Pith. sign in

REVIEW 5 major objections 6 minor 17 references

Teaching Physical Awareness to LLMs through Sounds

T0 review · 5 major / 6 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 By training on audio convolved with simulated physical channels, ACORN claims to give LLMs the ability to detect line-of-sight paths, estimate Doppler shifts, sound direction, multipath severity, and range, with that ability surviving…

desk verdict A sensible first step toward giving LLMs acoustic physical awareness, but the evaluation does not yet rule out level/energy cues; worth refereeing with requests for baselines and stronger real-world tests. read the letter →

arxiv 2506.08524 v2 pith:HHYNX4CG submitted 2025-06-10 cs.SD cs.AIcs.MMcs.ROeess.AS

Weiguo Wang , Andy Nie , Wenrui Zhou , Yi Kai , Chengchen Hu This is my paper
classification cs.SDcs.AIcs.MMcs.ROeess.AS
keywords physicalawarenessaudiolargelanguagemodelschannelimpulseresponsesimulationDopplereffectestimationdirectionofarrivalmultipathanalysiszero-shotreal-worldtransferquestionanswering
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

The paper is trying to establish that a large language model can acquire physical awareness from sound without needing a large corpus of real, instrumented acoustic recordings. Its recipe is to split any received sound into a sound source and a physical channel, simulate the channel with signal-processing models for direct paths, reflections, reverberation, Doppler shifts, and microphone delay, and convolve those simulated channels with real audio clips to generate an annotated training set of a million question-answer pairs. A phase-aware audio encoder feeds magnitude and phase into the LLM, on the argument that physical effects such as arrival times and frequency shifts live in phase relationships that magnitude-only encoders overlook. The reported results show accuracy above 0.9 on several simulated sensing tasks and above-chance zero-shot accuracy on real recordings inside a vehicle cabin for line-of-sight detection and left/right direction. If the claim holds, it offers a scalable route to giving voice assistants, vehicles, and robots a sense of the acoustic environment around them.

What carries the argument

The central object is the channel impulse response $h$ in the convolution $y = h \circledast s$, which represents everything the physical environment does to a source signal $s$ before it reaches the microphone. The simulator builds $h$ from damped-sinusoid impulse paths for the direct and reflected components, subband-filtered Gaussian noise with exponential decay for reverberation, a time-varying delay resampled at rate $1 - v/c$ for the Doppler effect, and inter-microphone delays for direction of arrival. The audio encoder's load-bearing novelty is to pass the STFT magnitude together with sine and cosine of the phase as three input channels, so the LLM can see the timing and phase relationships that carry physical cues rather than only the spectral content of the source.

What would settle it

Collect a real-room or real-vehicle test set with independently measured ground truth for line-of-sight, source velocity, direction, multipath severity, and range, train ACORN only on simulated channels as described, and measure zero-shot accuracy; if any task drops to chance (for example, line-of-sight accuracy near 0.5), the simulator is not carrying that physical cue. A second decisive check is to set the two phase branches of the audio encoder to zero and retrain; if all five metrics are unchanged, the phase mechanism is not the active ingredient.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that an LLM can be taught to read physical scene properties from audio by supervised fine-tuning on simulator-generated data, and that the phase of the spectrum is the key information channel. The ACORN model, built by adding a phase-aware audio encoder to Llama3.1-8B or Qwen2-7B and fine-tuning with LoRA on AQA-PHY, reports 0.924 binary accuracy for line-of-sight detection, 0.181 mean absolute error in Doppler shift percentage, 0.907 mean absolute error in time-difference-of-arrival samples, 0.903 accuracy for multipath severity, and 1.599 percent relative error for range estimation on simulated test sets. In zero-shot tests inside a real vehicle cabin, the same simulator-only models report 0.845 to 0.870 accuracy for detecting whether the sound source has a direct path and 0.8975 to 0.925 accuracy for left/right direction.

Load-bearing premise

The load-bearing premise is that the simulator's simplified channels—short synthetic echoes, noise-based reverberation, and a constant-speed Doppler resampling—are close enough to real sound propagation that a model trained only on them keeps working on real recordings in a vehicle cabin.

