Pith. sign in

REVIEW 4 major objections 5 minor 45 references

Steering dense music retrieval with open-vocabulary concept discovery

T0 review · 4 major / 5 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read Treating concept attribution as a geometric inverse problem recovers the audio-side features that text-aligned neuron ranking misses.

desk verdict Solid empirical paper on sparse-inversion concept attribution for music retrieval, but the missing nearest-audio-support baseline leaves the central claim underdetermined. read the letter →

arxiv 2608.08757 v1 pith:BOTOQJTX submitted 2026-08-09 cs.SD eess.AS

Julien Guinot , Alain Riou , Elio Quinton , György Fazekas This is my paper · ORCID
classification cs.SDeess.AS
keywords sparseautoencodersconceptattributiondensemusicretrievalcontrollableopen-vocabularyconceptsinversionmultimodalembeddingsmodalitygap
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

This paper targets a concrete failure in controllable music retrieval: when a user wants to find tracks that are 'more jazzy' or 'without guitar,' the system must decide which sparse audio features to amplify or suppress. The paper shows that the usual approach—ranking SAE neurons by cosine similarity to the concept's text embedding—picks neurons that match the wording but not the audio examples that actually express the concept, because text and audio embeddings occupy different regions of the shared space and concepts are split across many neurons. It proposes a training-free sparse inversion: solve for a sparse code whose decode reconstructs the text concept direction while a Mahalanobis term keeps it near the empirical audio distribution. The recovered supports overlap far more with concept-bearing audio (10 to 70 percent higher bundle recall), lift concept-detection AUROC from roughly 50 percent to 65–80 percent, and move the edit–preservation Pareto frontier outward. If right, this gives open-vocabulary, real-time concept steering of dense music search without paired supervision or SAE retraining.

What carries the argument

The load-bearing object is the sparse-inversion objective of Eq. (4): minimize cosine distance between the text concept embedding and the decode of a sparse code, penalized by the Mahalanobis distance to the empirical audio embedding distribution. The sparsifying operator $\Pi$ is the same one used in SAE training, so the solution respects native sparsity; the decoder $D$ is the audio SAE dictionary, so the support is expressed in the basis actually used for steering. Two solvers are used, gradient descent with Adam and FISTA, both initialized from the sparse code of the nearest audio neighbour to the text embedding to avoid convergence drift, followed by an IDF weighting that downweights over-frequent neurons. This mechanism replaces independent neuron ranking with joint reconstruction, which is what lets text concepts bridge to audio-side structure.

What would settle it

Take open-vocabulary concepts that have no audio-side realization in the training data—for example, 'quantum computer sounds' or 'neon dystopia'—and measure bundle recall of the recovered support against tagged audio examples plus the drift of edited retrievals; if the recovered supports score no better than random neurons or the edits degrade preservation, the claimed audio-faithful inversion mechanism is not doing the work.

Watch

Extended reading notes

Core claim

The central claim is that concept attribution in multimodal sparse spaces should be treated as a geometric inverse problem rather than a marginal neuron-ranking heuristic. Given a free-form concept with text embedding $z_c$ and an SAE trained on audio embeddings with decoder $D$ and sparsifying map $\Pi$, the paper solves $u_c^* = \arg\min_u d_{\cos}(z_c, D(\Pi(u))) + \gamma \Gamma(\mathcal{A}, D(\Pi(u)))$, where $\Gamma$ is a Mahalanobis distance to the empirical audio distribution. The selected support $\mathcal{K}(c)$ is the set of active, IDF-weighted coordinates of this optimized code. Across sparsity levels, these supports achieve 10 to 70 percent higher bundle recall, better bundle hit rate, and roughly 50 percent to 65–80 percent probe AUROC compared with cardinality-matched Discover-Then-Name cosine probing, and they yield stronger amplification and suppression edits at matched preservation cost. The paper concludes that faithful concept control depends less on finding individually aligned neurons than on recovering the sparse support that instantiates the concept in the target modality.

