Pith. sign in

REVIEW 3 major objections 5 minor 31 references

Grounding fine-tuning improves how decoders read event evidence that already exists in audio tokens.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 02:39 UTC pith:YLZJPUF6

load-bearing objection Useful diagnostic case study, but the readout result needs a random-state control and the token-to-time mapping an honest validation. the 3 major comments →

arxiv 2607.25355 v1 pith:YLZJPUF6 submitted 2026-07-28 cs.SD

From Semantics to Readout: Mechanistic Understanding of Audio Tokens after Fine-Tuning for Temporal Audio Grounding

Yujian Ma , Jinqiu Sang , Ruizhe Li , Jiaao Yu , Ang Li This is my paper
classification cs.SD
keywords temporal audio groundingaudio tokensmechanistic interpretabilitydecoder readoutfine-tuninglarge audio-language modelshidden statessemantics-to-readout
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper studies what happens inside large audio-language models when they are fine-tuned for temporal audio grounding—locating when a queried sound event occurs. By analyzing the hidden states of native audio tokens layer by layer in two models, it finds that the base, untuned checkpoints already contain latent evidence about queried events, positioned at largely stable times. Fine-tuning does not create that evidence; it makes it more accessible to the language decoder, and the main mechanism is decoder-side adaptation rather than changes in the audio-token states themselves. The paper also shows that fine-tuning aligns token-level evidence with the model's own predicted time windows, and that erasing updates at those windows degrades timestamp generation more than random erasure. A sympathetic reader would care because this reframes grounding failures as a readout problem rather than a missing-information problem, and offers four reusable diagnostics for studying audio-token representations.

Core claim

The central claim is a 'semantics-to-readout' account: temporal grounding fine-tuning improves how the decoder reads event evidence that already exists in native audio-token representations, and couples that evidence to temporal outputs. The base checkpoint's top event-aligned audio tokens appear at nearly the same temporal locations as after fine-tuning (Wasserstein distance 0.80s vs 1.60s random), and linear probes show annotated-window membership is recoverable before fine-tuning. After fine-tuning, calibrated readout gains increase primarily in early and middle layers, and a cross-checkpoint swap control shows the gain comes from the decoder, not from the supplied state. Predicted-window

What carries the argument

The central object is the layerwise hidden state of native audio tokens—the representations that carry acoustic evidence into the language decoder. Four diagnostics operate on these states: (1) query-conditioned token semantics, which scores each token by cosine similarity to a text-side embedding of the queried event phrase; (2) calibrated token readout, which patches a token state into a diagnostic prompt and measures whether the decoder can verbalize the event; (3) temporal-window probes, linear classifiers that test whether token states linearly encode membership in annotated or predicted time windows; and (4) residual-delta erasure, which ablates a layer's update at predicted-window tok

Load-bearing premise

The load-bearing premise is that audio tokens map uniformly to time (center time T*r/(n-1) and window width roughly T/n); all temporal diagnostics assume this linear correspondence, and if the real token-to-time alignment is nonlinear, the paper's stability and alignment results could be artifacts.

What would settle it

One concrete check: obtain token-level temporal alignments from the audio tower's internal frame rates or via forced alignment on a small labeled set, then re-run the top-20 hit ratios and window probes with those alignments. If the base-versus-fine-tuned stability gap disappears or the predicted-window probe gains shrink to chance, the central account is undermined. Alternatively, fine-tune the same base model on an equally demanding but non-temporal task—such as audio captioning—and test whether the readout gains still appear; if they do, the gains are not specific to temporal grounding.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Grounding gains come primarily from decoder-side adaptation, so improving temporal grounding can be achieved by changing how the decoder consumes audio tokens, not by re-encoding the audio.
  • Because base checkpoints already contain latent event evidence at stable locations, poor grounding before fine-tuning reflects a weak mapping from evidence to temporal output, not missing acoustic information.
  • Fine-tuning increases consistency between audio-token states and the model's own predicted windows more than recoverability of annotated windows, implying the model becomes self-consistent in its temporal outputs.
  • Removing residual updates within predicted windows degrades timestamp generation more than random erasure, showing those token updates are functionally used in generating the answer.

