pith. sign in

arxiv: 2606.15088 · v2 · pith:OQC3CFB7new · submitted 2026-06-13 · 💻 cs.SD · cs.CL· eess.AS

When the Same Musical Knowledge Forgets Differently: A Clean Probe of Pathway-Dependent Forgetting

Yu Liu , Zhiwei Yang , Wenxiao Zhang , Cong Cao , Fangfang Yuan , Kun Peng , Haimei Qin , Lei Jiang
show 3 more authors
Jin B. Hong Hao Peng Yanbing Liu
This is my paper

Pith reviewed 2026-06-27 04:48 UTC · model grok-4.3

classification 💻 cs.SD cs.CLeess.AS
keywords pathway-dependent forgettingmultimodal modelsaudio-language modelsmusic understandingacquisition routePaired Pathway Controlled Protocolforgetting asymmetry
0
0 comments X

The pith

Text-acquired musical knowledge forgets more than the same knowledge acquired via audio under identical pressure.

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

The paper challenges the assumption that the path by which knowledge enters a model does not affect later forgetting. It uses music clips and aligned text descriptions to deliver identical content through listening or reading routes in audio-language models. A controlled three-phase protocol matches the pathways and applies the same adaptation pressure, revealing that text-route knowledge is lost more readily. This holds across models even after controls for depth and other factors, so acquisition route becomes a variable that must be tracked in forgetting studies.

Core claim

Across multiple architecturally distinct audio-language models, text-pathway knowledge is forgotten more than matched audio-pathway knowledge under identical adaptation pressure. To attribute this to route rather than confounds, the Paired Pathway Controlled Protocol establishes matched pathway baselines, activates both pathways under symmetric supervision on the same knowledge pool, and applies identical forgetting pressure to both pathways. The gap is stable across models and gain-controlled analyses, persists when contradictory overwrite is replaced by correct-label cross-domain learning, remains under single-modality pressure, and is not removed by lightweight replay. Two independent rou

What carries the argument

The Paired Pathway Controlled Protocol (PPCP), a three-phase design that matches pathways and applies symmetric forgetting pressure to isolate acquisition route effects.

Load-bearing premise

The Paired Pathway Controlled Protocol successfully attributes the observed asymmetry to the acquisition route rather than confounds such as architectural depth or data differences.

What would settle it

A replication using the same PPCP protocol across several models that finds equal forgetting rates for text and audio pathways would show the route does not drive the difference.

Figures

Figures reproduced from arXiv: 2606.15088 by Cong Cao, Fangfang Yuan, Haimei Qin, Hao Peng, Jin B. Hong, Kun Peng, Lei Jiang, Wenxiao Zhang, Yanbing Liu, Yu Liu, Zhiwei Yang.

Figure 1
Figure 1. Figure 1: Schematic of path-dependent forgetting: for the [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of PPCP. The same musical knowledge unit is presented to both the audio pathway (A2T) and the text [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Trajectory of modality-specific performance and degradation across training steps, where the upper bars represent [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Per-pathway forgetting trajectory during Phase 2. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Exploratory gradient analysis on Qwen2-Audio [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
read the original abstract

A model can learn that the piano piece F\"ur Elise is calm and reflective by listening to the audio or by reading a text description, but does it matter which route that knowledge took when it is later at risk of being forgotten? Forgetting research in multimodal models measures what knowledge is lost under adaptation, yet has not asked whether acquisition route affects how easily that knowledge is forgotten. We call this untested premise the Pathway-Invariant Assumption. Music understanding enables a clean test because a music clip and a canonical text description can be aligned to the same perceptual content, allowing the same knowledge unit to enter a model through listening or reading while the target remains fixed. Across multiple architecturally distinct audio-language models, we observe a consistent asymmetry: text-pathway knowledge is forgotten more than matched audio-pathway knowledge under identical adaptation pressure. To attribute this effect to route rather than confounds, we introduce the Paired Pathway Controlled Protocol (PPCP), a three-phase design that establishes matched pathway baselines, activates both pathways under symmetric supervision on the same knowledge pool, and applies identical forgetting pressure to both pathways. The gap is stable across models and gain-controlled analyses, persists when contradictory overwrite is replaced by correct-label cross-domain learning, remains under single-modality pressure, and is not removed by lightweight replay. Two independent routing-depth controls confirm that the effect is not explained by architectural depth, pointing to input representation as the dominant factor. Under PPCP, our results demonstrate that forgetting is highly route-dependent, establishing acquisition route as a new analytical dimension for forgetting research and multimodal system design.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

0 major / 2 minor

Summary. The paper claims that the acquisition route of knowledge affects forgetting rates in multimodal audio-language models, violating the Pathway-Invariant Assumption. Using music clips and aligned text descriptions as matched knowledge units, it reports that text-pathway knowledge is forgotten more readily than audio-pathway knowledge under identical adaptation pressure. The Paired Pathway Controlled Protocol (PPCP) is introduced to establish matched baselines, apply symmetric supervision on the same knowledge pool, and impose identical forgetting pressure while controlling for architectural depth and data mismatch. The asymmetry persists across architecturally distinct models, gain-controlled analyses, single-modality pressure, replay, cross-domain overwrite, and two independent routing-depth controls, attributing the effect primarily to input representation.

