Pith. sign in

REVIEW 4 major objections 6 minor 30 references

SpectroStream: A Versatile Neural Codec for General Audio

T0 review · 4 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read SpectroStream, a 2D time-frequency neural codec, claims full-band 48 kHz stereo music coding at 4–16 kbps with quality above a strong baseline at every bit rate tested.

desk verdict A plausible new 2D time-frequency codec architecture with a serious evaluation gap: the 48 kHz claim is never actually tested because MUSDB18 is 44.1 kHz and no resampling is stated. read the letter →

arxiv 2508.05207 v1 pith:JAOKW7PI submitted 2025-08-07 cs.SD cs.AIeess.AS

Yunpeng Li , Kehang Han , Brian McWilliams , Zalan Borsos , Marco Tagliasacchi This is my paper
classification cs.SDcs.AIeess.AS
keywords SpectroStreamneuralaudiocodectime-frequencyrepresentationstereomusiccompressionresidualvectorquantizationdelayedfusionmulti-scaleSTFTdiscriminatorlow-bitrate
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

SpectroStream is a neural audio codec that targets full-band 48 kHz stereo music at 4–16 kbps, extending the SoundStream line beyond 24 kHz mono. The paper's central claim is that operating a 2D-convolutional encoder–decoder directly on STFT spectrograms, rather than on raw waveforms, yields better reconstruction quality at high sample rates. For multi-channel audio, a delayed-fusion encoder and early-splitting decoder let channels share later layers only, which the authors argue balances per-channel fidelity with cross-channel phase coherence. The reported result is a quality advantage over Descript Audio Codec (DAC) at every tested bit rate, most clearly at 2.7 kbps per channel, where SpectroStream scores 3.21 versus 1.47 on ViSQOL and is preferred by 76.3% of listeners in an A/B test. If these numbers hold, very low bitrate coding of stereo music, and a compact token representation for generative audio, become practical.

What carries the argument

The central object is the complex STFT spectrogram as the codec's native representation: real and imaginary parts become the channel dimension of 2D convolutional encoder and decoder, replacing the 1D waveform convolutions of SoundStream. The other load-bearing pieces are the delayed-fusion/early-splitting channel strategy, which keeps stereo channels separate until a chosen layer and jointly processes them afterward; residual vector quantization with biased quantizer dropout and full quantizer bypass, which makes the bit rate tunable from 4 to 16 kbps; and a multi-scale STFT discriminator that adjudicates realism at six window lengths. Together they let the model stream with 80 ms architect

What would settle it

Run a preregistered blind A/B listening test on the same MUSDB clips at 2.7 kbps per channel with matched loudness: if listener preference is near chance rather than the reported 76.3%, the perceptual claim fails. Separately, recompute Table 1 with ViSQOL's audio mode; if SpectroStream no longer leads DAC by more than a point, the headline gain is metric-dependent.

Watch

Extended reading notes

Core claim

The paper is trying to establish that a neural codec built on time-frequency representations can compress full-band stereo music to rates around 4–16 kbps without sacrificing quality relative to a strong waveform-based codec. SpectroStream encodes each channel by STFT, treats the real and imaginary parts as separate input channels, and downsamples the spectrogram with strided 2D convolutions to 25 embeddings per second. A residual vector quantizer with 64 levels, each over a vocabulary of 1024 codes, converts these embeddings into discrete tokens, and the bit rate is adjustable at inference by truncating the number of quantizers. The delayed-fusion encoder separates channels in early layers

Load-bearing premise

The whole quality comparison leans on ViSQOL's speech mode being a reliable perceptual proxy for full-band stereo music, because the paper found the audio mode saturating and set it aside; if that metric favors SpectroStream, the objective margins in Table 1 do not represent perceived quality.

Editorial extensions

If this is right

  • At 2.7 kbps per channel, stereo music is reported at ViSQOL 3.21 versus DAC's 1.47, so very low bitrate music streaming is a realistic target rather than a speech-only niche.
  • Because bit rate is selected by truncating RVQ levels, one trained model covers the full 4–16 kbps range with no retraining or architecture change.
  • The 80 ms architectural latency and CPU-only real-time inference make the codec suitable for streaming and interactive applications.
  • The delayed-fusion/early-splitting principle gives other multi-channel neural coders a concrete design rule: fuse channels only at the point where phase coherence can be learned without sacrificing per-channel fidelity.
  • The discrete 25 Hz token stream is a candidate tokenizer for language-model-based stereo music generation, extending the role SoundStream played for mono audio.