Load-bearing premise

The method assumes that any free-form text concept can be represented as a sparse combination of audio SAE dictionary atoms after the Mahalanobis regularizer pulls the solution toward the audio manifold; if a concept's text embedding points to a direction the audio dictionary cannot reach, the inversion in Eq. (4) returns arbitrary supports.

Editorial extensions

If this is right

  • Open-vocabulary concept control in music retrieval becomes feasible with no paired audio-text supervision and no SAE retraining: any text phrase can be mapped to an audio-side support in roughly 20 ms per inversion.
  • Amplification and suppression edits are more efficient: at matched edit strength, inversion moves the query closer to the target concept while staying closer to the audio manifold, so users get stronger concept change with less off-target tag drift.
  • Concept support is treated as a bundle rather than a single neuron, making the support size an explicit control whose optimal value varies from concept to concept.
  • Because the method is applied post hoc to any SAE in a shared audio-text space, it can be layered onto existing dense retrieval systems without changing the embedding model.

Reading between the lines

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

  • The same inversion recipe should transfer to other multimodal pairs with a modality gap, such as image-text retrieval, where text-aligned neurons are also likely to miss vision-side support; a testable extension is to run the identical protocol on joint image-text encoders.
  • A natural use of the recovered bundles is estimating a concept's support size from the data, replacing the fixed cardinality $|\mathcal{K}(c)|=L_0$ with a per-concept cut-off based on reconstruction loss or outlier structure.
  • If bundle recall is the right criterion, then SAE features that appear text-aligned but audio-inactive could be downweighted or removed before training the decoder, which suggests a data-cleaning loop between inversion and SAE training.
  • The IDF weighting and Mahalanobis regularizer together act like a prior for audio-likeness; one could strengthen the method by learning the regularizer from a larger held-out audio set rather than fixing a single scalar $\gamma$.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper addresses open-vocabulary concept attribution in sparse autoencoders (SAEs) for steerable music retrieval. It shows that Discover-Then-Name (DTN) cosine probing selects decoder atoms aligned with the text embedding of a concept but not active in audio examples expressing that concept. The authors propose a training-free alternative, sparse inversion: optimize a latent pre-activation u so that the decoded sparse code D(Π(u)) reconstructs the concept text embedding z_c under cosine distance, subject to a Mahalanobis regularizer Γ toward the empirical audio distribution, with nearest-audio-neighbor initialization, Adam or FISTA optimization, and IDF weighting. They evaluate the recovered supports against DTN on bundle recall/hit rate with positive audio examples, probe AUROC/mAP for concept classification, and retrieval edit–preservation Pareto frontiers. The paper concludes that concept attribution in multimodal sparse spaces should be treated as a geometric inverse problem rather than a marginal neuron-ranking heuristic.

Significance. If the reported effects are robust, the paper makes a useful and practical contribution: it provides a lightweight, training-free way to map free-form text concepts to audio-faithful SAE supports, with code released and an evaluation that uses external tags, held-out classifiers, and cardinality matching. The reframing of attribution as joint sparse reconstruction is coherent, and the experiments consistently support the direction of the effect. The main caveats are that the inversion objective is not isolated from other audio-side components and that the quantitative claims lack uncertainty quantification. These are addressable in revision.

