Pith. sign in

REVIEW 3 major objections 5 minor 34 references

Dual-BEATs: Unlocking Zero-Shot Stereo Audio Perception in Audio Large Language Models via Dithering

T0 review · 3 major / 5 minor · reviewed 2026-07-13 · grok-4.5

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

Pith's one-line read Uncorrelated dither lets ordinary dual audio encoders give multimodal LLMs native stereo localization without specialized spatial modules.

desk verdict Clean dual-encoder + uncorrelated dither result that really does rescue center-pan localization and zero-shot panning transfer; the overclaim is only that this equals generalized stereo understanding. read the letter →

arxiv 2607.08800 v1 pith:Z5GIDVFZ submitted 2026-07-09 cs.SD

Shuo-Chun Lin , Hen-Hsen Huang This is my paper
classification cs.SD
keywords Dual-BEATsstereoaudioperceptionditheringnormalizationbottleneckzero-shotspatiallocalizationmultimodalLLMsinter-channelleveldifferenceStochasticResonance
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

Multimodal audio LLMs understand what a sound is, but they are built on mono pipelines and therefore cannot hear where it is. The paper shows that this spatial blindness is not inevitable: simply routing left and right channels through two identical frozen semantic encoders fails only because internal normalization layers erase the tiny inter-channel amplitude differences that encode panning. Adding a static, uncorrelated dither noise floor before encoding creates a macro-variance that carries those differences through the normalizers, so ordinary cross-attention can read left-right geometry. On a ternary Left/Center/Right task the dithered models reach roughly 97 percent accuracy even at subtle 0.5 panning and generalize zero-shot to panning levels never seen in training. The practical claim is that, with this acoustic regularization, standard off-the-shelf multimodal models already possess generalized stereo understanding and do not need custom geometry-aware encoders.

What carries the argument

Dual-BEATs with uncorrelated dither: left and right waveforms each receive independent Gaussian noise scaled by a fixed dithering amplitude (DA=0.05), are passed through identical frozen BEATs encoders, feature-concatenated, and projected into the LLM; the uncorrelated noise floor "smuggles" spatial geometry across LayerNorm/RMSNorm.

What would settle it

Train and evaluate the same Dual-BEATs + dither pipeline on true binaural recordings that include HRTFs and multi-source overlap; if directional accuracy collapses to chance while the mono-panned condition remains high, the central claim that the method unlocks general stereo perception fails.

Watch

Extended reading notes

Core claim

Standard multimodal audio LLMs can perform continuous stereo localization once left and right channels are independently encoded and an uncorrelated dither floor is injected; the dither establishes a macro-variance that prevents internal normalizers from equalizing inter-channel amplitude differences, unlocking up to 97 percent ternary accuracy at subtle panning and robust zero-shot transfer to unseen spatial configurations.

Load-bearing premise

Deterministic single-source amplitude panning of mono clips is treated as a faithful enough stand-in for the spatial cues that real stereo or binaural audio would present to the same encoders and normalizers.

Editorial extensions

If this is right

  • Stereo localization becomes a lightweight adapter step rather than a full re-design of the audio encoder.
  • Existing mono-pretrained multimodal models can be given spatial awareness without discarding their semantic weights.
  • Zero-shot generalization across continuous panning amplitudes becomes the expected default once the noise floor is present.
  • The same dual-encoder-plus-dither pattern can be tested on larger backbones and denser token rates without new spatial pre-training.
  • Accessibility tools that need to name left-right events can reuse ordinary open multimodal stacks.

Reading between the lines

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

  • The same variance-floor idea may protect other continuous multimodal differentials that normalizers currently erase, such as mild spectral or temporal asymmetries.
  • If multi-source or true binaural tests succeed, the method could become a default preprocessing step for any audio LLM rather than a specialized research module.
  • Adversarial control of the dither seed is an immediate security surface once the technique moves to edge devices.
  • A lower DA schedule might shrink the observed semantic tax while still spanning the normalization threshold.
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. The paper proposes Dual-BEATs: left and right stereo channels are routed independently through two frozen identical BEATs encoders, feature-concatenated, and projected into an LLM (Gemma-3-1B or OLMo-3-7B) via QLoRA. The authors identify internal variance normalizers (LayerNorm/RMSNorm) as a bottleneck that erases subtle inter-channel level differences (ICLD), and introduce static uncorrelated Gaussian dither (DA=0.05) prior to encoding as a Stochastic Resonance bridge that preserves spatial geometry. On ternary Left/Center/Right classification of amplitude-panned AudioSet clips, dithered OLMo-3-7B reaches ~97% accuracy at PA=0.50 and shows continuous zero-shot generalization across unseen panning amplitudes (Fig. 2), while undithered models collapse to chance near center pan (Table 1). A modest semantic F1 tax is reported (Fig. 3). The authors conclude that standard multimodal models are natively capable of generalized stereo understanding once geometry is protected from normalization.

