{"id":"83360940-1b2f-44a9-a911-72f728ebb261","arxiv_id":"2608.08757","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Sparse inversion of text concepts into audio SAE feature space yields audio-faithful concept supports that improve edit-preservation trade-offs in steerable music retrieval.","lead":"This paper introduces a training-free method for finding which audio-side sparse features correspond to a text concept, by inverting the concept embedding through a sparse autoencoder trained on audio. The method improves steerable music retrieval, letting users amplify or suppress concepts like 'piano' or 'ambient' with less drift than text-alignment baselines.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Missing audio-side baseline: reported gains may come from nearest-audio-neighbor initialization rather than from sparse inversion, so the central 'geometric inverse problem' claim is not yet isolated.","rationale":"The reader's weakest assumption concerns whether a text concept embedding z_c can be representably reconstructed from the audio dictionary. My concern is complementary but arguably more load-bearing: even granting representability, the experiments do not establish that the sparse inversion, rather than the audio-side priors bundled into it, produces the gains. All three extra ingredients (Gamma, neighbor initialization, IDF) are audio-grounded, whereas the only baseline is a text-side ranking method. If a simple nearest-audio-neighbor support baseline performs comparably, then the 'inverse problem' is a complicated way to approximate retrieval-by-example support, and the conceptual claim is not supported. I do not see a reason to reject the paper, but this control is needed before the strongest claim can be accepted as stated. The reader already issued CONDITIONAL, and the missing baseline refines that condition without moving the verdict; hence UNCHANGED. Credit is due for releasing code, using parameter-free experimental controls like cardinality matching, and acknowledging the Mahalanobis diagnostic circularity; the concern is specifically about an omitted comparison rather than an internal inconsistency.","tokens_in":11054,"tokens_out":5056,"duration_ms":58281,"concrete_test":"Re-run the Section 4 and Section 5 protocols with two additional baselines. First, use a 'nearest-neighbor support' baseline: for each concept c, take the audio embedding in the evaluation/test set nearest to z_c under cosine similarity, encode it with the SAE, use its active support (cardinality matched to L0), apply the same IDF weighting, and evaluate bundle recall, probe AUROC, and the edit-preservation frontiers. Second, ablate Eq. (4): set gamma=0 and, separately, initialize u from a random audio embedding instead of the nearest neighbor. If the nearest-neighbor support baseline matches the inversion results within error bars, then the geometric-inverse framing is not the source of the gains. If the baseline is clearly worse and removing the neighbor initialization degrades performance, the concern does not land.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing weakness is that the paper never isolates the inversion as the causal ingredient. The proposed pipeline adds audio-side information at three points: (i) the Mahalanobis regularizer Gamma(A, D(Pi(u))) in Eq. (4), estimated from test-set audio embeddings; (ii) the initialization of u from the sparse code of the nearest audio neighbor to z_c, described in Section 3.3; and (iii) the IDF reweighting in Eq. (5), computed from audio activation frequencies. The experiments compare only against DTN cosine probing, which uses no audio-side information. Hence the reported 10-70% bundle-recall gains and improved Pareto frontiers could be produced entirely by the nearest-neighbor initialization or by the distributional regularizer, rather than by the sparse inversion objective itself. In the extreme case, taking the SAE support of the nearest audio embedding to z_c and discarding Eq. (4) entirely might match or beat the inversion results. That would not invalidate the practical recipe, but it would undermine the paper's central claim that concept attribution 'should be treated as a geometric inverse problem'. The paper contains no ablation that removes Gamma, removes the neighbor initialization, or compares with such an audio-side support baseline. This is an experimental omission, not an internal contradiction, but it is directly load-bearing for the strongest claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":11347,"tokens_out":7770,"duration_ms":82546,"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":[{"comment":"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.","section":"§3.3, Eq. (4), §5.3"},{"comment":"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.","section":"§4.2, Figs. 3–5"},{"comment":"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.","section":"§3.3"},{"comment":"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.","section":"§3.3, §5.2"}],"minor_comments":[{"comment":"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.","section":"§5.2, §5.3"},{"comment":"The constant ϵ in the denominator df(k)+ϵ is never defined; please specify its value.","section":"§3.3, Eq. (5)"},{"comment":"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.","section":"§4, Fig. 3"},{"comment":"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.","section":"§5.1, Eqs. (6)–(7)"},{"comment":"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.","section":"§4.1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within scope and the core methodological idea is interesting. The revision should focus on isolating the inversion objective from the initialization/regularizer and on addressing the test-set leakage in the Mahalanobis estimate; with those addressed, I expect the claims to be supportable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Guinot et al. treat concept attribution in audio SAE spaces as a sparse inverse problem: instead of ranking decoder atoms by cosine alignment to the text embedding (DTN), they solve a small optimization that reconstructs the text direction from audio-side atoms under a Mahalanobis manifold regularizer. The headline result is that their recovered supports overlap much better with the active supports of concept-bearing audio examples (10-70% higher bundle recall) and yield better edit-preservation frontiers in steerable retrieval. That is a genuine, useful empirical contribution, and the evaluation is otherwise careful: cardinality-matched supports, external tag-based metrics, and a held-out audio-vs-text classifier for preservation. The two inversion variants (Adam and FISTA) both beat DTN, which suggests something about the inversion objective is helping, not just a single implementation choice.