Significance. If the results hold under the described controls, the work supplies a clean empirical demonstration that acquisition route is a load-bearing factor in forgetting, adding a new analytical dimension to multimodal forgetting research and system design. The music testbed enables precise alignment of perceptual content across modalities, and the PPCP framework (matched baselines + symmetric supervision + identical pressure plus persistence checks) is a reusable methodological contribution. Explicit credit is due for the stability of the text-vs-audio gap across the listed controls and the routing-depth verifications, which directly address the main confounds.

minor comments (2)
  1. [Abstract] Abstract: the claim of stability 'across models' would be strengthened by stating the number of models and the typical magnitude of the asymmetry (e.g., percentage-point difference in forgetting rate).
  2. [Methods (PPCP)] The description of the PPCP three-phase design would benefit from an explicit diagram or table summarizing the matched conditions for each pathway at each phase.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for the positive assessment, accurate summary of our contributions, and recommendation of minor revision. The referee correctly identifies the core claim, the role of the PPCP protocol, and the robustness checks across models and controls. No specific major comments were raised in the report.

Circularity Check

0 steps flagged

Empirical protocol with no derivation chain

full rationale

The paper is a controlled empirical study introducing the PPCP protocol to isolate acquisition-route effects on forgetting. No equations, derivations, fitted parameters, or self-citation chains are present that could reduce any claim to its own inputs by construction. The central attribution rests on experimental controls (matched baselines, symmetric supervision, routing-depth checks) rather than definitional equivalence or renamed fits. This matches the default expectation of no significant circularity for non-derivational work.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The claim rests on the effectiveness of the introduced PPCP protocol to isolate route effects and the assumption of matched knowledge across pathways. Full verification requires the methods section.

axioms (1)
  • domain assumption A music clip and its canonical text description align to the same perceptual content allowing matched knowledge units
    This is the premise enabling the clean test of pathway effects as stated in the abstract.

pith-pipeline@v0.9.1-grok · 5856 in / 1194 out tokens · 43666 ms · 2026-06-27T04:48:06.180358+00:00 · methodology

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

42 extracted references · 13 canonical work pages · 5 internal anchors

  1. [1]

    Arslan Chaudhry, Marcus Rohrbach, Mohamed Elhoseiny, Thalaiyasingam Ajber, and Stefano Soatto. 2019. Continual learning with tiny episodic memories. In Workshop on Multi-Task and Lifelong Reinforcement Learning

    2019

  2. [2]

    Qizhou Chen, Chengyu Wang, Dakan Wang, Taolin Zhang, Wangyue Li, and Xiaofeng He. 2025. Lifelong knowledge editing for vision language models with low-rank mixture-of-experts. InProceedings of the Computer Vision and Pattern Recognition Conference. 9455–9466

    2025

  3. [3]

    Sanyuan Chen, Yu Wu, Chengyi Wang, Shujie Liu, Daniel Tompkins, Zhuo Chen, and Furu Wei. 2023. BEATs: Audio Pre-Training with Acoustic Tokenizers. In Proceedings of the 40th International Conference on Machine Learning (ICML), Vol. 202. PMLR, 5178–5193

    2023

  4. [4]

    Gonzalez, Ion Stoica, and Eric P

    Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. 2023. Vicuna: An Open-Source Chatbot Impressing GPT-4 with 90%* ChatGPT Quality. https://lmsys.org/blog/2023-03-30-vicuna/

    2023

  5. [5]

    Yunfei Chu, Jin Xu, Qian Yang, Haojie Wei, Xipin Wei, Zhifang Guo, Yichong Leng, Yuanjun Lv, Jinzheng He, Junyang Lin, et al. 2024. Qwen2-Audio technical report.arXiv preprint arXiv:2407.10759(2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  6. [6]

    William G. Cochran. 1957. Analysis of Covariance: Its Nature and Uses.Biometrics 13, 3 (1957), 261–281

    1957

  7. [7]

    1988.Statistical Power Analysis for the Behavioral Sciences(2nd ed.)