Editorial extensions

If this is right

  • Physical-awareness training can be scaled cheaply: any existing audio corpus can be convolved with parameterized channel models to synthesize labeled examples, avoiding the need to instrument real rooms and vehicles for every new task.
  • Voice-controlled systems could use line-of-sight detection to reject commands originating outside a secured space, such as a vehicle cabin or a smart home.
  • The encoder's benefit appears across two different 7B-to-8B language models, suggesting the phase-aware representation rather than a particular LLM carries the improvement.
  • Open-form question answering shows the model can not only classify but also perform multi-step physical calculations, converting measured delays into angles or distances.
  • Active range estimation from simulated chirp echoes points toward LLMs acting as sonar-like sensors that transmit a pulse, listen for the reflection, and compute the distance.

Reading between the lines

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

  • The real-world zero-shot results leave open whether the model has learned general acoustic physics or a narrow statistical match to the simulator; recording real impulse responses with independent ground truth and testing the same model on them would separate the two.
  • The phase-and-simulation recipe is likely portable to channel-borne cues the paper does not test, such as source height, reflecting surface material, moving scatterers, or non-uniform sound speed; finding these work would support the broader physical-awareness reading.
  • Because the training sources come from one large general audio dataset, performance may degrade for acoustic source types it never saw; a test using rarely represented or synthetic sources would bound how much of the claimed awareness is source-specific.
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

5 major / 6 minor

Summary. The paper presents ACORN, a framework that teaches LLMs physical awareness from audio by decomposing a received sound into a source and a physical channel. It builds a channel simulator covering LOS paths, early reflections, reverberation, Doppler, and microphone-array TDoA, synthesizes a 1M-pair audio QA dataset (AQA-PHY) by convolving AudioSet sources with simulated CIRs, and trains a Whisper-initialized phase-aware audio encoder with Llama3.1-8B and Qwen2-7B via LoRA on closed- and open-form QA pairs. On five simulated tasks the paper reports strong point estimates, including 0.924 LOS detection accuracy, 0.181 Doppler MAE, 0.907 DoA MAE, 0.903 multipath accuracy, and 1.599 range REP, plus zero-shot real-vehicle accuracies of 0.845-0.870 for LOS detection and 0.8975-0.925 for left/right direction.

Significance. If the sim-to-real generalization claim held, this would be a useful step toward audio-grounded physical reasoning in LLMs, with concrete applications in voice-command security and embodied AI. The paper has real strengths: the explicit source/channel decomposition, a phase-aware (sin/cos) encoder rather than magnitude-only processing, a large synthetic QA corpus, open-form multi-step reasoning examples, and a genuine zero-shot real-vehicle deployment. The main credibility gap is that the current evidence does not yet rule out exploitation of level or energy cues, and the simulated evaluation does not compare against classical physics-based estimators. These gaps are addressable and should be fixed before the central claim is accepted.

