencoders,” inInternational Conference on Learning Representations, vol. 2025, 2025, pp. 26 721–26 754

    2025

  11. [19]

    Automatic Assessment of English as a Second Language,

    B. Davies, “Automatic Assessment of English as a Second Language,” Department of Engineering, University of Cambridge, Fourth-Year Project Report, 2024

    2024

  12. [20]

    The BULATS Online Speaking Test,

    L. Chambers and K. Ingham, “The BULATS Online Speaking Test,”Research Notes, 2011. [Online]. Available: https://www. cambridgeenglish.org/Images/23161-research-notes-43.pdf

    2011

  13. [21]

    Cambridge: Cambridge University Press, 2001

    Council of Europe,Common European Framework of Reference for Languages: Learning, Teaching, Assessment. Cambridge: Cambridge University Press, 2001. [Online]. Available: https: //rm.coe.int/1680459f97

    2001

  14. [22]

    The Speak & Improve Corpus 2025: an L2 English Speech Corpus for Language Assessment and Feedback,

    K. Knill, D. Nicholls, M. J. Gales, M. Qian, and P. Stroinski, “The Speak & Improve Corpus 2025: an L2 English Speech Corpus for Language Assessment and Feedback,” 2025. [Online]. Available: https://doi.org/10.17863/CAM.114333

    2025 doi

  15. [23]

    Introducing the Speak & Improve Corpus 2025: an L2 English Speech Corpus for Language Assessment and Feedback,

    K. M. Knill, D. Nicholls, M. J. F. Gales, M. Qian, and P. Stroinski, “Introducing the Speak & Improve Corpus 2025: an L2 English Speech Corpus for Language Assessment and Feedback,” inProceedings of the 10th Workshop on Speech and Language Technology in Education (SLaTE), 2025...

    2025

  16. [24]

    ECAPA-TDNN: Emphasized Channel Attention, Propagation and Aggregation in TDNN Based Speaker Verification,

    B. Desplanques, J. Thienpondt, and K. Demuynck, “ECAPA-TDNN: Emphasized Channel Attention, Propagation and Aggregation in TDNN Based Speaker Verification,” inInterspeech 2020, 2020, pp. 3830–3834

    2020

  17. [25]

    V oice Gender Classifier,

    J. Huh, “V oice Gender Classifier,” https://github.com/JaesungHuh/ voice-gender-classifier, gitHub repository. Accessed November 2025

    2025

  18. [26]

    Speak & improve challenge 2025: Tasks and baseline systems,

    M. Qian, K. Knill, S. Banno, S. Tang, P. Karanasou, M. J. Gales, and D. Nicholls, “Speak & improve challenge 2025: Tasks and baseline systems,”arXiv preprint arXiv:2412.11985, 2024

    2025 arXiv

  19. [27]

    Speak & Improve Challenge 2025,

    M. Qian, K. M. Knill, S. Bann `o, S. Tang, P. Karanasou, M. J. Gales, and D. Nicholls, “Speak & Improve Challenge 2025,” in10th Workshop on Speech and Language Technology in Education (SLaTE), 2025, pp. 41–45

    2025

Pith tools

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