{"id":"e2c6b2dd-73c3-41c9-b5f0-4387c1a544bb","arxiv_id":"2501.09877","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"CLAP-S and CLAP-S+ adapt CLAP models to fiber-optic acoustic recognition by combining support-set retrieval with a fine-tuned adapter, reporting improved few-shot classification accuracy over existing methods.","lead":"This paper introduces CLAP-S and CLAP-S+, methods that adapt a pretrained audio-language model to classify sounds recorded by fiber-optic sensors using a small support set. It reports accuracy gains over several baselines on new fiber-optic ESC-50 and gunshot-firework datasets and concludes that zero-shot text knowledge can harm performance in this domain.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'highest accuracy' claim rests on unreported tuning of α and β in Eq. (1); if these were chosen on the test sets, the reported superiority of CLAP-S+ could be inflated.","rationale":"The reader's named weakest assumption—that the CLAP embedding geometry remains class-discriminative after adapter fine-tuning—is partially supported by the paper's own ablations: Table V shows that adding support-set retrieval to the Adapter raises average accuracy from 76.4% to 77.6% with text-aligned keys and to 78.8% with task-aligned keys. If the geometry had collapsed, the retrieval term would not help, so similarity structure transfer is at least indirectly demonstrated at the level of final accuracy. A less protected premise is the selection of α and β in Eq. (1). These scalars determine the magnitude of the retrieval contribution, and the paper reports neither their values nor the selection protocol, despite having a validation split (Table III). If α/β were tuned per dataset on the test labels, the reported average could be inflated, and the small gains over Tip-Adapter-F on FMO/FCO could disappear under an honest protocol. The Prompt Tuning baseline on real tasks is implausibly low (below chance for 8 classes), which suggests a configuration issue, but removing it would not change the result that CLAP-S+ beats Adapter and Tip-Adapter-F; it does, however, strengthen the case for full disclosure of training and tuning procedures. The proposed concrete test would settle whether the central claim is an artifact. The reader's CONDITIONAL verdict remains appropriate, so the verdict is unchanged.","tokens_in":9079,"tokens_out":12091,"duration_ms":125026,"concrete_test":"Obtain the α and β values used for each dataset, and rerun the Table II comparison with α and β selected on the validation split rather than on the test set, using the same adapter training (e.g., grid search over α ∈ {0, 0.2, 0.5, 0.8, 1} and β ∈ {1, 10, 20, 50} by validation accuracy). If the resulting average accuracy of CLAP-S+ across ECM, FM, FC, FMO, and FCO falls below Adapter or Tip-Adapter-F, the 'highest accuracy among all baseline methods' claim is not supported. A sensitivity table around the chosen α and β would further clarify whether the result is robust.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim is the Table II result that CLAP-S+ has the highest average accuracy among all studied baselines. This is only trustworthy if the two scalars in Eq. (1), α and β, were selected without using test labels. The paper defines α as 'a tuned parameter' and β as 'the sharpness parameter' (Section II) but never reports their values, never states whether they were chosen on the validation split, the test split, or fixed across datasets, and gives no sensitivity analysis. Since p_support = e^{-β(1-uF_train^T)} L_train^T, β controls the sharpness of the retrieval weights and α controls the interpolation weight, so small changes in either can change accuracy by several points on the low-shot real tasks, where the advantage over Tip-Adapter-F is only +1.2 to +2.4 percentage points. If α and β were tuned per dataset on the test set, the 78.8% average in Table II could be an optimistic selection artifact, and CLAP-S+ might no longer beat Adapter (76.4%) or Tip-Adapter-F (76.5%) under a validation-based protocol. The suspiciously low Prompt Tuning baseline on the real tasks (4.0±4.0%, 5.0±5.2%, below 12.5% chance for 8 classes) does not by itself overturn the claim because CLAP-S+ also beats the stronger baselines, but it reinforces the need for a disclosed and fair hyperparameter selection procedure.