Where Pith is reading between the lines

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

  • The uniform token-to-time mapping is an assumption; if audio towers use variable-rate tokenization, the temporal-stability results should be re-validated with alignment-derived times.
  • The same semantics-to-readout pattern may appear in other modalities—for example, video tokens in video-language models—suggesting a general principle for grounding fine-tuning.
  • A practical extension: because decoder readout is the bottleneck, training a lightweight readout head or using better attention to audio tokens could improve grounding without full fine-tuning.
  • The diagnostics are cheap enough to run on a subset; they could serve as a monitoring tool during fine-tuning to see when readout gains saturate.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper studies how LoRA fine-tuning for temporal audio grounding changes the internal audio-token representations of Qwen2.5-Omni and Qwen2-Audio. It proposes four diagnostics: query-conditioned token semantics, a calibrated hidden-state readout, temporal-window probes, and residual-delta erasure. The main claims are that base checkpoints already contain latent event evidence at largely stable temporal locations; that fine-tuning improves decoder readout of this evidence primarily through decoder adaptation; that fine-tuning improves consistency with the model's own predicted windows more than annotated-window recoverability; and that residual updates in predicted windows are functionally relevant to timestamp generation. The paper reports large grounding gains after fine-tuning (e.g., Qwen2.5-Omni mIoU 0.37→0.68) and a cross-architecture replication.

Significance. The semantics-to-readout distinction is a valuable framing for mechanistic interpretability of LALMs. The paper's strengths include multiple converging diagnostics, a paired cross-checkpoint swap, a position-only probe control, paired bootstrap confidence intervals, and a second-model replication. The appendices contain substantial implementation detail, including explicit confounding checks (e.g., the conservative random-erasure fallback). If the two load-bearing methodological gaps identified below are addressed, the paper would provide a useful and reusable methodology for studying how task fine-tuning changes internal audio-token states. As it stands, the central RQ2 claim (decoder adaptation) and the temporal-alignment claims are not fully established.

major comments (3)
  1. [§3, Calibrated Token Readout; Figure 3 and Table 3] Δ_readout is defined against a no-patch baseline only. A positive gain does not distinguish 'this audio-token state contains event information the decoder can express' from 'the decoder has a prior to produce event-compatible completions regardless of the patched state.' Grounding fine-tuning strengthens exactly such a prior because the model is trained to emit event phrases. The cross-checkpoint control in Table 3 cannot rule this out: a decoder that largely ignores the patched state and emits generic event-compatible text would produce the observed pattern (large DecoderSwap, near-zero StateSwap). Please add control patches at the same [Sound] site — e.g., norm-matched Gaussian noise, another audio-token state from a different position/clip, or a text-token state — and report Δ_readout for these controls, together with the swap analysis under a control patch. Without this, the RQ2 conc
  2. [Appendix E.3; Figures 2, 4; Table 3] All temporal comparisons (top-20 hit ratios, window-probe labels, erasure masks, W1 stability) assume a uniform token-to-time map: token r is assigned center time T*r/(n-1) and width approximately T/n. Qwen audio towers may apply strided or variable-rate tokenization; if the true token-to-time correspondence is nonlinear, the reported hit ratios, probe AUROCs, and erasure gaps could reflect misalignment artifacts rather than genuine temporal evidence. The position-only probe control is not sufficient because it uses token rank, not time. Please validate the mapping against the model's actual audio encoder (e.g., downsampling rate, frame boundaries) or report sensitivity analyses under alternative monotonic maps. This is load-bearing for RQ1/RQ3/RQ4 and for the 'stable temporal locations' claim.
  3. [§3 Query-Conditioned Token Semantics; Figure 2] The semantic reference bank is built from each checkpoint's own text pathway, and predicted-window labels are derived from each checkpoint's own generated timestamps. The paper acknowledges the checkpoint-relative nature, but the interpretation in RQ1 ('latent event evidence already present') and RQ3 ('alignment with predicted support') is affected by this self-reference. For example, an increase in predicted-window hit ratio after fine-tuning could simply reflect the FT model's windows being more accurate, not states being more aligned in an objective sense. Please (i) report match coverage and handling of unmatchable queries for the matching cascade, (ii) add a control using a fixed external text encoder (e.g., SBERT) for the semantic references, and (iii) show that the predicted-window probe results hold when labels are constructed from oracle windows or from the other checkpoint's pr
minor comments (5)
  1. [Appendix C.2, Table 7] The no-patch baseline outputs (e.g., 'mechanisms whispering', 'mechanisms laugh') suggest the diagnostic prompt itself induces a generic 'mechanisms' prior. This is worth discussing explicitly, as it interacts with the readout-control issue in the major comments.
  2. [Figure 2 and Figure 7] The unit 'pp' (percentage points) is used in axis labels but not consistently defined in captions. Please state the unit once and define it.
  3. [§3, temporal-window probes] The L2 regularization α=1e-4 is only given in Appendix E.3. Add a pointer in the main text to the exact probe equation and training protocol, since the main text refers to 'standard diagnostic linear classifiers' without enough detail.
  4. [§3, Calibrated Token Readout] The term 'calibrated' may be confused with probability calibration. Consider renaming to 'baseline-subtracted readout gain' or defining the intended meaning explicitly in the main text.
  5. [Appendix E.3] The mapping assumes n audio tokens; if n=1 the expression T*r/(n-1) divides by zero. More generally, state how n is determined for Qwen2.5-Omni vs Qwen2-Audio and whether it is fixed across layers and checkpoints.