major comments (4)
  1. [§3.3, Eq. (4), §5.3] The reported gains are not yet attributable to the sparse inversion objective itself. The full pipeline adds audio-side information in three places: the Mahalanobis regularizer Γ(A,D(Π(u))) in Eq. (4), the initialization of u from the sparse code of the nearest audio neighbor to z_c (Section 3.3), and the IDF weighting in Eq. (5) computed from audio activation frequencies. The DTN baseline uses none of these, so the 10–70% bundle-recall gains and improved Pareto frontiers could in principle be produced by any of the three components alone. Please add ablations that (i) remove Γ (γ=0), (ii) replace the nearest-neighbor initialization with a random or zero initialization, and (iii) compare against an audio-side support baseline that takes the SAE support of the nearest audio embedding to z_c and applies Eq. (5) without running Eq. (4). Without such controls, the central claim that concept attribution should be treated as a geometric inverse problem is not isolated.
  2. [§4.2, Figs. 3–5] No error bars, confidence intervals, or multiple-seed results are reported. Headline quantities such as “10–70% better bundle recall” and “~50% AUROC to ~65–80%” are point estimates over a small set of top-20 tags; the paper should provide bootstrap or per-concept variability and a significance test (e.g., paired across concepts) to establish that the improvements are not driven by a few concepts or a single SAE training run.
  3. [§3.3] The FISTA variant is introduced by citing Beck and Teboulle’s algorithm for linear inverse problems, but Eq. (4) is not a convex linear inverse problem because Π is the BatchTopK sparsifying operator, which is nonlinear and non-smooth. The convergence guarantees of FISTA therefore do not automatically transfer. Please justify the adaptation—for example, by proving that the composite objective remains convex in the relevant regime, or by reporting empirical convergence curves and a sensitivity analysis over FISTA parameters—or present FISTA explicitly as a heuristic whose behavior is validated only empirically.
  4. [§3.3, §5.2] The Mahalanobis regularizer Γ(A, ·) is estimated from “10k test-set MaxCaps audio embeddings,” and the same 10k set is later used as the retrieval pool and for the |ΔΓ| diagnostic. Fitting the regularizer on the evaluation set is a form of transductive leakage; it can inflate the manifold-adherence advantage of the proposed method over DTN, which uses no such estimate. Please estimate Γ from a held-out training split only, or clearly justify why this transductive use is acceptable.
minor comments (5)
  1. [§5.2, §5.3] The audio-manifold adherence claim relies partly on a circular diagnostic: |ΔΓ(z,A)| is the same Mahalanobis distance used in the Eq. (4) objective. The authors disclose this and label it as a diagnostic, but §5.3 still highlights that inverse variants stay “markedly closer to the audio manifold in Mahalanobis distance” as evidence. Please make the held-out audio-vs-text classifier p_A(z′) the primary audio-likeness metric, and report Γ changes only as descriptive statistics.
  2. [§3.3, Eq. (5)] The constant ϵ in the denominator df(k)+ϵ is never defined; please specify its value.
  3. [§4, Fig. 3] The text alternates between “cardinality-matched” and “cosine-matched” baselines; please use one term consistently and state explicitly whether the DTN support is obtained by top-k cosine ranks or by the 4σ outlier rule from Figure 2.
  4. [§5.1, Eqs. (6)–(7)] The equations do not show the L2 normalization of Π(u_c^*) described in the text; clarify the exact edit direction used in the retrieval experiments.
  5. [§4.1] Please report the number of concepts that survive the ≥10-positive-examples filter and the number of SAE training seeds used for Figure 3, as this information is needed to interpret the aggregate curves.

Circularity Check

1 steps flagged · score 2.0 of 10

No load-bearing circularity found; the only disclosed shared-metric diagnostic is explicitly labeled, and the central claim is supported by independent benchmarks.

  1. self definitional [Section 5.2 (Metrics), final paragraph; see also Eq. (4)]
    "We also report |∆Γ(z,A)|, the change in Mahalanobis distance to the empirical audio distribution, but treat it as a diagnostic as the same distance is used in the inversion objective."

    The inversion support is defined by minimizing dcos(z_c, D(Π(u))) + γ Γ(A, D(Π(u))) in Eq. (4). Therefore Γ(A, ·) is part of the very objective that selects the recovered coordinates. Reporting |ΔΓ| between edited and original embeddings as evidence that inversion stays closer to the audio manifold compares each method on a quantity that inversion optimizes by construction, so inversion is rewarded on this metric independently of whether it recovers better audio-faithful supports. The paper explicitly labels this metric a diagnostic and does not use it for the main Pareto claims, which rely on externally defined neighbor-tag drift, a held-out audio-vs-text classifier, and classifier probability changes; the circularity is disclosed and non-load-bearing.