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CLAP-S and CLAP-S+, two adaptation methods for using the CLAP audio-language model on fiber-optic acoustic recognition, a domain with significant distribution shift and limited labeled data. CLAP-S performs training-free classification by retrieving labels from a support set via cross-attention between test and training audio embeddings. CLAP-S+ combines a fine-tuned two-layer MLP adapter with support-set retrieval, interpolating the adapter's prediction and the retrieval prediction through a weighted sum. The authors evaluate on laboratory-recorded fiber-optic versions of ESC-50 and a real-world gunshot-firework dataset, reporting that CLAP-S+ achieves the highest average accuracy among the studied baselines (Prompt Tuning, Adapter, Treff, Tip-Adapter, Tip-Adapter-F), along with efficiency comparisons and ablations on the contributions of zero-shot knowledge and task-aligned embeddings.","tokens_in":9349,"tokens_out":3351,"duration_ms":33619,"significance":"If the empirical claims hold, the paper offers a simple, practical recipe for adapting a general-purpose audio foundation model to a sensor domain with a large domain gap: fine-tune a small adapter on the support set and use the task-aligned embeddings for both classification and support-set retrieval. The paper also introduces a new real-world DAS gunshot-firework dataset and releases code, which are valuable resources. The conceptual framing of 'implicit' versus 'explicit' knowledge is useful, and the ablation showing that zero-shot text-aligned knowledge can hurt in this domain is a noteworthy caveat for practitioners. However, the strength of the central 'highest accuracy' claim is currently limited by unreported hyperparameter selection, a suspiciously weak Prompt Tuning baseline, and the lack of statistical significance assessment for several small margins.","major_comments":[{"comment":"The interpolation weight alpha and the kernel sharpness beta are described only as \"tuned parameter\" and \"sharpness parameter\" with no reported values, no statement of whether they were selected on the validation splits described in Table III, and no sensitivity analysis. This is load-bearing because the claimed advantages over Tip-Adapter-F are only +2.4 and +1.2 percentage points on FMO and FCO, so the choice of alpha and beta could change the ranking. Please report the per-dataset (or global) values used, specify the selection protocol, and provide a sensitivity study over alpha and beta.","section":"Section II, Eq. (1)"},{"comment":"The Prompt Tuning baseline achieves 4.0±4.0% and 5.0±5.2% on the real-world FMO and FCO tasks, which is below the 12.5% chance level for an 8-class problem. Such a result strongly suggests a misconfigured or undertrained baseline, not a faithful comparison. The paper's hypothesis that Prompt Tuning is ineffective on unseen tasks does not explain performance below chance. As presented, the baseline comparison is unfair, and the statement that CLAP-S+ \"achieves the highest accuracy among all the baseline methods\" is weakened. Please verify the Prompt Tuning implementation (e.g., prompt length, learning rate, epochs, or per-class text initialization) or re-run with a properly tuned configuration.","section":"Section III-B, Table II"},{"comment":"Several of the reported accuracy differences between CLAP-S+ and the strongest baselines are within one standard deviation of the reported runs; for example, on FM the result is 70.0±0.8 versus Adapter's 68.8±1.0 and Tip-Adapter-F's 68.6±1.0, and on FCO it is 92.0±1.7 versus 90.2±0.7. The paper does not report per-seed results, confidence intervals, or any paired significance test. The claim that CLAP-S+ is the best method across all five datasets therefore needs statistical support; please add significance tests or at least report the individual run outcomes so the variability can be assessed.","section":"Section III-B, Table II"}],"minor_comments":[{"comment":"The method is inconsistently referred to as both \"CLAP-S+\" and \"CLAP-S-+\"; please unify the notation.","section":"Throughout"},{"comment":"There is a typo in \"We focusing on how the pre-trained knowledge\" — the verb form should be fixed.","section":"Section I"},{"comment":"The table header contains \"Trainin-Free\" — this should be corrected to \"Training-Free.