Circularity Check

1 steps flagged

Minor self-referential reference bank, but central semantics-to-readout claims have independent external anchors.

specific steps
  1. self definitional [Section 3, Query-Conditioned Token Semantics; RQ1 answer in abstract]
    "The diagnostic is checkpoint-relative because both the audio-token states and text-side event references for e(q) are extracted from the corresponding checkpoint; it therefore measures within-checkpoint audio-text alignment rather than isolating changes in audio states alone."

    RQ1 ('latent event evidence is already present before fine-tuning') is operationalized as high cosine similarity between audio-token states and a semantic vector z_l(q) built from the same checkpoint's own text-pathway hidden states. For any internally consistent model, audio and text representations of the same prompt will tend to align, so the base-checkpoint 'event evidence' is partly a self-referential measurement rather than an independent discovery about audio tokens. The circularity is limited because the paper acknowledges the checkpoint-relative nature and additionally corroborates RQ1 with annotated-window overlaps and temporal-window probes using external ground truth.

full rationale

The paper's central claims are not forced by construction. The readout diagnostic has a no-patch baseline and a cross-checkpoint decoder/state swap (Appendix C.1), the temporal-window probes use external annotated windows alongside checkpoint-specific predicted windows, and residual-delta erasure is a size-matched causal intervention. Fine-tuning gains are benchmarked against external annotations (Table 2). The one self-referential design choice—the event-phrase semantic bank built from the same checkpoint being analyzed—is disclosed and does not by itself determine the main conclusions, which are also supported by annotated-window probes and cross-checkpoint controls. Thus the overall circularity is minor (score 2), not a case of predictions reducing to inputs by construction.

Axiom & Free-Parameter Ledger

4 free parameters · 7 axioms · 0 invented entities

The paper introduces no new physical entities, forces, or conserved quantities. Its load-bearing assumptions are measurement and modeling assumptions about token-time alignment, semantic reference validity, patch-based readout, SBERT scoring, and single-seed training. The free parameters listed are analysis thresholds rather than fitted model parameters; robustness to their exact values is partially checked.