full rationale

The derivation chain is otherwise self-contained. The recovered concept support K(c) is produced by sparse inversion (Eqs. 3–5) and compared against DTN cosine probing on three independent yardsticks: (i) bundle recall and hit-rate against the active supports S(x) of validation-set positive audio examples; (ii) probe AUROC/mAP of an MLP trained on held-out concept tags using only the recovered coordinates; and (iii) retrieval-editing Pareto frontiers using neighbor tag counts, classifier probability deltas, and a held-out audio-vs-text classifier. None of these metrics is a function of the inversion objective, so the reported 10–70% bundle-recall gains and improved Pareto frontiers are not forced by construction. The IDF reweighting in Eq. (5) is applied equally to DTN and inversion, so it does not bias the comparison. The only self-citation, reference [30], supports a background claim about the modality gap alongside external citations [26,29] and is not load-bearing. The main correctness risk—not a circularity—is the absence of ablations isolating the nearest-audio-neighbor initialization and the Mahalanobis regularizer; either could explain part of the gains without the inversion objective doing the work, but this is an experimental omission rather than a definitional reduction.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The method depends on a small number of hand-chosen hyperparameters (gamma, L0, IDF epsilon implicit) and on the domain assumptions that SAE features are controllable and that the text-to-audio inversion is well-posed under the audio manifold. No new entities or physical parameters are introduced.

free parameters (3)
  • gamma (manifold regularizer weight) = 1e-4 (fixed after sweep)
    Regularizer weight controlling manifold adherence in Eq. (4); chosen by hyperparameter sweep, not derived from data.
  • L0 sparsity level = 5, 10, 20, 50, 100
    Training-time sparsity of the SAE; also used to match concept support cardinality in experiments. The choice affects supports and retrieval behavior.
  • 4-sigma Gaussian outlier threshold = 4 standard deviations
    Ad hoc threshold for defining DTN supports in the descriptive analysis (Section 3.2), not used in the main method.
assumptions (5)
  • domain assumption SAE features are effective control units for retrieval edits.
    The entire steering mechanism assumes editing sparse codes in the SAE basis translates into meaningful dense embedding changes.
  • domain assumption Text concept embeddings can be approximated by sparse combinations of audio decoder atoms under audio-manifold regularization.
    Eq. (3)-(4) in Section 3.3 treat z_c as a target direction for the decoded sparse code; if the modality gap is too large, the recovered support is not audio-faithful.
  • domain assumption The Mahalanobis distance computed from 10k test-set audio embeddings approximates the audio manifold.
    Used in Eq. (4) and also as a diagnostic metric in Section 5.2.
  • ad hoc to paper FISTA applies to the inversion objective despite the nonlinear sparsifying operator Pi.
    Section 3.3 invokes linearity of D to justify FISTA, but Pi (TopK or BatchTopK) is nonlinear, so convergence is not guaranteed.
  • domain assumption IDF weighting suppresses generic neurons and improves concept specificity.
    Eq. (5) applies log N/df weighting to both DTN and inversion supports; treats frequent neurons as less concept-specific.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Steering dense music retrieval with open-vocabulary concept discovery." pith.science (2026). https://pith.science/paper/BOTOQJTX

@misc{pith2026260808757,
  author       = {Pith},
  title        = {Pith review of: Steering dense music retrieval with open-vocabulary concept discovery},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BOTOQJTX}},
  note         = {Machine review of arXiv:2608.08757}
}
read the original abstract

Controllable music retrieval lets users find music that is, for example, more ambient, less distorted, or without guitar while preserving the other semantic content of an original seed query. Sparse autoencoders (SAEs) are a promising interface for this kind of concept-level control, but a key problem remains: given a free-form text concept, which sparse features should be edited? In shared multimodal embedding spaces, standard attribution methods often select neurons that match the concept's wording but not the audio examples that express it. This leads to weak or unstable edits: relevant features are missed when concepts are distributed across neurons, while others are selected due to text alignment rather than audio-side structure. We address this with a lightweight, training-free method that recovers a sparse set of audio features whose decoded representation reconstructs the target concept while remaining consistent with audio-space geometry. This reframes concept attribution as a sparse inversion problem rather than a text-side neuron-ranking heuristic. The method requires neither paired audio-text supervision nor SAE retraining. We evaluate this approach in steerable music retrieval and show that the recovered supports align more closely with concept-bearing audio examples and achieve a stronger trade-off between edit strength and preservation than alignment baselines, enabling more precise concept amplification and suppression with reduced drift on preservation metrics.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