\"","section":"Section III-B, Table II"},{"comment":"The relation with Tip-Adapter and Treff would be clearer if the table indicated which embedding is used for the query versus keys in each method; the current notation \"u0\" and \"uf\" is not defined inside the table itself.","section":"Section II, Table I"},{"comment":"The claim that task-aligned embeddings are more effective for retrieval is only indirectly supported by final accuracy; reporting a direct retrieval metric (e.g., nearest-neighbor accuracy of the support-set keys) would strengthen this ablation.","section":"Section III-E, Table V"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for an audio/signal-processing venue and the dataset release is a positive contribution. The main concerns are empirical rigor: hyperparameter selection, baseline fairness, and statistical significance. These are fixable with additional experiments and reporting, so I do not recommend rejection. The self-citation to the authors' own DAS dataset is appropriate given that the dataset is genuinely theirs, but the Prompt Tuning baseline issue should be addressed carefully because it invites skepticism about the experimental protocol."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, the real content is two new fiber-optic acoustic datasets (a lab re-recording of ESC-50 through fiber sensors and an outdoor gunshot-firework set from telecom cables) plus a clean ablation showing that zero-shot text-aligned CLAP predictions hurt in this domain. Second, the method itself is Tip-Adapter/Treff with the query and key switched to task-aligned adapter embeddings and an interpolation weight; that combination is sensible and the paper shows it works, but it is not a new mechanism.\n\nWhat is good: the datasets are a real contribution; code and data are linked; the comparison across training-free and training-required baselines is thorough; the ablation in Table V directly tests which embedding to use for retrieval, which is the right question. The finding that zero-shot knowledge can lower accuracy in a shifted domain is worth reporting even if it is domain-specific.\n\nSoft spots, in proportion. The main one is the missing hyperparameter disclosure. The stress-test note is on target: alpha and beta in Eq. (1) are called tuned but no values, no selection protocol, no sensitivity analysis. The gains over Adapter and Tip-Adapter-F are +1.2 to +2.4 points on the real tasks, so if alpha/beta were chosen per dataset on the test set, the headline 'highest accuracy' could be an artifact. That needs to be fixed before the claim is trustworthy. Second, the Prompt Tuning baseline at 4-5% on an 8-class real task is below chance, which makes me think the prompt tuning configuration was broken, not just weak; the authors should show the setup and the learned prompts. It doesn't kill the paper because CLAP-S+ also beats stronger baselines, but it lowers confidence in the whole comparison. Third, several reported advantages are within one standard deviation of the runner-up; some significance or effect-size reporting would help. The untested assumption about embedding geometry is real but minor: final accuracy on unseen classes is the practical test, and it passes.\n\nWho this is for: people working on DAS/fiber-optic audio and anyone doing few-shot adaptation of audio-language models in shifted domains. It deserves a serious referee, but the referee should ask for the hyperparameter protocol and a corrected prompt-tuning baseline. My verdict would be revise, not reject.","headline":"A useful incremental paper with real new datasets and a sensible retrieval-plus-adapter combination; the headline accuracy claim currently rests on undisclosed tuning of alpha and beta.","tokens_in":9862,"tokens_out":1980,"would_cite":true,"duration_ms":20242,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"For fiber-optic acoustic recognition, the best CLAP adaptation combines a fine-tuned adapter with support-set retrieval on task-aligned embeddings.","keywords":["fiber-optic acoustic recognition","distributed acoustic sensing","CLAP","support set adaptation","few-shot learning","domain adaptation","sound event classification","adapter fine-tuning"],"falsifier":"Using the released gunshot-firework dataset and the same train/validation/test splits, rerun CLAP-S+ with the retrieval branch disabled (alpha=0) and with support keys replaced by text-aligned embeddings; if either variant matches or beats the full CLAP-S+, or if the same-class minus different-class cosine similarity among fine-tuned test embeddings is not positive, the paper's claim that task-aligned support-set interpolation drives the gains is refuted.","tokens_in":8848,"feed_emoji":"🔊","tokens_out":10352,"duration_ms":92376,"temperature":0.7,"pith_summary":"The paper tries to establish that the best way to adapt a CLAP audio model to fiber-optic acoustic sensing is not fine-tuning alone and not retrieval alone, but a linear interpolation of a fine-tuned adapter with a support-set retrieval branch. This matters because fiber-optic distributed acoustic sensing