Reading between the lines

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

  • If the ViSQOL speech-mode assumption is valid, the low-bitrate gap is large enough that competing codecs should be re-evaluated at 2.7 kbps; if not, the objective margin may shrink under a metric that does not saturate.
  • The paper leaves the optimal fusion point as an empirical finding; a testable extension is to sweep the fusion layer for 5.1 or binaural audio, where cross-channel phase matters differently.
  • Biased quantizer dropout toward low truncation levels is a training trick that could be ported to other RVQ-based codecs to improve their low-bitrate regimes.
  • The 25 Hz token rate matches the frame rate of many audio language models, so SpectroStream codes could be plugged into existing generative pipelines; whether that improves music generation is a direct experiment the paper does not run.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. SpectroStream is a neural audio codec that operates on STFT spectrograms with a 2D convolutional encoder/decoder, RVQ quantization, a multi-scale STFT discriminator, and a delayed-fusion strategy for stereo channels. The paper claims full-band 48 kHz stereo music compression at 4–16 kbps with low architectural latency, and reports objective (ViSQOL) and subjective (A/B listening) comparisons against DAC at 2.7, 5.3, and 8 kbps per channel. The central claim is that SpectroStream outperforms DAC consistently, with particularly large gains at low bit rates.

Significance. If the results hold, the architecture makes useful contributions: operating directly on time-frequency representations with 2D convolutions, the delayed-fusion/early-splitting design for stereo phase consistency, and the emphasis on causal streaming with 80 ms latency are all sensible and potentially valuable for audio generation and compression. The paper also gives a fairly detailed description of the loss formulation and quantizer dropout scheme. However, the evidence base is currently too narrow and contains a load-bearing sample-rate ambiguity: the evaluation uses MUSDB18 (44.1 kHz) while the system and the headline claim are for 48 kHz full-band audio. The objective metric choice is post hoc, and no confidence intervals or significance tests are reported. The central claim is therefore not yet established to the standard expected for a systems paper.

major comments (4)
  1. [Section 4 / Section 2] The full-band 48 kHz claim is not operationalized. MUSDB18 is distributed as 44.1 kHz stereo audio, but Section 4 never states whether the 49 selected tracks were resampled to 48 kHz. If they were not, the STFT parameters in Section 2 (48 kHz, window 960, hop 480) are applied to a signal at the wrong sample rate, making the comparison ill-defined. If they were resampled, the test material still has no content above 22.05 kHz, so the 'full-band 48 kHz' behavior is never exercised. Either way, the Table 1 gains and Table 2 preferences do not establish the headline claim. The paper must state the native sample rate of every test item, the resampling procedure, and include at least one natively 48 kHz full-band evaluation set.
  2. [Section 4, footnote 1] The choice of ViSQOL's speech mode as the primary objective metric is post hoc: the audio mode was tried and discarded because it was 'more prone to saturation and not as informative.' No evidence is provided that speech-mode ViSQOL is a valid perceptual proxy for full-band music, and no correlation with the listening test is reported. The large ViSQOL advantage at 2.7 kbps (3.21 vs. 1.47) may be an artifact of the metric. Please report both audio- and speech-mode ViSQOL (or a second objective metric such as a mel distance), and validate the chosen metric against the subjective preferences on a per-condition basis.
  3. [Section 4, Tables 1 and 2] Tables 1 and 2 report point estimates only. With N=49 tracks for objective scores and 20 clips/60 raters for listening, there are no confidence intervals, no significance tests, and no per-item analysis. At 5.3 and 8 kbps the listening preferences (55.0% and 50.8%) are plausibly not distinguishable from chance. Please provide confidence intervals, bootstrap or Wilcoxon tests, and the distribution of scores across items. This is essential to support the 'consistently superior' claim.
  4. [Section 4 / Section 5] The comparison is limited to a single pre-trained baseline, DAC. Section 5 claims comparison with 'state-of-the-art models with similar capabilities,' but FlowDec (cited as [7]) is a full-band general audio codec and Encodec [3] is a standard reference; neither is evaluated. Without at least one additional full-band baseline, the paper should temper the headline claim or add comparisons.
minor comments (6)
  1. [Section 2, Figure 1] The labels 'STFT (100Hz, 480, 2)' are unclear: is 100 Hz the frame rate, 480 the window/hop, and 2 the number of channels? Please define the tuple explicitly.
  2. [Section 3.1.1] The biased quantizer dropout probabilities are described qualitatively ('twice as high', 'four times as high'); specify the exact sampling distribution, including the normalization over the three ranges.
  3. [Section 4] State the native sample rate and bit depth of the DAC pre-trained model and the exact DAC version/settings used. DAC supports different sample rates and bitrates, and this affects the fairness of the comparison.
  4. [Section 4] Please clarify the listening test protocol: how many clips per condition, whether each rater heard all clips, whether the 'original reference' was used as a hidden anchor, and what instructions were given. Inter-rater agreement would also be helpful.
  5. [Abstract / Section 5] The paper claims 'general audio' but evaluates only stereo music. Consider qualifying the claim to stereo music, or add a small speech/general-sound evaluation.
  6. [Throughout] Minor typos: 'autogressive' in the abstract should be 'autoregressive'; 'a versatile' in the title is fine but check hyphenation; Equation (6) uses B_s and alpha_s without defining them immediately before the formula.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: central results are externally measured against DAC, with an independent subjective listening test; self-citations are lineage only.