    Jacob Cohen. 1988.Statistical Power Analysis for the Behavioral Sciences(2nd ed.). Lawrence Erlbaum Associates

    1988

  8. [8]

    Bradley Efron. 1979. Bootstrap Methods: Another Look at the Jackknife.The Annals of Statistics7, 1 (1979), 1–26

    1979

  9. [9]

    Robert M French. 1999. Catastrophic forgetting in connectionist networks.Trends in Cognitive Sciences3, 4 (1999), 128–135

    1999

  10. [10]

    Sreyan Ghosh, Arushi Goel, Lasha Koroshinadze, Sang-gil Lee, Zhifeng Kong, Joao Felipe Santos, Ramani Duraiswami, Dinesh Manocha, Wei Ping, Mohammad Shoeybi, and Bryan Catanzaro. 2025. Music Flamingo: Scaling Music Under- standing in Audio Language Models.arXiv preprint arXiv:2511.10289(2025)

    work page arXiv 2025

  11. [11]

    Arushi Goel, Sreyan Ghosh, Jaehyeon Kim, Sonal Kumar, Zhifeng Kong, Sang-gil Lee, Chao-Han Huck Yang, Ramani Duraiswami, Dinesh Manocha, Rafael Valle, and Bryan Catanzaro. 2025. Audio Flamingo 3: Advancing Audio Intelligence with Fully Open Large Audio Language Models.arXiv preprint arXiv:2507.08128 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  12. [12]

    Tianyu Huai, Jie Zhou, Xingjiao Wu, Qin Chen, Qingchun Bai, Ze Zhou, and Liang He. 2025. Cl-moe: Enhancing multimodal large language model with dual momentum mixture-of-experts for continual visual question answering. InProceedings of the computer vision and pattern recognition conference. 19608– 19617

    2025

  13. [13]

    James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. 2017. Overcoming catastrophic forgetting in neural networks.Proceedings of the National Academy of Sciences114, 13 (2017), 3521– 3526

    2017

  14. [14]

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. 2023. BLIP-2: Boot- strapping Language-Image Pre-Training with Frozen Image Encoders and Large Language Models. InInternational Conference on Machine Learning. PMLR, 19730– 19742

    2023

  15. [15]

    Shansong Liu, Atin Sakkeer Hussain, Chenshuo Sun, and Ying Shan. 2024. Music understanding llama: Advancing text-to-music generation with question an- swering and captioning. InICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 286–290

    2024

  16. [16]

    Xiaohao Liu, Xiaobo Xia, See-Kiong Ng, and Tat-Seng Chua. 2025. Continual multimodal contrastive learning.arXiv preprint arXiv:2503.14963(2025)

    work page arXiv 2025

  17. [17]

    Pratyush Maini, Zhili Feng, Avi Schwarzschild, Zachary Chase Lipton, and J Zico Kolter. 2024. TOFU: A Task of Fictitious Unlearning for LLMs. InAdvances in Neural Information Processing Systems

    2024

  18. [18]

    Michael McCloskey and Neal J Cohen. 1989. Catastrophic interference in con- nectionist networks: The sequential learning problem.Psychology of Learning and Motivation24 (1989), 109–165

    1989

  19. [19]

    Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. 2022. Locating and Editing Factual Associations in GPT. InAdvances in Neural Information Processing Systems

    2022

  20. [20]

    Eric Mitchell, Charles Lin, Antoine Bosselut, Chelsea Finn, and Christopher D Manning. 2023. Mass-Editing Memory in a Transformer. InProceedings of the 11th International Conference on Learning Representations (ICLR)

    2023

  21. [21]

    Shuai Pan et al. 2025. Audio-Language Models for Audio-Centric Tasks: A Survey. arXiv preprint arXiv:2501.15177(2025)

    work page arXiv 2025

  22. [22]

    Weiguo Pian, Shijian Deng, Shentong Mo, Yunhui Guo, and Yapeng Tian. 2025. Modality-Inconsistent Continual Learning of Multimodal Large Language Mod- els. InProceedings of the 39th AAAI Conference on Artificial Intelligence

    2025

  23. [23]

    Soujanya Poria, Devamanyu Hazarika, Navonil Majumder, Gautam Naik, Erik Cambria, and Rada Mihalcea. 2019. MELD: A Multimodal Multi-Party Dataset for Emotion Recognition in Conversations. InProceedings of the 57th Annual Meeting of the Association for Computational Linguistics. 527–536

    2019

  24. [24]

    Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. 2023. Robust Speech Recognition via Large-Scale Weak Super- vision. InProceedings of the 40th International Conference on Machine Learning (ICML), Vol. 202. PMLR, 28492–28518

    2023

  25. [25]

    Nils Reimers and Iryna Gurevych. 2019. Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks. InProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing

    2019

  26. [26]