is a low-label, high-domain-shift setting where standard microphone-pretrained representations transfer poorly. The authors build a fiber-optic version of ESC-50 and use a real gunshot-versus-firework dataset collected from telecom cables. Their results show the hybrid CLAP-S+ outperforms prompt tuning, adapter, Treff, and Tip-Adapter-F, and that the training-free CLAP-S outperforms Tip-Adapter. They also find that zero-shot text-aligned knowledge can hurt and that task-aligned embeddings give better retrieval.","feed_headline":"Retrieval-plus-adapter beats all baselines on fiber-optic audio","feed_subtitle":"CLAP-S+ pairs fine-tuned embeddings with support-set matching, raising average accuracy to 78.8 percent.","key_machinery":"The machinery is the support set: the $NK$ labeled training audios are stored as key embeddings $F_{\\text{train}}\\in\\mathbb{R}^{NK\\times C}$ with one-hot label values $L_{\\text{train}}$. A test query $u$ produces a retrieval distribution $p_{\\text{support}}(y|x,u)=e^{-\\beta(1-uF_{\\text{train}}^T)}L_{\\text{train}}^T$ via cross-attention, and a two-layer MLP adapter maps the text-aligned embedding $u_0$ to a task-aligned embedding $u_f$. CLAP-S+ sets both query and keys to $u_f$ and outputs $(1-\\alpha)p_{\\text{clap}}+\\alpha p_{\\text{support}}$, while CLAP-S uses $u_0$ with $\\alpha=1$ and no training. This machinery carries the argument because the same labeled samples feed both the adapter and the retrieval memory, and because the ablation ties the gain to using task-aligned, not text-aligned, keys.","core_discovery":"The paper's central claim is that CLAP can be adapted to fiber-optic distributed acoustic sensing by using the small labeled set twice: once to fine-tune a two-layer MLP adapter and once as a key-value support set for retrieval. The final prediction is a linear blend of the adapter's class distribution and the retrieval distribution. The authors report that when query and keys both use the task-aligned adapter embeddings, this blend, CLAP-S+, is the most accurate of the studied methods, averaging 78.8% accuracy across five datasets versus 76.5% for Tip-Adapter-F, the strongest baseline; with no training at all, the pure-retrieval CLAP-S reaches 71.6% and beats Tip-Adapter's 69.2%.","pith_inferences":["A natural extension the paper does not run is a systematic sweep of domain-shift severity; the larger gains on the outdoor real-world tasks suggest the retrieval branch matters most when CLAP's text-aligned space is most distorted.","Because alpha is selected on validation labels, a deployment version would need a label-free rule for alpha; one testable option is to set it from retrieval confidence or from the entropy of the support-set distribution.","If the mechanism transfers, the same adapter-plus-support-set recipe could be tried for other acoustic domains far from microphone pretraining, such as hydrophone or respiratory sound classification, where labels are scarce and language descriptions are weak.","The paper's negative result for zero-shot knowledge implies that removing the text branch entirely, or replacing class names with learned or noisy captions, could further improve adaptation; that is a direct testable consequence."],"forward_implications":["On the five datasets reported, CLAP-S+ raises average accuracy to 78.8%, the best among prompt tuning, adapter, Treff, and Tip-Adapter-F.","Without any training, CLAP-S reaches 71.6% average accuracy and beats Tip-Adapter (69.2%) while using no extra parameters and slightly less inference time.","Task-aligned adapter embeddings outperform text-aligned embeddings as support-set keys and queries in this domain.","Adding zero-shot CLAP predictions to either the support-set or adapter predictions can reduce accuracy, so zero-shot knowledge is not always beneficial under large domain shift.","Training one adapter jointly on multiple fiber-optic device domains outperforms training separate adapters per domain."],"supporting_citations":[{"why":"Introduces the CLAP contrastive language-audio pretraining model whose frozen encoders are the starting point for every variant.","marker":"[13]"},{"why":"Describes natural-language-supervised audio representations; supplies the text-aligned embedding space that CLAP-S and the zero-shot baselines rely on.","marker":"[14]"},{"why":"Prompt tuning baseline that the paper compares against and finds ineffective on the real-world fiber-optic tasks.","marker":"[15]"},{"why":"Adapter baseline whose two-layer MLP architecture CLAP-S+ extends with the support-set retrieval branch.","marker":"[19]"},{"why":"Tip-Adapter, the