full rationale

SpectroStream is an empirical systems paper rather than a derivation: there is no equation in which the reported outcome is constructed from its inputs. Table 1 reports ViSQOL scores computed with an external metric ([23,24]) on 49 MUSDB tracks, and Table 2 reports an A/B preference test with 60 raters. The baseline DAC is an external pretrained model ([4]). These are independent measurements, not fitted quantities renamed as predictions. The self-citations to SoundStream [1] and SEANet [9] describe architectural lineage and training recipe ('The quantizer is trained broadly in the same methodology as [1]'), but they do not carry the argument: the claim that SpectroStream outperforms DAC rests on the measured comparisons, not on SoundStream's conclusions. Hyperparameter choices (STFT frame 960/480, embedding dimension 256, one-embedding look-ahead, biased quantizer dropout) are empirical ablations, not parameters fitted to the test set. Two concerns are validity issues, not circularity. First, footnote 1 selects ViSQOL speech mode after finding the audio mode 'more prone to saturation and not as informative,' which is a metric-choice limitation. Second, the paper claims '48 kHz stereo music' while selecting '49 tracks from the MUSDB dataset [25]' without stating that the MUSDB18 44.1 kHz material was resampled, so the full-band 48 kHz claim is under-supported. Neither concern makes the result equal to its input by construction.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central claim is empirical and depends on a set of hand-tuned architecture hyperparameters and two domain assumptions: that ViSQOL speech mode measures music quality, and that the undisclosed training data do not contaminate the MUSDB evaluation. No new physical entities are introduced.