free parameters (4)
  • Top-M audio-token selection size = 20
    Chosen for the semantics diagnostic; robustness to Top-5/10/20/50 is checked in Appendix E.1.
  • Degradation threshold for residual-delta erasure = 0.05
    A sample counts as degraded when timestamp IoU drops by more than 0.05; Appendix E.2 reports that thresholds 0, 0.05, and 0.10 give similar conclusions.
  • Start-time recovery tolerance = 1 second
    Used in the activation-patching ShiftScore in Appendix F; peripheral to the main claims.
  • Probe L2 regularization = 1e-4
    Standard choice for the logistic probes; the claimed pattern is layerwise and consistent, so this value is not load-bearing.
axioms (7)
  • domain assumption Uniform token-to-time mapping: audio token with rank r is assigned center time T*r/(n-1) and width ~T/n.
    Used in Appendix E.3 to construct all temporal window labels, top-20 hit ratios, and erasure masks. If token times are nonuniform, temporal alignment results may be artifacts.
  • domain assumption The event-phrase semantic bank built from each checkpoint's own text pathway is a valid semantic reference for the queried event.
    The query-conditioned semantics diagnostic compares audio states to same-checkpoint text states; this measures within-model alignment, not external event semantics. Appendix B validates category structure but not external meaning.
  • domain assumption Patching an audio-token hidden state into a text placeholder and decoding is a valid probe of decoder readability.
    The calibrated readout diagnostic follows the Patchscopes paradigm; if inserted states are out-of-distribution at the placeholder, readout gains may be artifacts.
  • domain assumption SBERT (paraphrase-multilingual-mpnet-base-v2) cosine similarity captures semantic match between decoded readout phrases and target event phrases.
    All readout gains are computed in this external embedding space; a poor SBERT fit would distort the measured gains.
  • domain assumption A single LoRA fine-tuning run per model is representative of grounding fine-tuning.
    No repeated seeds or training runs are reported; cross-checkpoint conclusions may be sensitive to seed variation.
  • domain assumption Qwen3-Max normalization of source event phrases preserves the semantic content needed for grounding.
    Dataset construction in Appendix A normalizes phrases with Qwen3-Max at temperature 0; errors in normalization would systematically alter the QA instances and semantic bank.
  • domain assumption Linear logistic probes on frozen hidden states, with a position-only control, measure recoverability of temporal membership.
    The temporal-window probe conclusions depend on this recoverability framing and on the adequacy of the position-only control.

pith-pipeline@v1.3.0-alltime-deepseek · 23421 in / 12737 out tokens · 131018 ms · 2026-08-01T02:39:28.964199+00:00 · methodology

0 comments
read the original abstract

Large audio-language models (LALMs) convey acoustic evidence to language decoders through native audio tokens, yet the internal roles of these tokens remain poorly understood. Using temporal audio grounding as a diagnostic setting, we examine how language-model fine-tuning affects the layerwise semantics, decoder accessibility, and temporal output alignment of native audio-token states through four complementary analyses: query-conditioned token semantics, calibrated token readout, temporal-window probes, and residual-delta erasure during generation. Alongside substantial improvements in temporal localization, semantic analysis of Qwen2.5-Omni shows that latent evidence for queried events is already present before fine-tuning and that the audio tokens most strongly aligned with the queried event appear at similar temporal positions before and after fine-tuning. After fine-tuning, event-related information in audio tokens becomes more accessible to the decoder, especially in early and middle layers, and a cross-checkpoint control shows that this improvement arises primarily from decoder adaptation. Temporal probes show that the base checkpoint already contains recoverable information about annotated windows and that fine-tuning mainly improves alignment with each checkpoint's own predicted temporal support. Residual-delta erasure further shows that removing audio-token updates within predicted windows harms timestamp generation more than removing the same number of randomly selected updates. The same broad improvements in decoder readability and prediction alignment also appear in Qwen2-Audio. Together, these results support a semantics-to-readout account in which grounding fine-tuning helps the decoder read existing event evidence and connect it more reliably to temporal outputs.

Figures

Figures reproduced from arXiv: 2607.25355 by Ang Li, Jiaao Yu, Jinqiu Sang, Ruizhe Li, Yujian Ma.

Figure 1
Figure 1. Figure 1: Overview of the core audio-token semantics-to-readout analysis pipeline. (1) [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Audio-token semantics analysis. (a1) Top-20 overlap with annotated and predicted temporal support. (a2) Inside– [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Audio-token readout analysis. (a1) Mean calibrated readout gain for base and fine-tuned models. (a2) Percentage of [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Output-facing temporal diagnostics. (a1,a2) Layer [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 1
Figure 1. Figure 1: Category-wise semantic structure of the event-phrase bank. Related/confusable pairs remain more [PITH_FULL_IMAGE:figures/full_fig_p013_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Layerwise event-phrase bank similarity. Each panel reports lower-triangle cosine similarity among [PITH_FULL_IMAGE:figures/full_fig_p013_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Layer-to-layer CKA of the event-phrase bank. Higher values indicate that two layers preserve more [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Layerwise version of the cross-checkpoint readout control summarized in Table 3 of the main [PITH_FULL_IMAGE:figures/full_fig_p016_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Temporal stability of high-scoring query-aligned audio tokens across checkpoints. (a) Base–FT [PITH_FULL_IMAGE:figures/full_fig_p019_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Layerwise activation patching ShiftScore. Mean-phrase patching tests the target event phrase [PITH_FULL_IMAGE:figures/full_fig_p023_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Qwen2-Audio aggregate audio-token semantics diagnostic. The panels compare top-20 [PITH_FULL_IMAGE:figures/full_fig_p024_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Qwen2-Audio matched top-20 audio-token semantics case. The panel visualizes where the [PITH_FULL_IMAGE:figures/full_fig_p025_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Qwen2-Audio audio-token readout diagnostic. The aggregate panels compare average and top [PITH_FULL_IMAGE:figures/full_fig_p026_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Qwen2-Audio temporal window probe results. Curves report held-out token-level [PITH_FULL_IMAGE:figures/full_fig_p026_10.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