Significance. If the mechanism transfers beyond the synthetic proxy, the work offers a modular, encoder-agnostic path to stereo perception in audio LLMs without custom geometry-aware encoders or room simulation, which would be practically valuable for accessibility and scene analysis. Strengths include clean dithered-vs-undithered contrasts (Table 1), zero-shot heatmaps (Fig. 2), explicit seed isolation and uncorrelated-noise controls against shortcut learning (Appendix A.3–A.4), and transparent documentation of the semantic tax and limitations. The normalization-bottleneck diagnosis and signal-level dither intervention are concrete and falsifiable within the ICLD setting. The contribution is therefore of genuine interest to the audio-LLM community, provided claims are scoped to what the experiments actually demonstrate.

major comments (3)
  1. [Abstract, §1, §6 vs §5 / A.3] Abstract, §1, and §6 claim “true spatial abstraction,” “generalized stereo audio understanding,” and that standard models are “natively capable” of stereo perception. All positive results (Table 1, Fig. 2) are obtained exclusively under deterministic single-source amplitude panning of mono AudioSet clips (ICLD only; §3.3, A.3). There is no HRTF, ITD/phase, multi-source overlap, or real stereo/binaural evaluation. §5 correctly flags this as a limitation, but the framing in the abstract and conclusion overstates the evidence. Either (i) temper the central claim to “ICLD-based directional perception under amplitude panning” or (ii) add at least one transfer experiment on real stereo mixes or simple binaural renderings. Without one of these, the load-bearing “native generalized stereo” claim is not supported.
  2. [§3.2, §4.1, §5] The dither amplitude is fixed at DA=0.05 with no systematic ablation of the noise floor (only noted as future work in §5). Because the rescue effect is attributed to establishing a macro-variance floor that survives normalization, the result is sensitive to this free parameter. A minimal DA sweep (e.g., 0.01–0.10) on at least one backbone, reporting both directional accuracy and semantic F1, is needed to show that the Stochastic Resonance bridge is robust rather than a single-point hyperparameter success.
  3. [§2, §4] There is no head-to-head comparison against a specialized spatial encoder (e.g., Spatial-AST / BAT or a simple dual-channel baseline with phase features) on the same ternary task and AudioSet subset. Without this, it is hard to judge whether Dual-BEATs + dither is competitive or merely “works when specialized methods are unavailable.” A single controlled baseline would strengthen the accessibility claim in the introduction.
minor comments (5)
  1. [Table 1, §4.1] Table 1: Gemma-3-1B shows non-monotonic and sometimes degraded accuracy when dither is on at PA=0.00 (e.g., Direction-First 66.8% → 47.7%). A short discussion of noise-to-signal sensitivity for smaller models would help readers interpret the architecture’s capacity dependence.
  2. [Figure 3] Figure 3: the mono BEATs baseline (40.6%) and “Music; Speech” prior (28.73%) are useful anchors; state explicitly in the caption whether the mono baseline was trained under the same QLoRA recipe or is an off-the-shelf reference.
  3. [Throughout] Notation: P A and D A are written with spaces inconsistently (PA / P A, DA / D A). Standardize to PA and DA throughout.
  4. [Appendix A.1] Appendix A.1: the frozen March 2024 AudioSet snapshot is good for reproducibility; consider releasing the exact YouTube ID list or a DOI-backed artifact so others can reconstruct the 18,373 / 17,148 splits.
  5. [§2] Related work: “The World is Not Mono” [18] is cited; a one-sentence contrast with Dual-BEATs (dual semantic encoders + dither vs. heterogeneous semantic+spatial encoders) would clarify positioning.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; claims rest on empirical dithered-vs-undithered ablations and zero-shot matrices, not self-referential definitions or fitted predictions.