major comments (5)
  1. [§4.1, Table 4; §6.2, Table 2] The central claim that ACORN imparts physical awareness is not yet separable from the claim that the model fits the simulator's own channel statistics. Because the same simulator (Section 4) generates both the training data and the held-out labels for Table 2, high simulated scores partly measure how well the model inverts the simulator's channel model. This concern is aggravated by the parameter ranges in Table 4: LOS amplitudes are 5.0-20.0 while early-reflection amplitudes are 0.1-0.5, so the LOS/non-LOS distinction is available as a 10-200x energy or onset-level gap. The evaluation should include controls that remove this cue, such as per-sample gain randomization, amplitude normalization, or a first-arrival energy-ratio baseline, to show that the model uses delay and multipath structure rather than level cues.
  2. [§6.4, Appendix F] The real-world evidence is limited to two binary tasks, LOS detection and left/right classification, and both can be solved by level or proximity cues. The four microphones are placed at fixed left/right positions in the cabin, and no gain calibration or loudness equalization is reported; a speaker placed near a microphone will produce a larger signal in that channel, making left/right detection possible through inter-microphone level differences alone. Similarly, interior/exterior LOS detection can be answered by overall loudness or low-frequency attenuation. The zero-shot vehicle results are a useful sanity check, but they do not yet demonstrate physical channel understanding. Please add real-world Doppler, multipath, or range measurements, or at least a symmetric loudness-matched left/right protocol with per-condition accuracy.
  3. [§6.1-6.2, Table 2] The empirical evaluation omits essential methodology: there is no train/test split description, no indication of whether the same AudioSet source clips appear in both training and test utterances, no error bars or seed variance, and no statistical significance testing. The claims that the ACORN encoder is superior (e.g., 0.924 vs 0.867 BCA with Llama, 0.181 vs 1.042 Doppler MAE with Qwen) and model-agnostic rest on single point estimates. Reporting means and standard deviations over at least three seeds, with a source-disjoint split, is necessary to make Table 2 interpretable.
  4. [§6.1, Table 2; §7] There is no comparison against classical physics-based estimators, which are the natural and much cheaper baselines for these tasks: cross-correlation TDoA for DoA, spectral-peak or phase-derivative frequency-shift estimation for Doppler, and FMCW or matched-filter peak detection for range. Given that the paper's value proposition is that an LLM can learn physical awareness from data, Table 2 should include these reference numbers; otherwise the absolute values are difficult to contextualize and the reader cannot tell whether the LLM adds value over well-understood signal processing.
  5. [Appendix H; §8] The simulator-fidelity validation is qualitative: Appendix H displays the nearest retrieved synthetic CIR for a few real CIRs, but nearest-neighbor retrieval is by construction biased toward similarity and the figures are not quantified. Since the simulator's representativeness is load-bearing for the sim-to-real claim, the paper should provide quantitative comparisons with real recordings, for example RT60 distributions, energy-decay curves, or per-component CIR error on held-out real measurements. The Limitations section explicitly acknowledges the limited real-world scope, and this comment is a request to close exactly that gap.
minor comments (6)
  1. [Appendix B.1] The heading 'LOS ad Early Reflections' should read 'LOS and Early Reflections'.
  2. [§6.1] The phrase 'window size of 254 1' appears to be a typo; specify the units (samples) and the intended window length.
  3. [§6.4] The sentence 'Figure 8 presents the experimental results' should reference Figure 9, since Figure 8 is the SNR ablation.
  4. [Table 10] In the DoA open-form example, the formula reads 'θ = arccos((π× c) / d)', but the subsequent numeric substitution uses the delay τ; replace π with τ.
  5. [Appendix B.3] 'For connivent' should read 'For convenience'.
  6. [§3.2] The phrase '3x3 1D convolution' mixes 2D and 1D terminology; please clarify the exact kernel shape and the tensor dimensions being convolved.

Circularity Check

0 steps flagged · score 0.0 of 10

No load-bearing circularity: the simulated training/evaluation loop is self-consistent by design, and the real-world zero-shot tests provide independent grounding for the central transfer claim.

full rationale

ACORN's derivation chain is a standard source/channel decomposition (Eq. 1), an explicit channel simulator (Eqs. 4-17), and a supervised SFT dataset whose labels are the simulator's physically meaningful parameters (Section 4.5, Table 1). Training and evaluating on held-out samples from the same simulator is a conventional synthetic benchmark; it tests whether the model can decode the simulator's channel model, and the paper does not present those simulated numbers as if they were independent physical evidence. The real-world zero-shot experiments in Section 6.4 (LOS detection at 0.845-0.870 and left/right DoA at 0.8975-0.925) provide an external check that breaks any circularity for the central transfer claim. The only self-citations (Wang et al. 2020/2022) support a standard far-field TDoA formula and contextualize prior acoustic-sensing work; they are not load-bearing for ACORN's main results. No fitted parameter is relabeled as a prediction, no uniqueness theorem is imported from the authors, and no target quantity is defined in terms of the model's own output. The acknowledged limitations in Section 8 and the qualitative real-vs-synthetic CIR comparison in Appendix H are validation-coverage concerns, not circularity: the Appendix H retrieval is illustrative, not the basis of any metric, and the limitations simply state that more real-world testing is needed. Therefore no specific circular step is exhibited, and the derivation chain is self-contained apart from its stated dependence on the simulator's own physical assumptions.