free parameters (6)
  • STFT window length / hop size = 960 / 480 samples at 48 kHz
    Chosen as a 'good balance' between time and frequency resolution (Section 2), not derived. Determines the 25 Hz embedding rate.
  • Embedding dimension D = 256
    Set after finding 128 worse and 512 no better (Section 2).
  • Number of RVQ quantizers = 64
    Sets max bitrate to 16 kbps and enables flexible bitrate by truncation (Section 2).
  • Feature loss weight lambda_feat = 100
    Set empirically, all other loss weights to 1 (Section 3).
  • Delayed fusion point = not specified numerically
    Called 'crucial' for quality and phase consistency, but no ablation is shown and the exact layer index is not stated (Section 2, Figure 1).
  • Biased quantizer dropout probabilities = 2x and 4x sampling ratio for lower ranges
    Set by hand to emulate a quasi-exponential density (Section 3.1.1).
assumptions (4)
  • domain assumption ViSQOL speech mode is a valid objective proxy for perceived quality of 48 kHz stereo music.
    Invoked in Section 4, footnote 1, where the authors discard the audio mode because it saturates. All objective conclusions in Table 1 depend on this.
  • domain assumption The generic music training dataset does not overlap with the MUSDB evaluation tracks.
    Stated in Section 4, but the training dataset is not named or described, so the independence of train and test data cannot be verified.
  • domain assumption The 49 MUSDB tracks are representative of general full-band stereo audio.
    The paper generalizes to 'general audio' in the title and abstract, but evaluates only music from MUSDB (Section 4).
  • ad hoc to paper The delayed fusion point, chosen near the bottleneck, balances per-channel quality and cross-channel phase consistency.
    Asserted in Section 2 without an ablation or quantitative comparison of fusion positions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SpectroStream: A Versatile Neural Codec for General Audio." pith.science (2026). https://pith.science/paper/JAOKW7PI

@misc{pith2026250805207,
  author       = {Pith},
  title        = {Pith review of: SpectroStream: A Versatile Neural Codec for General Audio},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JAOKW7PI}},
  note         = {Machine review of arXiv:2508.05207}
}
read the original abstract

We propose SpectroStream, a full-band multi-channel neural audio codec. Successor to the well-established SoundStream, SpectroStream extends its capability beyond 24 kHz monophonic audio and enables high-quality reconstruction of 48 kHz stereo music at bit rates of 4--16 kbps. This is accomplished with a new neural architecture that leverages audio representation in the time-frequency domain, which leads to better audio quality especially at higher sample rate. The model also uses a delayed-fusion strategy to handle multi-channel audio, which is crucial in balancing per-channel acoustic quality and cross-channel phase consistency.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

30 extracted references · 23 canonical work pages

  1. [7]

    Audiolm: A language modeling ap- proach to audio generation,

    Z. Borsos, R. Marinier, D. Vincent, E. Kharitonov, O. Pietquin, M. Sharifi, D. Roblek, O. Teboul, D. Grangier, M. Tagliasac- chi, and N. Zeghidour, “Audiolm: A language modeling ap- proach to audio generation,” IEEE Transactions on Audio, Speech, and Language Processing , vol. 31, pp. 2523–2533, 2023

    work page 2023

  2. [3]

    Unlike [17], we do not use the wave-based discrimina- tor [18]

    LEARNING The model is trained with a combination of adversarial and reconstruction losses similar to [1], as illustrated in Figure 3. Unlike [17], we do not use the wave-based discrimina- tor [18]. Instead, we create a multi-scale STFT-based discrim- inator consisting of a collection of individual discriminators, each operating at a different STFT window ...

    work page

  3. [1]

    The de- velopment of SoundStream [1] was a key development that unlocked the potential of autogressive audio generation with standard language-modeling objectives [2]

    INTRODUCTION A high-fidelity audio codec with a compact representation is crucial for audio compression and generation. The de- velopment of SoundStream [1] was a key development that unlocked the potential of autogressive audio generation with standard language-modeling objectives [2]. Since then, many new codecs have been proposed that pushed the fronti...

    work page

  4. [2]

    Additionally, a discriminator is used solely for adver- sarial learning, which we will describe in Section 3

    MODEL ARCHITECTURE SpectroStream consists of an encoder, a decoder, and a quan- tizer. Additionally, a discriminator is used solely for adver- sarial learning, which we will describe in Section 3. The input audio is first converted into spectrograms by applying short-time Fourier transform (STFT) to each of its stereo channels. The real and imaginary comp...

    arXiv 2025

  5. [4]

    Both the generator and the discriminators are updated exactly once in each step

    EXPERIMENTS We trained our model on a generic music dataset for 2 million steps with a batch size of 128 and training example length of 1.28 seconds. Both the generator and the discriminators are updated exactly once in each step. We compare SpectroStream with Descript Audio Codec (DAC) [4], which provides a pre-trained model. We use the open-source imple...

    work page

  6. [5]

    A key innovation of our model lies in its 2D con- volutional architecture, which operates directly on the time- frequency representation