45 extracted references · 36 canonical work pages

  1. [1]

    Steering dense music retrieval with open-vocabulary concept discovery

    INTRODUCTION Dense retrieval underpins modern music search and rec- ommendation systems, where tracks are embedded into a shared space and retrieved by similarity to text or au- dio queries. For example, a user might search for “am- bient electronic music” or retrieve tracks similar to a ref- erence clip. While effective at capturing overall seman- tic si...

    work page Pith review arXiv 2026

  2. [2]

    BACKGROUND 2.1 Sparse Autoencoders Letf:X → A ⊂R d be a pretrained representation learning model. A Sparse Autoencoder (SAE) consists of a trainable encoder-decoder pair(E,D), and a sparsity con- straintΠsuch that, for all embeddingsz=f(x)∈ A: u=E(z)∈R m, s= Π(u)∈R m,ˆz=D(s)∈R d. Typicallym≫d. The training objective is: L(z,ˆz, s) =∥z−ˆz∥2 2 +λΩ(s),(1) wh...

    work page

  3. [3]

    Audio clips are encoded with the audio towers fA of CLAP [28] and MuQ [33]; from each 10-second frame we extract an embeddingz a ∈R 512

    METHOD AND OBSERV ATIONS 3.1 SAE Training We train SAEs on audio embeddings from JamendoMax- Caps [32]. Audio clips are encoded with the audio towers fA of CLAP [28] and MuQ [33]; from each 10-second frame we extract an embeddingz a ∈R 512. We train BatchTopK SAEs3 [2] withm= 4096latent features (ex- pansion factor 8), sparsity levelsL0 ∈ {5,10,20,50,100}...

    work page

  4. [4]

    piano”.Starting from an audio embeddingz, we navigate on the hypersphere by amplifying the concept “piano

    CONCEPT ATTRIBUTION EXPERIMENTS Do the neurons selected for a concept actually correspond to the sparse support used by audio examples that instanti- ate it? A useful concept-to-support mapping should satisfy two properties: it should overlap with the active supports of positive audio examples, and those restricted activations should retain enough signal ...

    work page

  5. [5]

    I want similar songs without the guitar

    SPARSE STEERABLE RETRIEV AL We now test whether improved concept attribution also im- proves retrieval editing. The target use case is user-facing control such as “I want similar songs without the guitar” or “I want a track where the vocalist sounds more like this artist.” In this section, we ask whether sparse autoencoders support steerable music retriev...

    work page

  6. [6]

    LIMITATIONS, DISCUSSION, CONCLUSION Our method is constrained by the underlying text–music embedding space and SAE fidelity. If a concept is weakly grounded, rare, abstract, or diffusely represented across modalities, inversion may fail to recover reliable support, inheriting reconstruction error, feature splitting, and fea- ture absorption. Performance a...

    work page

  7. [7]

    ACKNOWLEDGEMENT This work is supported by the EPSRC UKRI Centre for Doctoral Training in Artificial Intelligence and Music (EP/S022694/1) and Universal Music Group

    work page

  8. [8]

    AI assistance was also used sparingly in writing this paper to tighten phrasing, reword for conci- sion, and smooth out mathematical notation consistency

    AI USAGE STATEMENT AI LLM assistants were used sparingly in the development of this project to accelerate the implementation of evalu- ation pipelines. AI assistance was also used sparingly in writing this paper to tighten phrasing, reword for conci- sion, and smooth out mathematical notation consistency

    work page

Show all 45 references
  1. [9]

    Jumping ahead: Improving reconstruction fidelity with jumprelu sparse autoencoders,

    S. Rajamanoharan, T. Lieberum, N. Sonnerat, A. Conmy, V . Varma, J. Kramár, and N. Nanda, “Jumping ahead: Improving reconstruction fidelity with jumprelu sparse autoencoders,”arXiv preprint arXiv:2407.14435, 2024