training-free and trainable support-set baselines that CLAP-S and CLAP-S+ are directly compared with and improve upon.","marker":"[20]"},{"why":"Treff, the few-shot audio-learner baseline that motivates using task-aligned embeddings for adaptation.","marker":"[25]"},{"why":"ESC-50, the environmental sound benchmark re-recorded through fiber-optic sensors to build the laboratory datasets.","marker":"[26]"},{"why":"The real-world gunshot-firework classification dataset from telecom distributed acoustic sensing, used for FMO and FCO evaluation.","marker":"[27]"},{"why":"Cross-attention mechanism used to turn query-key similarities into the support-set retrieval distribution.","marker":"[28]"}],"fun_headline_variants":["Support-set adaptation boosts fiber-optic acoustic recognition","CLAP-S+ mixes adapter and support set for 78.8% accuracy","Retrieval-augmented adapter tops fiber-optic audio benchmarks","Support set as memory improves CLAP for fiber-optic sensing","Blending memory and adapter lifts fiber-optic audio accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that after adapter fine-tuning, the CLAP audio encoder's embedding space keeps same-class fiber-optic sounds closer together than different-class sounds, so nearest-neighbor retrieval from the support set is a trustworthy signal.","fun_headline_variants_meta":{"raw":{"variants":["Support-set adaptation boosts fiber-optic acoustic recognition","CLAP-S+ mixes adapter and support set for 78.8% accuracy","Retrieval-augmented adapter tops fiber-optic audio benchmarks","Support set as memory improves CLAP for fiber-optic sensing","Blending memory and adapter lifts fiber-optic audio accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001089,"raw_usage":{"total_tokens":4526,"prompt_tokens":897,"completion_tokens":3629,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":513,"completion_tokens_details":{"reasoning_tokens":3542}},"tokens_in":513,"tokens_out":3629,"duration_ms":27481,"temperature":1.0,"reasoning_tokens":3542,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T19:35:16.928682+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Using the released gunshot-firework dataset and the same train/validation/test splits, rerun CLAP-S+ with the retrieval branch disabled (alpha=0) and with support keys replaced by text-aligned embeddings; if either variant matches or beats the full CLAP-S+, or if the same-class minus different-class cosine similarity among fine-tuned test embeddings is not positive, the paper's claim that task-aligned support-set interpolation drives the gains is refuted.","supporting_citations":[{"cited_title":"Clap learning audio concepts from natural language supervision,","cited_arxiv_id":null,"evidence_quote":"Introduces the CLAP contrastive language-audio pretraining model whose frozen encoders are the starting point for every variant."},{"cited_title":"Natural language supervision for general-purpose audio representations,","cited_arxiv_id":null,"evidence_quote":"Describes natural-language-supervised audio representations; supplies the text-aligned embedding space that CLAP-S and the zero-shot baselines rely on."},{"cited_title":"Learning to prompt for vision-language models,","cited_arxiv_id":null,"evidence_quote":"Prompt tuning baseline that the paper compares against and finds ineffective on the real-world fiber-optic tasks."},{"cited_title":"Clip-adapter: Better vision-language models with feature adapters,","cited_arxiv_id":null,"evidence_quote":"Adapter baseline whose two-layer MLP architecture CLAP-S+ extends with the support-set retrieval branch."},{"cited_title":"Adapting Language-Audio Models as Few-Shot Audio Learners","cited_arxiv_id":"2305.17719","evidence_quote":"Treff, the few-shot audio-learner baseline that motivates using task-aligned embeddings for adaptation."},{"cited_title":"ESC: Dataset for Environmental Sound Classification,","cited_arxiv_id":null,"evidence_quote":"ESC-50, the environmental sound benchmark re-recorded through fiber-optic sensors to build the laboratory datasets."},{"cited_title":"Deep learning-based intrusion detection and impulsive event classification for distributed acoustic sensing across telecom networks,","cited_arxiv_id":null,"evidence_quote":"The real-world gunshot-firework classification dataset from telecom distributed acoustic sensing, used for FMO and FCO evaluation."},{"cited_title":"Attention is all you need,","cited_arxiv_id":null,"evidence_quote":"Cross-attention mechanism used to turn query-key similarities into the support-set retrieval distribution."}],"review_version":1}