full rationale

The paper's load-bearing chain is architectural (Dual-BEATs feature-concat of two frozen BEATs encoders) plus a signal-level intervention (uncorrelated Gaussian dither of fixed DA=0.05 prior to encoding) followed by QLoRA instruction tuning and ternary directional evaluation on amplitude-panned AudioSet. Accuracy numbers (e.g., 97.1% at PA=0.50) and the continuous off-diagonal blocks in the zero-shot heatmaps of Figure 2 are measured experimental outcomes under explicit seed isolation (train seed 42, eval seed 1337) and uncorrelated noise, not quantities obtained by fitting a parameter on a subset and then 'predicting' a related quantity, nor by defining X in terms of Y. No uniqueness theorem or ansatz is imported via self-citation; references are to external works (BEATs, LayerNorm, Spatial-AST, etc.). The undithered baselines collapsing to chance (~33%) while dithered models do not constitutes an independent ablation, not a tautology. Limitations (ICLD-only panning) are stated but do not create circularity in the reported results. Honest non-finding: score 0, empty steps.

Assumptions & free parameters 3 free parameters · 4 assumptions · 2 invented entities

The claim rests on a small set of modeling choices (dual identical frozen BEATs, feature-concat, fixed DA, amplitude-panning proxy) plus standard LLM components. No new physical entities are postulated; the “Stochastic Resonance bridge” is a framing of ordinary additive Gaussian noise. Free parameters are few and explicitly stated.

free parameters (3)
  • Dithering Amplitude DA = 0.05
    Fixed at 0.05 (~-26 dB) for all primary results; chosen as a “robust initial baseline” without a reported sweep that would show the minimal viable floor or the semantic-tax trade-off surface.
  • Panning Amplitude PA training/eval grid = grid listed in Fig. 2
    Discrete set {0.00, 0.01, 0.05, 0.10, 0.25, 0.50, 0.80} used both for training single-PA models and for the zero-shot matrix; the continuous spatial claim is interpolated from these points.
  • QLoRA rank / alpha / target modules
    Adapter configuration (Table 5) is required for the LLM to learn to read the preserved inter-channel delta; exact rank values are not numerically listed in the provided text but are part of the adaptation that makes spatial tokens usable.
assumptions (4)
  • domain assumption Internal variance normalizers (LayerNorm/RMSNorm) act as aggressive compressors that erase subtle continuous inter-channel amplitude differentials unless a macro-variance floor is supplied.
    Stated as the core hypothesis in §3.2 and used to motivate dither; supported by the undithered collapse but not derived from first principles of the normalizer equations.
  • ad hoc to paper Deterministic amplitude panning of mono AudioSet sources produces inter-channel level differences that are a valid proxy for the spatial geometry present in real stereo or binaural recordings.
    Used for all training and evaluation (§3.3, A.3); explicitly listed as a limitation in §5.
  • domain assumption Uncorrelated Gaussian noise (NL ≠ NR) prevents shortcut learning via identity comparison while still floating the true ICLD across normalizers.
    Justified by early ablation observations in A.4 and by analogy to Stochastic Resonance; enforced by construction in the method.
  • domain assumption Frozen BEATs encoders retain sufficient semantic fidelity under the added dither for the downstream LLM to perform joint semantic+spatial generation.
    Assumed by keeping encoders frozen (§3.3); measured only via the residual semantic F1 after the “tax.”
invented entities (2)
  • Dual-BEATs architecture
    purpose: Route left and right channels through two identical frozen semantic encoders, feature-concatenate, and project into an LLM so that spatial geometry is available to cross-attention without a specialized spatial encoder.
    The concrete routing topology (Fig. 1) is introduced by the paper; independent evidence is the empirical accuracy tables themselves.
  • Stochastic Resonance bridge (uncorrelated dither floor)
    purpose: Supply an artificial macro-variance that normalizers lock onto, thereby preserving the delicate inter-channel amplitude delta.
    Framing of ordinary additive noise; the paper does not claim a new physical mechanism beyond the engineering intervention.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Dual-BEATs: Unlocking Zero-Shot Stereo Audio Perception in Audio Large Language Models via Dithering." pith.science (2026). https://pith.science/paper/Z5GIDVFZ