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

No new physical entities are introduced; the novelties are a simulator, a dataset, and an encoder. The free parameters are the hand-chosen distributions of the simulator, which directly determine the ground-truth labels. The axioms are the physical simplifications that make the simulator tractable, and all of them are potential sources of sim-to-real mismatch.

free parameters (4)
  • Simulator subband decay rate λ_i (reverberation) = Uniform 1.5 to 60 (Table 5)
    Hand-chosen range defines simulated RT60 and the multipath Rich/Moderate/Negligible labels; training and simulated evaluation share this distribution.
  • LOS and early-reflection path parameters = LOS amplitude 5-20, early amplitude 0.1-0.5, decay modification 1.0-1.5, base frequency 2000 Hz (Table 4)
    Damped-sinusoid paths are a hand-designed simplification of real CIRs; all LOS and echo labels derive from these parameter draws.
  • Inter-microphone spacing d_mic = 8-15 cm uniform (Table 7)
    Sets the TDoA scale for direction-of-arrival labels; no calibration against the real vehicle microphone geometry is reported.
  • Doppler relative speed and initial distance ranges = v in [-50, 50] m/s, d0 in [0.5, 100] m (Table 6)
    Defines the Doppler label distribution; reported Doppler accuracy is only within this synthetic range.
assumptions (4)
  • domain assumption Received sound is exactly the convolution of an independent source with a linear channel: y = h ⊛ s (Eq. 1).
    Underlies the entire simulator and dataset; real-world audio is assumed separable into clean source and channel, ignoring nonlinearities, source-channel coupling, and device noise.
  • domain assumption Reverberation is representable as bandpass-filtered Gaussian noise subbands with exponential decay (Eq. 13).
    This simplified model ignores true room geometry, modal resonances, and direction-dependent reflections; if it misses cues, the model will not generalize to real spaces.
  • domain assumption Doppler is fully captured by time-domain resampling with factor (1 - v/c) at constant relative velocity (Eq. 17).
    Assumes constant velocity and no frequency-dependent propagation; real broadband acoustic Doppler in moving vehicles is more complex.
  • domain assumption Far-field plane-wave TDoA model holds for all DoA data: τθ = d cos θ / c (Eq. 6).
    Used to generate DoA labels and to evaluate; the real-vehicle left/right test is a much simpler binary proxy of this assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Teaching Physical Awareness to LLMs through Sounds." pith.science (2026). https://pith.science/paper/HHYNX4CG

@misc{pith2026250608524,
  author       = {Pith},
  title        = {Pith review of: Teaching Physical Awareness to LLMs through Sounds},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HHYNX4CG}},
  note         = {Machine review of arXiv:2506.08524}
}
read the original abstract

Large Language Models (LLMs) have shown remarkable capabilities in text and multimodal processing, yet they fundamentally lack physical awareness--understanding of real-world physical phenomena. In this work, we present ACORN, a framework that teaches LLMs physical awareness through sound, focusing on fundamental physical phenomena like the Doppler effect, multipath effect, and spatial relationships. To overcome data scarcity, ACORN introduce a physics-based simulator combining real-world sound sources with controlled physical channels to generate diverse training data. Using this simulator, we build AQA-PHY, a comprehensive Audio Question-Answer dataset, and propose an audio encoder that processes both magnitude and phase information. By connecting our audio encoder to state-of-the-art LLMs, we demonstrate reasonable results in both simulated and real-world tasks, such as line-of-sight detection, Doppler effect estimation, and Direction-of-Arrival estimation, paving the way for enabling LLMs to understand physical world.

Figures

Figures reproduced from arXiv: 2506.08524 by the authors.