\n\nThe soft spot is the missing ablation. The pipeline adds audio-side information in three places: the Mahalanobis regularizer, the initialization from the nearest audio neighbor's sparse code, and IDF reweighting. The experiments compare only against DTN, which uses none of these. So the reported gains could in principle come entirely from the nearest-audio-neighbor initialization or the regularizer, and a baseline that simply takes the sparse support of the nearest audio embedding to the text query, then applies IDF and cardinality matching, is not in the paper. That baseline is cheap and would isolate the inversion objective. Without it, the central claim \"attribution should be treated as a geometric inverse problem\" is underdetermined. The paper says FISTA also works, but FISTA is still an inversion method; it doesn't remove the initialization dependency. This is an experimental omission, not an internal contradiction, but it is directly load-bearing.\n\nOther issues are minor by comparison: no error bars or multiple seeds, so we can't assess variance; the FISTA application to a non-convex objective (because of the sparsifying map) lacks theoretical justification; the internal taxonomy used to select concepts is not released. The disclosed Mahalanobis circularity is genuinely minor because it's used as a diagnostic and the main metrics are independent.\n\nBottom line: this paper deserves a serious referee. It's a solid empirical study with a clear formulation and a plausible central claim, but the missing baseline needs to be run before the claim is fully established. If the nearest-audio-support baseline matches the inversion, the contribution shrinks to a practical recipe; if it doesn't, the geometric-inversion framing is validated. Either way, the paper is worth discussing.","headline":"Solid empirical paper on sparse-inversion concept attribution for music retrieval, but the missing nearest-audio-support baseline leaves the central claim underdetermined.","tokens_in":11835,"tokens_out":3026,"would_cite":true,"duration_ms":30584,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Treating concept attribution as a geometric inverse problem recovers the audio-side features that text-aligned neuron ranking misses.","keywords":["sparse autoencoders","concept attribution","dense music retrieval","controllable retrieval","open-vocabulary concepts","sparse inversion","multimodal embeddings","modality gap"],"falsifier":"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.","tokens_in":10880,"feed_emoji":"🎧","tokens_out":7485,"duration_ms":66517,"temperature":0.7,"pith_summary":"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.","feed_headline":"Sparse inversion finds the audio features text probing misses","feed_subtitle":"Concept edits like 'more piano' or 'no guitar' work better when supports are recovered by geometry, not wording.","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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 $\\gamma$."],"forward_implications":["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."],"supporting_citations":[{"why":"The Discover-Then-Name cosine-probing baseline that the paper's inversion is compared against and shown to improve on.","marker":"[21]"},{"why":"Establishes the modality gap used to explain why text-side neuron ranking fails to recover audio-side support.","marker":"[26]"},{"why":"Supplies the BatchTopK sparse autoencoder variant whose decoder dictionary is the basis for the recovered supports.","marker":"[2]"},{"why":"Supplies the lightweight Mahalanobis manifold regularizer that keeps inverted codes near the audio distribution.","marker":"[35]"},{"why":"Provides the FISTA solver used for the fast sparse linear inversion in Eq. (4).","marker":"[37]"},{"why":"Provides the music-caption dataset and tags used to train the SAEs and evaluate concept supports.","marker":"[32]"},{"why":"Supplies one of the joint music-text encoders whose audio embeddings define the sparse space.","marker":"[33]"},{"why":"Supplies the contrastive language-audio encoder used for the shared text and audio embeddings.","marker":"[28]"}],"fun_headline_variants":["Concept edits recover audio geometry, not just wording","Sparse inversion beats text matching for music concept edits","Geometry, not wording, finds audio features for concept control","Solving sparse inversion recovers concept-supporting audio features","Audio-space sparse inversion outperforms text-side probes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["Concept edits recover audio geometry, not just wording","Sparse inversion beats text matching for music concept edits","Geometry, not wording, finds audio features for concept control","Solving sparse inversion recovers concept-supporting audio features","Audio-space sparse inversion outperforms text-side probes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00057,"raw_usage":{"total_tokens":2727,"prompt_tokens":1003,"completion_tokens":1724,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":619,"completion_tokens_details":{"reasoning_tokens":1649}},"tokens_in":619,"tokens_out":1724,"duration_ms":11454,"temperature":1.0,"reasoning_tokens":1649,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T04:24:52.420887+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":"Interpreting the linear structure of vision- language model embedding spaces,","cited_arxiv_id":null,"evidence_quote":"Establishes the modality gap used to explain why text-side neuron ranking fails to recover audio-side support."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the BatchTopK sparse autoencoder variant whose decoder dictionary is the basis for the recovered supports."},{"cited_title":"Un- derstanding the modality gap in clip,","cited_arxiv_id":null,"evidence_quote":"Provides the FISTA solver used for the fast sparse linear inversion in Eq. (4)."},{"cited_title":"A is for absorption: Studying feature splitting and absorption in sparse au- toencoders,","cited_arxiv_id":null,"evidence_quote":"Provides the music-caption dataset and tags used to train the SAEs and evaluate concept supports."},{"cited_title":"Decomposing multimodal embedding spaces with group-sparse au- toencoders,","cited_arxiv_id":null,"evidence_quote":"Supplies the contrastive language-audio encoder used for the shared text and audio embeddings."}],"review_version":1}