@misc{pith2026260708800,
  author       = {Pith},
  title        = {Pith review of: Dual-BEATs: Unlocking Zero-Shot Stereo Audio Perception in Audio Large Language Models via Dithering},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Z5GIDVFZ}},
  note         = {Machine review of arXiv:2607.08800}
}
read the original abstract

Multimodal Large Language Models (LLMs) have remarkable semantic audio understanding, yet they remain "spatially agnostic" due to their reliance on mono-channel audio representations. Currently, spatial audio perception methods mainly focus on complex room simulations and custom-trained, geometry-aware stereo encoders, which limits their accessibility and generalizability. In this paper, we introduce the Dual-BEATs architecture, in which the left and right audio channels are routed independently through two identical semantic encoders as an alternative to specialized spatial modules. To circumvent the architectural bottleneck where internal normalization otherwise erases the inter-channel variance of stereo audio, we inject a static, uncorrelated dithering noise floor prior to encoding. This dithering intervention establishes a macro-variance floor that "smuggles" spatial geometry across the normalization layers. Evaluated on a ternary directional classification task (Left, Center, Right), we demonstrate that dithered models achieve exceptional spatial resolution--reaching up to 97.2% localization accuracy even on subtle 0.5 panning amplitudes--and demonstrates robust, zero-shot generalization to entirely unseen spatial configurations. Our results suggest that with the appropriate acoustic regularization, standard multimodal models are natively capable of generalized stereo audio understanding.

Figures

Figures reproduced from arXiv: 2607.08800 by the authors.

Figure 1
Figure 1. The Dual-BEATs Architecture. Uncorrelated dithering noise ( [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Zero-shot spatial generalization heatmaps (List-First condition). Undithered baselines [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Zero-shot semantic retention evaluated under the strictly isolated “List-First” condition for [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

34 extracted references · 12 linked inside Pith

  1. [1]

    Audio flamingo: A novel audio language model with few-shot learning and dialogue abilities

    Arda Enver et al. Audio flamingo: A novel audio language model with few-shot learning and dialogue abilities. InForty-first International Conference on Machine Learning, 2024

    2024

  2. [2]

    Salmonn: Towards generic hearing abilities for large language models.arXiv preprint arXiv:2310.13289, 2023

    Changli Tang et al. Salmonn: Towards generic hearing abilities for large language models.arXiv preprint arXiv:2310.13289, 2023

    arXiv 2023

  3. [3]

    Qwen2-audio technical report.arXiv preprint arXiv:2407.10759, 2024

    Yunfei Chu, Jin Xu, Qian Yang, Xipin Wei, et al. Qwen2-audio technical report.arXiv preprint arXiv:2407.10759, 2024

    arXiv 2024

  4. [4]

    Music flamingo: Scaling music understanding in audio language models

    Sreyan Ghosh, Arushi Goel, Lasha Koroshinadze, Sang-gil Lee, Zhifeng Kong, João Felipe Santos, Ramani Duraiswami, Dinesh Manocha, Wei Ping, Mohammad Shoeybi, and Bryan Catanzaro. Music flamingo: Scaling music understanding in audio language models. InThe Fourteenth International Conference on Learning Representations, 2026. URLhttps://openreview.net/forum...

    2026

  5. [5]

    Robust speech recognition via large-scale weak supervision

    Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. Robust speech recognition via large-scale weak supervision. InProceedings of the 40th International Conference on Machine Learning, volume 202 ofProceedings of Machine Learning Research, pages 28492–28518. PMLR, 2023

    2023

  6. [6]

    CLAP learning audio concepts from natural language supervision

    Benjamin Elizalde, Soham Deshmukh, Mahmoud Al Ismail, and Huaming Wang. CLAP learning audio concepts from natural language supervision. InICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1–5. IEEE, 2023

    2023

  7. [7]

    Gpt-4 technical report.arXiv preprint arXiv:2303.08774, 2023

    OpenAI. Gpt-4 technical report.arXiv preprint arXiv:2303.08774, 2023

    arXiv 2023

  8. [8]

    Gpt-4o system card.arXiv preprint arXiv:2410.21276, 2024

    OpenAI. Gpt-4o system card.arXiv preprint arXiv:2410.21276, 2024

    arXiv 2024

Show all 34 references
  1. [9]