Figure 1
Figure 1. Model architecture and audio encoder design . (a) Model architecture: Integrates audio and text encoders to generate unified token sequences for LLMs, enabling physical reasoning (e.g., detecting LOS). (b) Audio encoder: Processes magnitude and phase components of audio to capture physical phenomena such as Doppler effects and multipath. 3.2. Audio Encoder To effectively understand complex physical phenom￾ena—such a… view at source ↗
Figure 2
Figure 2. Illustration of Acoustic Channel. the physical phenomena, isolating them from source s. Our channel simulator models five main components: (1) LOS paths and (2) early reflections, (3) reverberation, (4) Doppler effects, and (5) microphone array reception. Each can be independently controlled to generate diverse physical phenomena. For more details, please refer to Appendix B. 4.1. LOS Path and Early Reflections We b… view at source ↗
Figure 3
Figure 3. Microphone Array Model. 4.4. Microphone Array Sound Direction, formally known as Direction of Arrival (DoA), can be estimated by measuring the Time Difference of Arrival (TDoA) between microphones. Due to their spa￾tial separation, each microphone receives sound waves at slightly different times, enabling DoA estimation [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Synthesis Workflow of <Audio, Question, Answer>. For Doppler Estimation, we vary source speeds to generate different frequency shifts. For DoA Estimation, we adjust arrival directions and microphone spacing to create TDoA variations. For Multipath Analysis, we modify d…
Figure 5
Figure 5. Figure 5: Loss History. chitecture. The model is trained on 4 NVIDIA A100 GPUs with batch size 32 and completes after 7 epochs. The total training time is about 61 hours. For response generation, the decoding parameters are set with temperature 1, top-p 1, and top-k 50. Appendix…
Figure 6
Figure 6. Figure 6: Impact of LOS. W/ Doppler W/O Doppler 0.5 0.6 0.7 0.8 0.9 1.0 BCA Llama Qwen [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 10
Figure 10. Figure 10: Ranging Estimation Illustration. A.2. Active Sensing Tasks Range Estimation explores active sensing capabilities, requiring LLMs to initiate measurements by transmitting sound pulses and analyzing their echoes. Unlike passive tasks that only observe existing sounds, t…
Figure 11
Figure 11. Figure 11: Impact of Doppler Speed on LOS detection. F. Real-World Deployment The data collection setup, as illustrated in [PITH_FULL_IMAGE:figures/full_fig_p017_11.png]
Figure 12
Figure 12. Figure 12: Illustration of microphone and speaker deployment in the NIO ES6 vehicle. Microphones (Mic. 1–4) are positioned in the front and back seat areas, while the speaker is placed at various locations for replaying audios. <|im_end|>{\n} <|im_start|>assistant{\n} {Answer} <…
Figure 13
Figure 13. Figure 13: Comparison between real and synthesized CIRs in both LOS and NLOS scenarios. For each example, we select a real CIR (left) and retrieve the most similar synthesized CIR (right) using a similarity search. Two LOS and two NLOS cases are shown to demonstrate the diversit…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

17 extracted references · 3 canonical work pages

  1. [1]

    Do as i can, not as i say: Grounding language 9 Teaching Physical Awareness to LLMs from Sounds in robotic affordances.arXiv preprint arXiv:2204.01691,

    Ahn, M., Brohan, A., Brown, N., Chebotar, Y ., Cortes, O., David, B., Finn, C., Fu, C., Gopalakrishnan, K., Hausman, K., et al. Do as i can, not as i say: Grounding language 9 Teaching Physical Awareness to LLMs from Sounds in robotic affordances.arXiv preprint arXiv:2204.01691,

    arXiv

  2. [5]

    S., Lynch, C., Chowdhery, A., Ichter, B., Wahid, A., Tompson, J., Vuong, Q., Yu, T., et al

    Driess, D., Xia, F., Sajjadi, M. S., Lynch, C., Chowdhery, A., Ichter, B., Wahid, A., Tompson, J., Vuong, Q., Yu, T., et al. Palm-e: An embodied multimodal language model. arXiv preprint arXiv:2303.03378,

    arXiv

  3. [6]

    The llama 3 herd of models.arXiv preprint arXiv:2407.21783,

    Dubey, A., Jauhri, A., Pandey, A., Kadian, A., Al-Dahle, A., Letman, A., Mathur, A., Schelten, A., Yang, A., Fan, A., et al. The llama 3 herd of models.arXiv preprint arXiv:2407.21783,

    arXiv

  4. [7]

    H., Karlinsky, L., and Glass, J

    Gong, Y ., Luo, H., Liu, A. H., Karlinsky, L., and Glass, J. Listen, think, and understand.arXiv preprint arXiv:2305.10790,

    arXiv

  5. [8]

    Deepseek-r1: In- centivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948,

    Guo, D., Yang, D., Zhang, H., Song, J., Zhang, R., Xu, R., Zhu, Q., Ma, S., Wang, P., Bi, X., et al. Deepseek-r1: In- centivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948,

    arXiv

  6. [10]

    This pulse travels along two paths: a direct Line-of-Sight (LOS) path to the microphone, and a reflected ECHO path that bounces off the wall

    As shown in Figure 10(a), the basic deployment involves a system that actively transmits a sound pulse. This pulse travels along two paths: a direct Line-of-Sight (LOS) path to the microphone, and a reflected ECHO path that bounces off the wall. This configuration mirrors the principles used in natural and technological sonar systems. The spectrogram in F...