    CONCLUSION We have presented SpectroStream, a full-band, multi-channel neural audio codec designed for high-quality 48 kHz stereo music. A key innovation of our model lies in its 2D con- volutional architecture, which operates directly on the time- frequency representation. This, combined with a multi-scale STFT-based discriminator, leads to significant i...

    work page

  7. [6]

    Soundstream: An end-to-end neural audio codec,

    N. Zeghidour, A. Luebs, A. Omran, J. Skoglund, and M. Tagliasacchi, “Soundstream: An end-to-end neural audio codec,” IEEE/ACM Transactions on Audio, Speech, and Language Processing , vol. 30, pp. 495–507, 2021. [Online]. Available: https://ieeexplore.ieee.org/stamp/stamp. jsp?arnumber=9625818

    work page 2021

  8. [8]

    High fidelity neural audio compression,

    A. D ´efossez, J. Copet, G. Synnaeve, and Y . Adi, “High fidelity neural audio compression,”Transactions on Machine Learning Research, 2023

    2023

Show all 30 references
  1. [9]

    High-fidelity audio compression with improved RVQGAN,

    R. Kumar, P. Seetharaman, A. Luebs, I. Kumar, and K. Kumar, “High-fidelity audio compression with improved RVQGAN,” in Neural Information Processing Systems, 2023

    2023

  2. [10]

    Moshi: a speech-text foundation model for real-time dialogue,

    A. D ´efossez, L. Mazar ´e, M. Orsini, A. Royer, P. P ´erez, H. J´egou, E. Grave, and N. Zeghidour, “Moshi: a speech-text foundation model for real-time dialogue,” 2024

    2024

  3. [11]

    Fo- calcodec: Low-bitrate speech coding via focal modulation net- works,

    L. D. Libera, F. Paissan, C. Subakan, and M. Ravanelli, “Fo- calcodec: Low-bitrate speech coding via focal modulation net- works,” 2025

    2025

  4. [12]

    FlowDec: A flow- based full-band general audio codec with high percep- tual quality,

    S. Welker, M. Le, R. T. Q. Chen, W.-N. Hsu, T. Gerk- mann, A. Richard, and Y .-C. Wu, “FlowDec: A flow- based full-band general audio codec with high percep- tual quality,” in The Thirteenth International Conference on Learning Representations , 2025. [Online]. Available: http...

    2025

  5. [13]

    Streaming keyword spotting on mobile devices,

    O. Rybakov, N. Kononenko, N. Subrahmanya, M. Visontai, and S. Laurenzo, “Streaming keyword spotting on mobile devices,” CoRR, 2020. [Online]. Available: https://arxiv.org/ abs/2005.06720

    2020 arXiv

  6. [14]

    SEANet: A multi-modal speech enhancement network,

    M. Tagliasacchi, Y . Li, K. Misiunas, and D. Roblek, “SEANet: A multi-modal speech enhancement network,” in INTER- SPEECH, 2020

    2020

  7. [15]

    Soundstorm: Efficient parallel au- dio generation,

    Z. Borsos, M. Sharifi, D. Vincent, E. Kharitonov, N. Zeghi- dour, and M. Tagliasacchi, “Soundstorm: Efficient parallel au- dio generation,” arXiv preprint arXiv:2305.09636, 2023

    2023 arXiv

  8. [16]

    Musicrl: Aligning music generation to hu- man preferences,

    G. Cideron, S. Girgin, M. Verzetti, D. Vincent, M. Kastelic, Z. Borsos, B. McWilliams, V . Ungureanu, O. Bachem, O. Pietquin et al., “Musicrl: Aligning music generation to hu- man preferences,” arXiv preprint arXiv:2402.04229, 2024

    2024 arXiv

  9. [17]

    Musiclm: Generating music from text,

    A. Agostinelli, T. I. Denk, Z. Borsos, J. Engel, M. Verzetti, A. Caillon, Q. Huang, A. Jansen, A. Roberts, M. Tagliasacchi et al., “Musiclm: Generating music from text,” arXiv preprint arXiv:2301.11325, 2023

    2023 arXiv

  10. [18]