    2024 arXiv

  2. [10]

    Batchtopk sparse autoencoders,

    B. Bussmann, P. Leask, and N. Nanda, “Batchtopk sparse autoencoders,” inNeurIPS 2024 Workshop on Scientific Methods for Understanding Deep Learning, 2024

    2024

  3. [11]

    Learning multi-level features with ma- tryoshka sparse autoencoders,

    B. Bussmann, N. Nabeshima, A. Karvonen, and N. Nanda, “Learning multi-level features with ma- tryoshka sparse autoencoders,” inInternational Con- ference on Machine Learning. PMLR, 2025, pp. 6077–6101

    2025

  4. [12]

    From flat to hierarchical: Extracting sparse representations with matching pursuit,

    V . Costa, T. Fel, E. S. Lubana, B. Tolooshams, and D. E. Ba, “From flat to hierarchical: Extracting sparse representations with matching pursuit,” inThe Thirty- ninth Annual Conference on Neural Information Pro- cessing Systems, 2025

    2025

  5. [13]

    Scal- ing and evaluating sparse autoencoders,

    L. Gao, T. D. la Tour, H. Tillman, G. Goh, R. Troll, A. Radford, I. Sutskever, J. Leike, and J. Wu, “Scal- ing and evaluating sparse autoencoders,” inThe Thir- teenth International Conference on Learning Repre- sentations, 2024

    2024

  6. [14]

    Gemma scope: Open sparse autoen- coders everywhere all at once on gemma 2,

    T. Lieberum, S. Rajamanoharan, A. Conmy, L. Smith, N. Sonnerat, V . Varma, J. Kramár, A. Dragan, R. Shah, and N. Nanda, “Gemma scope: Open sparse autoen- coders everywhere all at once on gemma 2,” inPro- ceedings of the 7th BlackboxNLP Workshop: Analyz- ing and Interpreting Ne...

    2024

  7. [15]

    Sparse autoencoders learn monosemantic features in vision-language models,

    M. Pach, S. Karthik, Q. Bouniot, S. Belongie, and Z. Akata, “Sparse autoencoders learn monosemantic features in vision-language models,”arXiv preprint arXiv:2504.02821, 2025

    2025

  8. [16]

    Interpreting clip with hierarchical sparse autoencoders,

    V . Zaigrajew, H. Baniecki, and P. Biecek, “Interpreting clip with hierarchical sparse autoencoders,” inInter- national Conference on Machine Learning. PMLR, 2025, pp. 73 918–73 956

    2025

  9. [17]

    Eval- uating sparse autoencoders for controlling open-ended text generation,

    N. M. A. Makelov, N. Monson, and J. Adebayo, “Eval- uating sparse autoencoders for controlling open-ended text generation,” inSecond NeurIPS Workshop on At- tributing Model Behavior at Scale, 2024

    2024

  10. [18]

    Improving steer- ing vectors by targeting sparse autoencoder features,

    S. Chalnev, M. Siu, and A. Conmy, “Improving steer- ing vectors by targeting sparse autoencoder features,” arXiv preprint arXiv:2411.02193, 2024

    2024 arXiv

  11. [19]

    Denoising concept vectors with sparse autoencoders for improved language model steering,

    H. Zhao, X. Wu, F. Yang, B. Shen, N. Liu, and M. Du, “Denoising concept vectors with sparse autoencoders for improved language model steering,” inFindings of the Association for Computational Linguistics: EACL 2026, 2026, pp. 797–808

    2026

  12. [20]

    In- terpreting and steering llms with mutual information- based explanations on sparse autoencoders,

    X. Wu, J. Yuan, W. Yao, X. Zhai, and N. Liu, “In- terpreting and steering llms with mutual information- based explanations on sparse autoencoders,” in NeurIPS 2025 Workshop on Regulatable ML, 2025

    2025

  13. [21]

    Concept steerers: Lever- aging k-sparse autoencoders for test-time controllable generations,

    D. Kim and D. Ghadiyaram, “Concept steerers: Lever- aging k-sparse autoencoders for test-time controllable generations,”arXiv preprint arXiv:2501.19066, 2025

    2025

  14. [22]

    Interpret and control dense retrieval with sparse latent features,

    H. Kang, T. Wang, and C. Xiong, “Interpret and control dense retrieval with sparse latent features,” inProceed- ings of the 2025 Conference of the Nations of the Amer- icas Chapter of the Association for Computational Lin- guistics: Human Language Technologies (Volume 2: Short...