31 extracted references · 7 linked inside Pith

  1. [1]

    International Conference on Learning Representations , volume=

    Salmonn: Towards generic hearing abilities for large language models , author=. International Conference on Learning Representations , volume=

  2. [2]

    Chu, Yunfei and Xu, Jin and Yang, Qian and Wei, Haojie and Wei, Xipin and Guo, Zhifang and Leng, Yichong and Lv, Yuanjun and He, Jinzheng and Lin, Junyang and others , journal=

  3. [3]

    2024 , organization=

    Kong, Zhifeng and Goel, Arushi and Badlani, Rohan and Ping, Wei and Valle, Rafael and Catanzaro, Bryan , booktitle=. 2024 , organization=

    2024

  4. [4]

    Wang, Bin and Zou, Xunlong and Lin, Geyu and Sun, Shuo and Liu, Zhuohan and Zhang, Wenyu and Liu, Zhengyuan and Aw, AiTi and Chen, Nancy , booktitle=

  5. [5]

    Yang, Qian and Xu, Jin and Liu, Wenrui and Chu, Yunfei and Jiang, Ziyue and Zhou, Xiaohuan and Leng, Yichong and Lv, Yuanjun and Zhao, Zhou and Zhou, Chang and others , booktitle=

  6. [6]

    Ahia, Orevaoghene and Bartelds, Martijn and Ahuja, Kabir and Gonen, Hila and Hofmann, Valentin and Arora, Siddhant and Li, Shuyue Stella and Puttagunta, Vishal and Adeyemi, Mofetoluwa and Buchireddy, Charishma and others , journal=

  7. [7]

    arXiv preprint arXiv:2606.17417 , year=

    A Closer Look at Failure Modes in Temporal Understanding of Large Audio-Language Models , author=. arXiv preprint arXiv:2606.17417 , year=

    Pith/arXiv arXiv

  8. [8]

    arXiv preprint arXiv:2510.12185 , year=

    Not in Sync: Unveiling Temporal Bias in Audio Chat Models , author=. arXiv preprint arXiv:2510.12185 , year=

    arXiv

  9. [9]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Listening Between the Frames: Bridging Temporal Gaps in Large Audio-Language Models , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=. 2026 , doi=

    2026

  10. [10]

    arXiv preprint arXiv:2604.13715 , year=

    Towards Fine-grained Temporal Perception: Post-Training Large Audio-Language Models with Audio-Side Time Prompt , author=. arXiv preprint arXiv:2604.13715 , year=

    Pith/arXiv arXiv

  11. [11]

    arXiv preprint arXiv:2604.13023 , year=

    SpotSound: Enhancing Large Audio-Language Models with Fine-Grained Temporal Grounding , author=. arXiv preprint arXiv:2604.13023 , year=

    Pith/arXiv arXiv

  12. [12]

    arXiv preprint arXiv:2602.10230 , year=

    Frame-Level Internal Tool Use for Temporal Grounding in Audio LMs , author=. arXiv preprint arXiv:2602.10230 , year=

    Pith/arXiv arXiv

  13. [13]

    Locating and Editing Factual Associations in

    Meng, Kevin and Bau, David and Andonian, Alex and Belinkov, Yonatan , journal=. Locating and Editing Factual Associations in

  14. [14]

    Proceedings of the 41st International Conference on Machine Learning , series=

    Patchscopes: A Unifying Framework for Inspecting Hidden Representations of Language Models , author=. Proceedings of the 41st International Conference on Machine Learning , series=

  15. [15]

    arXiv preprint arXiv:2602.00462 , year=

    LatentLens: Revealing Highly Interpretable Visual Tokens in LLMs , author=. arXiv preprint arXiv:2602.00462 , year=