    Andrei A Rusu, Neil C Rabinowitz, Guillaume Desjardins, Hubert Soyer, James Kirkpatrick, Koray Kavukcuoglu, Razvan Pascanu, and Raia Hadsell. 2016. Pro- gressive neural networks.arXiv preprint arXiv:1606.04671(2016)

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  27. [27]

    Haizhou Shi, Zihao Xu, Hengyi Wang, Weiyi Qin, Wenyuan Wang, Yibin Wang, Zifeng Wang, Sayna Ebrahimi, and Hao Wang. 2025. Continual Learning of Large Language Models: A Comprehensive Survey.Comput. Surveys58, 5 (2025). doi:10.1145/3735633

    work page doi:10.1145/3735633 2025

  28. [28]

    Changli Tang, Wenyi Yu, Guangzhi Sun, Xianzhao Chen, Tian Tan, Wei Li, Lu Lu, Zejun Ma, and Chao Zhang. 2023. Salmonn: Towards generic hearing abilities for large language models.arXiv preprint arXiv:2310.13289(2023). Liu et al

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  29. [29]

    Lukas Thede, Karsten Roth, Matthias Bethge, Zeynep Akata, and Tom Hartvigsen

  30. [30]

    Wikibigedit: Understanding the limits of lifelong knowledge editing in LLMs.arXiv preprint arXiv:2503.05683(2025)

    work page arXiv 2025

  31. [31]

    Gido M van de Ven, Nicholas Soures, and Dhireesha Kudithipudi. 2025. Continual Learning and Catastrophic Forgetting. InLearning and Memory: A Comprehensive Reference(3rd ed.), John Wixted (Ed.). Vol. 1. Academic Press, 153–168

    2025

  32. [32]

    Liyuan Wang, Xingxing Zhang, Hang Su, and Jun Zhu. 2024. A Comprehensive Survey of Continual Learning: Theory, Method and Application.IEEE Trans- actions on Pattern Analysis and Machine Intelligence46, 8 (2024), 5362–5383. doi:10.1109/TPAMI.2024.3367329

    work page doi:10.1109/tpami.2024.3367329 2024

  33. [33]

    Jason Wei, Maarten Bosma, Vincent Y Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M Dai, and Quoc V Le. 2022. Finetuned Language Models Are Zero-Shot Learners. InProceedings of the 10th International Conference on Learning Representations (ICLR)

    2022

  34. [34]

    Xiwen Wei, Mustafa Munir, and Radu Marculescu. 2025. Mitigating Intra- and Inter-modal Forgetting in Continual Learning of Unified Multimodal Models. In Advances in Neural Information Processing Systems

    2025

  35. [35]

    Xiwen Wei, Mustafa Munir, and Radu Marculescu. 2025. Mitigating Intra-and Inter-modal Forgetting in Continual Learning of Unified Multimodal Models. arXiv preprint arXiv:2512.03125(2025)

    work page arXiv 2025

  36. [36]

    Frank Wilcoxon. 1945. Individual Comparisons by Ranking Methods.Biometrics Bulletin1, 6 (1945), 80–83

    1945

  37. [37]

    Jin Xu, Zhifang Guo, Jinzheng He, Hangrui Hu, Ting He, Shuai Bai, Keqin Chen, Jialin Wang, Yang Fan, Kai Dang, Bin Zhang, Xiong Wang, Yunfei Chu, and Jun- yang Lin. 2025. Qwen2.5-Omni Technical Report.arXiv preprint arXiv:2503.20215 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  38. [38]

    Yuexiang Zhai, Shengbang Tong, Xiao Li, Mu Cai, Qing Qu, Yong Jae Lee, and Yi Ma. 2024. Investigating the Catastrophic Forgetting in Multimodal Large Language Models. InConference on Parsimony and Learning

    2024

  39. [39]

    Qiang Zhang, Fanrui Zhang, Jiawei Liu, Ming Hu, Junjun He, and Zheng-Jun Zha. [n. d.]. Reliable Lifelong Multimodal Editing: Conflict-Aware Retrieval Meets Multi-Level Guidance. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems

  40. [40]

    Weinberger, and Yoav Artzi

    Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q. Weinberger, and Yoav Artzi

  41. [41]

    InProceedings of the 8th International Conference on Learning Representations (ICLR)

    BERTScore: Evaluating Text Generation with BERT. InProceedings of the 8th International Conference on Learning Representations (ICLR)

  42. [42]

    Junhao Zheng, Qianli Ma, Zhen Liu, Binquan Wu, and Huawen Feng. 2024. Beyond Anti-Forgetting: Multimodal Continual Instruction Tuning with Positive Forward Transfer.arXiv preprint arXiv:2401.09181(2024). When the Same Musical Knowledge Forgets Differently: A Clean Probe of Pathway-Dependent Forgetting A Training Hyperparameters Table 5 lists the full trai...

    work page arXiv 2024