    Gemini: A family of highly capable multimodal models.arXiv preprint arXiv:2312.11805, 2023

    Gemini Team et al. Gemini: A family of highly capable multimodal models.arXiv preprint arXiv:2312.11805, 2023

    2023 arXiv

  2. [10]

    Spatial-ast: Binaural audio spectrogram transformer.arXiv preprint arXiv:2402.01591, 2024

    Puyuan Peng et al. Spatial-ast: Binaural audio spectrogram transformer.arXiv preprint arXiv:2402.01591, 2024

    2024 arXiv

  3. [11]

    OWL: Geometry-aware spatial reasoning for audio large language models

    Subrata Biswas, Mohammad Nur Hossain Khan, and Bashima Islam. OWL: Geometry-aware spatial reasoning for audio large language models. InThe Fourteenth International Conference on Learning Representations, 2026. URLhttps://openreview.net/forum?id=zPv46YKv3w

    2026

  4. [12]

    Beats: Audio pre-training with acoustic tokenizers

    Sanyuan Chen et al. Beats: Audio pre-training with acoustic tokenizers. InProceedings of the Fortieth International Conference on Machine Learning, 2023

    2023

  5. [13]

    Layer normalization.arXiv preprint arXiv:1607.06450, 2016

    Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization.arXiv preprint arXiv:1607.06450, 2016

    2016 arXiv

  6. [14]

    Root mean square layer normalization.Advances in Neural Information Processing Systems, 32, 2019

    Biao Zhang and Rico Sennrich. Root mean square layer normalization.Advances in Neural Information Processing Systems, 32, 2019

    2019

  7. [15]

    GAMA: A large audio-language model with advanced audio understanding and complex reasoning abilities

    Sreyan Ghosh, Sonal Kumar, Ashish Seth, Chandra Kiran Reddy Evuru, Utkarsh Tyagi, S Sakshi, Oriol Nieto, Ramani Duraiswami, and Dinesh Manocha. GAMA: A large audio-language model with advanced audio understanding and complex reasoning abilities. InProceedings of the 2024 Confe...

    2024

  8. [16]

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

    Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. InAdvances in Neural Information Processing Systems, volume 33, pages 12449–12460, 2020

    2020

  9. [17]

    Hubert: Self-supervised speech representation learning by masked prediction of hidden units.IEEE/ACM Transactions on Audio, Speech, and Language Processing, 29:3451–3460, 2021

    Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrahman Mohamed. Hubert: Self-supervised speech representation learning by masked prediction of hidden units.IEEE/ACM Transactions on Audio, Speech, and Language Processing, 29:3...

    2021

  10. [18]

    The world is not mono: Enabling spatial understanding in large audio-language models.arXiv preprint arXiv:2601.02954, 2026

    Yuhuan You, Lai Wei, Xihong Wu, and Tianshu Qu. The world is not mono: Enabling spatial understanding in large audio-language models.arXiv preprint arXiv:2601.02954, 2026

    2026 arXiv

  11. [19]

    ASAudio: A survey of advanced spatial audio research.arXiv preprint arXiv:2508.10924, 2025

    Zhiyuan Zhu, Yu Zhang, Wenxiang Guo, Changhao Pan, and Zhou Zhao. ASAudio: A survey of advanced spatial audio research.arXiv preprint arXiv:2508.10924, 2025

    2025 arXiv

  12. [20]

    Stereo sound event localization and detection with onscreen/offscreen classification

    Kazuki Shimada et al. Stereo sound event localization and detection with onscreen/offscreen classification. arXiv preprint arXiv:2507.12042, 2025

    2025 arXiv

  13. [21]

    Can large language models understand spatial audio?arXiv preprint arXiv:2406.07914, 2024

    Changli Tang, Wenyi Yu, Guangzhi Sun, Xianzhao Chen, Tian Tan, Wei Li, Jun Zhang, Lu Lu, Zejun Ma, Yuxuan Wang, and Chao Zhang. Can large language models understand spatial audio?arXiv preprint arXiv:2406.07914, 2024

    2024 arXiv

  14. [22]

    SA VVY: Spatial awareness via audio-visual LLMs through seeing and hearing

    Mingfei Chen, Zijun Cui, Xiulong Liu, Jinlin Xiang, Caleb Zheng, Jingyuan Li, and Eli Shlizerman. SA VVY: Spatial awareness via audio-visual LLMs through seeing and hearing. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems, 2026. URL https://openre...