    2025

  15. [23]

    Decoding dense embed- dings: Sparse autoencoders for interpreting and dis- cretizing dense retrieval,

    S. Park, T. Kim, and Y . Ko, “Decoding dense embed- dings: Sparse autoencoders for interpreting and dis- cretizing dense retrieval,” inProceedings of the 2025 Conference on Empirical Methods in Natural Lan- guage Processing, 2025, pp. 26 479–26 496

    2025

  16. [24]

    Discovering and steering interpretable concepts in large generative mu- sic models,

    N. Singh, M. Cherep, and P. Maes, “Discovering and steering interpretable concepts in large generative mu- sic models,” inNeurIPS AI for Music Workshop, 2025

    2025

  17. [25]

    Universal sparse autoencoders: In- terpretable cross-model concept alignment,

    H. Thasarathan, J. Forsyth, T. Fel, M. Kowal, and K. G. Derpanis, “Universal sparse autoencoders: In- terpretable cross-model concept alignment,” inForty- second International Conference on Machine Learn- ing, 2025

    2025

  18. [26]

    Interpreting the linear structure of vision- language model embedding spaces,

    I. Papadimitriou, H. Su, T. Fel, S. M. Kakade, and S. Gil, “Interpreting the linear structure of vision- language model embedding spaces,” inSecond Con- ference on Language Modeling, 2025

    2025

  19. [27]

    Disentangling dense embeddings with sparse autoencoders,

    C. O’Neill, C. Ye, K. Iyer, and J. F. Wu, “Disentangling dense embeddings with sparse autoencoders,”arXiv preprint arXiv:2408.00657, 2024

    2024 arXiv

  20. [28]

    Decomposing multimodal embedding spaces with group-sparse au- toencoders,

    C. Kaushik, D. Barch, and A. Fanelli, “Decomposing multimodal embedding spaces with group-sparse au- toencoders,”arXiv preprint arXiv:2601.20028, 2026

    2026

  21. [29]

    Discover-then-name: Task-agnostic concept bottle- necks via automated concept discovery,

    S. Rao, S. Mahajan, M. Böhle, and B. Schiele, “Discover-then-name: Task-agnostic concept bottle- necks via automated concept discovery,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 444–461

    2024

  22. [30]

    From what to how: At- tributing clip’s latent components reveals unexpected semantic reliance,

    M. Dreyer, L. Hufe, J. Berend, T. Wiegand, S. La- puschkin, and W. Samek, “From what to how: At- tributing clip’s latent components reveals unexpected semantic reliance,”arXiv preprint arXiv:2505.20229, 2025

    2025 arXiv

  23. [31]

    Sparse autoencoders do not find canonical units of analysis,

    P. Leask, B. Bussmann, M. T. Pearce, J. I. Bloom, C. Tigges, N. Al Moubayed, L. Sharkey, and N. Nanda, “Sparse autoencoders do not find canonical units of analysis,” inThe Thirteenth International Conference on Learning Representations, 2025

    2025

  24. [32]

    A is for absorption: Studying feature splitting and absorption in sparse au- toencoders,

    D. Chanin, J. Wilken-Smith, T. Dulka, H. Bhatnagar, S. Golechha, and J. I. Bloom, “A is for absorption: Studying feature splitting and absorption in sparse au- toencoders,” inThe Thirty-ninth Annual Conference on Neural Information Processing Systems, 2024

    2024

  25. [33]

    Ortsae: Orthogonal sparse autoencoders uncover atomic features,

    A. Korznikov, A. Galichin, A. Dontsov, O. Rogov, E. Tutubalina, and I. Oseledets, “Ortsae: Orthogonal sparse autoencoders uncover atomic features,”arXiv preprint arXiv:2509.22033, 2025