    work page 2017

  7. [11]

    10 Teaching Physical Awareness to LLMs from Sounds Radford, A., Kim, J

    Accessed: 2024-12-01. 10 Teaching Physical Awareness to LLMs from Sounds Radford, A., Kim, J. W., Xu, T., Brockman, G., McLeavey, C., and Sutskever, I. Robust speech recognition via large- scale weak supervision. InProceedings of the 40th In- ternational Conference on Machine Learning (ICML),

    work page 2024

  8. [13]

    Qwen2 technical report.arXiv preprint arXiv:2407.10671,

    Yang, A., Yang, B., Hui, B., Zheng, B., Yu, B., Zhou, C., Li, C., Li, C., Liu, D., Huang, F., et al. Qwen2 technical report.arXiv preprint arXiv:2407.10671,

    arXiv

Show all 17 references
  1. [14]

    Speechgpt: Empowering large language models with intrinsic cross-modal conversational abilities

    Zhang, D., Li, S., Zhang, X., Zhan, J., Wang, P., Zhou, Y ., and Qiu, X. Speechgpt: Empowering large language models with intrinsic cross-modal conversational abilities. arXiv preprint arXiv:2305.11000,

    arXiv

  2. [15]

    Swift: a scalable lightweight infrastructure for fine-tuning.arXiv preprint arXiv:2408.05517,

    Zhao, Y ., Huang, J., Hu, J., Wang, X., Mao, Y ., Zhang, D., Jiang, Z., Wu, Z., Ai, B., Wang, A., et al. Swift: a scalable lightweight infrastructure for fine-tuning.arXiv preprint arXiv:2408.05517,

    arXiv

  3. [17]

    Ablation Study of LoRA Rank We keep the LoRA rank relatively low, 8, to preserve the LLM’s language ability, as most representation learning occurs in the audio encoder

    Table 8.Training Hyperparameters Parameter Value GPUs 4 NVIDIA A100 Global Batch Size 32 Epochs 7 Optimizer AdamW Optimizer Parametersβ 1 = 0.9,β 2 = 0.95,ϵ= 1e-8 Learning Rate Schedule WarmupDecayLR Weight Decay 0.1 Warm-up Min Learning Rate 0 Warm-up Max Learning Rate 0.0001...

  4. [2016]

    Simlm: Can language models infer parameters of physical systems?arXiv preprint arXiv:2312.14215,

    Memery, S., Lapata, M., and Subr, K. Simlm: Can language models infer parameters of physical systems?arXiv preprint arXiv:2312.14215,

    arXiv

  5. [2020]

    Wang, W., Li, J., He, Y ., Guo, X., and Liu, Y . Motorbeat: Acoustic communication for home appliances via vari- able pulse width modulation.Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technolo- gies, 6(1):1–24, 2022a. Wang, W., Li, J., He, Y ., and ...

    arXiv

  6. [2022]

    Llmphy: Complex physical reasoning using large lan- guage models and world models.arXiv preprint arXiv:2411.08027,

    Cherian, A., Corcodel, R., Jain, S., and Romeres, D. Llmphy: Complex physical reasoning using large lan- guage models and world models.arXiv preprint arXiv:2411.08027,

    arXiv

  7. [2023]

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

    Chu, Y ., Xu, J., Yang, Q., Wei, H., Wei, X., Guo, Z., Leng, Y ., Lv, Y ., He, J., Lin, J., et al. Qwen2-audio technical report.arXiv preprint arXiv:2407.10759,

    arXiv

  8. [2024]

    Qwen-audio: Advancing univer- sal audio understanding via unified large-scale audio- language models.arXiv preprint arXiv:2311.07919,

    Chu, Y ., Xu, J., Zhou, X., Yang, Q., Zhang, S., Yan, Z., Zhou, C., and Zhou, J. Qwen-audio: Advancing univer- sal audio understanding via unified large-scale audio- language models.arXiv preprint arXiv:2311.07919,

    arXiv

  9. [2025]

    and Gimpel, K

    Hendrycks, D. and Gimpel, K. Gaussian error linear units (gelus).arXiv preprint arXiv:1606.08415,

    arXiv

Pith tools

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