    2026

  15. [23]

    Phasecoder: Microphone geometry-agnostic spatial audio understanding for multimodal llms.arXiv preprint arXiv:2601.21124, 2026

    Artem Dementyev, Wazeer Zulfikar, Sinan Hersek, Pascal Getreuer, Anurag Kumar, and Vivek Kumar. Phasecoder: Microphone geometry-agnostic spatial audio understanding for multimodal llms.arXiv preprint arXiv:2601.21124, 2026

    2026

  16. [24]

    Spatial audio question answering and reasoning on dynamic source movements.arXiv preprint arXiv:2509.14666, 2025

    Arvind Krishna Sridhar, Yinyi Guo, and Erik Visser. Spatial audio question answering and reasoning on dynamic source movements.arXiv preprint arXiv:2509.14666, 2025

    2025

  17. [25]

    Bartoldson, Bhavya Kailkhura, Avi Schwarzschild, Aniruddha Saha, Micah Goldblum, Jonas Geiping, and Tom Goldstein

    Neel Jain, Ping-yeh Chiang, Yuxin Wen, John Kirchenbauer, Hong-Min Chu, Gowthami Somepalli, Brian R. Bartoldson, Bhavya Kailkhura, Avi Schwarzschild, Aniruddha Saha, Micah Goldblum, Jonas Geiping, and Tom Goldstein. NEFTune: Noisy embeddings improve instruction finetuning. InT...

    2024

  18. [26]

    Towards evaluating the robustness of neural networks

    Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. In2017 ieee symposium on security and privacy (sp), pages 39–57. IEEE, 2017

    2017

  19. [27]

    The mechanism of stochastic resonance.Journal of Physics A: mathematical and general, 14(11):L453, 1981

    Roberto Benzi, Alfonso Sutera, and Angelo Vulpiani. The mechanism of stochastic resonance.Journal of Physics A: mathematical and general, 14(11):L453, 1981

    1981

  20. [28]

    Dither in digital audio.Journal of the Audio Engineering Society, 35(12):966–975, 1987

    John Vanderkooy and Stanley P Lipshitz. Dither in digital audio.Journal of the Audio Engineering Society, 35(12):966–975, 1987

    1987

  21. [29]

    Sigmoid-weighted linear units for neural network function approximation in reinforcement learning.Neural Networks, 107:3–11, 2018

    Stefan Elfwing, Eiji Uchibe, and Kenji Doya. Sigmoid-weighted linear units for neural network function approximation in reinforcement learning.Neural Networks, 107:3–11, 2018

    2018

  22. [30]

    Gaussian error linear units (gelus).arXiv preprint arXiv:1606.08415, 2016

    Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus).arXiv preprint arXiv:1606.08415, 2016

    2016 arXiv

  23. [31]

    Shortcut learning in deep neural networks.Nature Machine Intelligence, 2(11):665–673, 2020

    Robert Geirhos, Jörn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Felix A Wichmann. Shortcut learning in deep neural networks.Nature Machine Intelligence, 2(11):665–673, 2020

    2020

  24. [32]

    Gemmeke, Daniel P

    Jort F. Gemmeke, Daniel P. W. Ellis, Dylan Freedman, Aren Jansen, Wade Lawrence, R. Channing Moore, Manoj Plakal, and Marvin Ritter. Audio set: An ontology and human-labeled dataset for audio events. In2017 IEEE International Conference on Acoustics, Speech and Signal Processi...

    2017 doi

  25. [33]

    QLoRA: Efficient finetuning of quantized LLMs

    Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. QLoRA: Efficient finetuning of quantized LLMs. InThirty-seventh Conference on Neural Information Processing Systems, 2023. URL https://openreview.net/forum?id=OUIFPHEgJU

    2023

  26. [34]

    Music” and “Speech

    Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. InInternational Conference on Learning Representations, 2022. URLhttps://openreview.net/forum?id=nZeVKeeFYf9. 11 A...

    2022

Pith tools

Reviewed July 13, 2026 · model on record in the stance chip above.