    2025

  26. [34]

    Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning,

    W. Liang, Y . Zhang, Y . Kwon, S. Yeung, and J. Zou, “Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning,” in Advances in Neural Information Processing Systems, vol. 35, 2022

    2022

  27. [35]

    Sparse autoencoders find highly inter- pretable features in language models,

    H. Cunningham, A. Ewart, L. Riggs, R. Huben, and L. Sharkey, “Sparse autoencoders find highly inter- pretable features in language models,”arXiv preprint arXiv:2309.08600, 2023

    2023 arXiv

  28. [36]

    Large-scale contrastive language- audio pretraining with feature fusion and keyword-to- caption augmentation,

    Y . Wu, K. Chen, T. Zhang, Y . Hui, T. Berg-Kirkpatrick, and S. Dubnov, “Large-scale contrastive language- audio pretraining with feature fusion and keyword-to- caption augmentation,” inICASSP 2023-2023 IEEE In- ternational Conference on Acoustics, Speech and Sig- nal Processi...

    2023

  29. [37]

    Un- derstanding the modality gap in clip,

    P. Y . Shi, M. Welle, M. Björkman, and D. Kragic, “Un- derstanding the modality gap in clip,” inInternational Conference on Learning Representations, 2023

    2023

  30. [38]

    Slap: Siamese language-audio pretraining without negative samples for music understanding,

    J. Guinot, A. Riou, E. Quinton, and G. Fazekas, “Slap: Siamese language-audio pretraining without negative samples for music understanding,” inProceedings of the 26th International Society for Music Information Retrieval Conference (ISMIR), Daejeon, South Korea, September 2025

    2025

  31. [39]

    An analysis of variance test for normality (complete samples),

    S. S. Shapiro and M. B. Wilk, “An analysis of variance test for normality (complete samples),”Biometrika, vol. 52, no. 3/4, pp. 591–611, 1965

    1965

  32. [40]

    Jamendo- maxcaps: A large scale music-caption dataset with im- puted metadata,

    A. Roy, R. Liu, T. Lu, and D. Herremans, “Jamendo- maxcaps: A large scale music-caption dataset with im- puted metadata,” in2025 International Joint Confer- ence on Neural Networks (IJCNN). IEEE, 2025, pp. 1–8

    2025

  33. [41]

    Muq: Self-supervised music representation learning with mel residual vector quan- tization,

    H. Zhu, Y . Zhou, H. Chen, J. Yu, Z. Ma, R. Gu, Y . Luo, W. Tan, and X. Chen, “Muq: Self-supervised music representation learning with mel residual vector quan- tization,”IEEE Transactions on Audio, Speech and Language Processing, 2025

    2025

  34. [42]

    Cross the gap: Exposing the intra- modal misalignment in clip via modality inversion,

    M. Mistretta, A. Baldrati, L. Agnolucci, M. Bertini, A. Bagdanovet al., “Cross the gap: Exposing the intra- modal misalignment in clip via modality inversion,” in13th International Conference on Learning Repre- sentations, ICLR 2025. International Conference on Learning Repres...

    2025

  35. [43]

    Training-free diffusion priors for text-to-image generation via optimization-based visual inversion,

    S. Dell’Erba and A. D. Bagdanov, “Training-free diffusion priors for text-to-image generation via optimization-based visual inversion,”arXiv preprint arXiv:2511.20821, 2025

    2025

  36. [44]

    Adam: A method for stochastic optimization,

    D. P. Kingma and J. L. Ba, “Adam: A method for stochastic optimization,” in3rd International Conference on Learning Representations, ICLR 2015 - Conference Track Proceedings, 12 2015. [Online]. Available: http://arxiv.org/abs/1412.6980

    2015 arXiv

  37. [45]

    A fast iterative shrinkage- thresholding algorithm for linear inverse problems,

    A. Beck and M. Teboulle, “A fast iterative shrinkage- thresholding algorithm for linear inverse problems,” Society for Industrial and Applied Mathematics, vol. 2, pp. 183–202, 2009

    2009

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.