    Weight normalization: A sim- ple reparameterization to accelerate training of deep neural net- works,

    T. Salimans and D. P. Kingma, “Weight normalization: A sim- ple reparameterization to accelerate training of deep neural net- works,” in Advances in Neural Information Processing Sys- tems, 2016, pp. 901–909

    2016

  11. [19]

    Layer normalization,

    J. L. Ba, J. R. Kiros, and G. E. Hinton, “Layer normalization,” arXiv preprint arXiv:1607.06450, 2016

    2016 arXiv

  12. [20]

    Fast and accurate deep network learning by exponential linear units (ELUs),

    D.-A. Clevert, T. Unterthiner, and S. Hochreiter, “Fast and accurate deep network learning by exponential linear units (ELUs),” in International Conference on Learning Represen- tations, 2016

    2016

  13. [21]

    Rectifier nonlinear- ities improve neural network acoustic models,

    A. L. Maas, A. Y . Hannun, and A. Y . Ng, “Rectifier nonlinear- ities improve neural network acoustic models,” inICML Work- shop on Deep Learning for Audio, Speech and Language Pro- cessing, 2013

    2013

  14. [23]

    Visqol v3: An open source production ready objective speech and audio metric,

    M. Chinen, F. S. C. Lim, J. Skoglund, N. Gureev, F. O’Gorman, and A. Hines, “Visqol v3: An open source production ready objective speech and audio metric,” 2020. [Online]. Available: https://arxiv.org/abs/2004.09584

    2020 arXiv

  15. [24]

    Mel- GAN: Generative adversarial networks for conditional wave- form synthesis,

    K. Kumar, R. Kumar, T. de Boissiere, L. Gestin, W. Z. Teoh, J. Sotelo, A. de Brebisson, Y . Bengio, and A. Courville, “Mel- GAN: Generative adversarial networks for conditional wave- form synthesis,” inAdvances in Neural Information Processing Systems, 2019

    2019

  16. [25]

    Hifi-gan: Generative adversarial networks for efficient and high fidelity speech synthesis,

    J. Kong, J. Kim, and J. Bae, “Hifi-gan: Generative adversarial networks for efficient and high fidelity speech synthesis,” in Advances in Neural Information Process- ing Systems , H. Larochelle, M. Ranzato, R. Hadsell, M. Balcan, and H. Lin, Eds., vol. 33. Curran As- sociates,...

    2020

  17. [26]

    Generative adversarial nets,

    I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde- Farley, S. Ozair, A. Courville, and Y . Bengio, “Generative adversarial nets,” in Neural Information Processing Systems (NIPS), 2014, pp. 2672–2680

    2014

  18. [27]

    A spectral energy distance for parallel speech synthesis,

    A. Gritsenko, T. Salimans, R. van den Berg, J. Snoek, and N. Kalchbrenner, “A spectral energy distance for parallel speech synthesis,” in Advances in Neural Information Process- ing Systems , H. Larochelle, M. Ranzato, R. Hadsell, M. Bal- can, and H. Lin, Eds., vol. 33. Curran...

    2020

  19. [28]

    Adam: A method for stochastic opti- mization

    D. P. Kingma and J. Ba, “Adam: A method for stochastic opti- mization.” in International Conference on Learning Represen- tations, Y . Bengio and Y . LeCun, Eds., 2015

    2015

  20. [30]

    Visqol: The virtual speech quality objective listener,

    A. Hines, J. Skoglund, A. Kokaram, and N. Harte, “Visqol: The virtual speech quality objective listener,” in IWAENC 2012; International Workshop on Acoustic Signal Enhance- ment, 2012

    2012

  21. [31]

    The MUSDB18 corpus for music separation,

    Z. Rafii, A. Liutkus, F.-R. St ¨oter, S. I. Mimilakis, and R. Bittner, “The MUSDB18 corpus for music separation,” Dec. 2017. [Online]. Available: https://doi.org/10.5281/ zenodo.1117372

    2017

  22. [2020]

    Available: https://arxiv.org/abs/2002.08933

    [Online]. Available: https://arxiv.org/abs/2002.08933

    2002 arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.