    Pith/arXiv arXiv

  16. [16]

    IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages=

    Text-to-Audio Grounding: Building Correspondence Between Captions and Sound Events , author=. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages=. 2021 , organization=

    2021

  17. [17]

    AudioCaps : Generating Captions for Audios in The Wild

    Kim, Chris Dongjoo and Kim, Byeongchang and Lee, Hyunmin and Kim, Gunhee. AudioCaps : Generating Captions for Audios in The Wild. 2019. doi:10.18653/v1/N19-1011

    doi:10.18653/v1/n19-1011 2019

  18. [18]

    2017 , organization=

    Gemmeke, Jort F and Ellis, Daniel PW and Freedman, Dylan and Jansen, Aren and Lawrence, Wade and Moore, R Channing and Plakal, Manoj and Ritter, Marvin , booktitle=. 2017 , organization=

    2017

  19. [19]

    Yang, An and Li, Anfeng and Yang, Baosong and Zhang, Beichen and Hui, Binyuan and Zheng, Bo and Yu, Bowen and Gao, Chang and Huang, Chengen and Lv, Chenxu and others , journal=

  20. [20]

    Xu, Jin and Guo, Zhifang and He, Jinzheng and Hu, Hangrui and He, Ting and Bai, Shuai and Chen, Keqin and Wang, Jialin and Fan, Yang and Dang, Kai and Zhang, Bin and Wang, Xiong and Chu, Yunfei and Lin, Junyang , journal=

  21. [21]

    and Shen, Yelong and Wallis, Phillip and Allen-Zhu, Zeyuan and Li, Yuanzhi and Wang, Shean and Wang, Lu and Chen, Weizhu , booktitle=

    Hu, Edward J. and Shen, Yelong and Wallis, Phillip and Allen-Zhu, Zeyuan and Li, Yuanzhi and Wang, Shean and Wang, Lu and Chen, Weizhu , booktitle=. 2022 , url=

    2022

  22. [22]

    arXiv preprint arXiv:2406.08056 , year=

    Cornell, Samuele and Ebbers, Janek and Douwes, Constance and Mart. arXiv preprint arXiv:2406.08056 , year=

    Pith/arXiv arXiv

  23. [23]

    Bhosale, Swapnil and Nag, Sauradip and Kanojia, Diptesh and Deng, Jiankang and Zhu, Xiatian , booktitle=

  24. [24]

    AudioLens: A Closer Look at Auditory Attribute Perception of Large Audio-Language Models , year=

    Yang, Chih-Kai and Ho, Neo and Lee, Yi-Jyun and Lee, Hung-Yi , booktitle=. AudioLens: A Closer Look at Auditory Attribute Perception of Large Audio-Language Models , year=

  25. [25]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Do Audio-Visual Large Language Models Really See and Hear? , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  26. [26]

    Beyond Transcription: Mechanistic Interpretability in

    Glazer, Neta and Segal-Feldman, Yael and Segev, Hilit and Shamsian, Aviv and Buchnick, Asaf and Hetz, Gill and Fetaya, Ethan and Keshet, Joseph and Navon, Aviv , booktitle=. Beyond Transcription: Mechanistic Interpretability in

  27. [27]

    Designing and interpreting probes with control tasks , author=. Proceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference on natural language processing (emnlp-ijcnlp) , pages=

    2019

  28. [28]

    Sentence-

    Reimers, Nils and Gurevych, Iryna , booktitle=. Sentence-. 2019 , doi=

    2019

  29. [29]

    arXiv preprint arXiv:2605.20266 , year=

    A Survey of Large Audio Language Models: Generalization, Trustworthiness, and Outlook , author=. arXiv preprint arXiv:2605.20266 , year=

    Pith/arXiv arXiv

  30. [30]

    arXiv preprint arXiv:2603.13768 , year=

    Causal Tracing of Audio-Text Fusion in Large Audio Language Models , author=. arXiv preprint arXiv:2603.13768 , year=

    arXiv

  31. [31]

    Behind the Scenes: Mechanistic Interpretability of

    Ma, Yujian and Lu, Xikun and Sang, Jinqiu and Jiang, Xianquan and Li, Ruizhe , booktitle=. Behind the Scenes: Mechanistic Interpretability of. 2026 , doi=

    2026