{"id":"26290a68-5f5b-4e5e-9fc6-a75b88b95675","arxiv_id":"2606.18833","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"Introduces the Perception algorithm for seed-guided semi-supervised clustering via a-contrario anomaly detection, defining clusters as anomaly-free subsets and achieving competitive performance with 10-30 seeds per cluster on benchmarks.","lead":"The paper presents a semi-supervised clustering method that defines clusters as maximal subsets containing no anomalies under a uniform randomness null hypothesis, using a seed-guided iterative exclusion process. A smart generalist might read it for a statistically grounded approach to noisy data grouping that requires few labels and avoids manual parameter tuning.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"Uniform randomness null may mismatch structured data, undermining anomaly-based cluster definition","rationale":"The reader's weakest_assumption is precisely the load-bearing statistical modeling choice; the empirical claims rest on it holding for the evaluated datasets. Full-text verification of the Perception algorithm would only strengthen this diagnosis. No other internal inconsistency is visible from the abstract-level description.","tokens_in":1767,"tokens_out":295,"duration_ms":16171,"concrete_test":"Generate synthetic data from a 2-D Gaussian mixture embedded in 50-D space (instead of uniform), run the Perception step with the paper's E<1 rule on the same seed sets, and measure the fraction of true inliers incorrectly excluded; if >15% of inliers are rejected the headline performance claims weaken.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central construction treats a cluster as a maximal subset containing no anomalies w.r.t. a uniform-randomness null, with the Perception algorithm admitting points only when E<1. For seed-guided expansion and noise isolation to succeed on image/text embeddings (as claimed), this null must produce anomaly scores that align with the actual data-generating process. In high-D or manifold-structured regimes the uniform model typically fails to capture local density variations or distance distributions, so the E<1 threshold can systematically misclassify inliers as anomalies (or vice versa), breaking the claimed duality and the low-seed performance.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper proposes a seed-guided semi-supervised clustering method based on a-contrario anomaly detection. Clusters are defined as maximal subsets containing no anomalies w.r.t. a uniform-randomness null hypothesis. The Perception algorithm uses an expectation-based threshold (E < 1) to admit points during iterative 'clustering-by-exclusion' expansion from seed-initialized medians. The central claim is that with 10-30 seeds per cluster the method achieves competitive or strong performance on synthetic and real (image/text embedding) benchmarks under low-tuning protocols while scaling linearly in observations and dimensionality (for fixed seeded clusters and iterations).","tokens_in":1888,"tokens_out":456,"duration_ms":14247,"significance":"If the central construction holds, the approach supplies a statistically grounded, low-parameter alternative to standard semi-supervised clustering that naturally isolates noise and unknown clusters. The reported low-seed performance and linear scaling would be practically relevant for embedding-based tasks.","major_comments":[{"comment":"The uniform-randomness null underlying the Perception algorithm and the E < 1 threshold (central to the cluster definition and duality claim) is load-bearing for all performance results. In high-dimensional or manifold-structured embeddings the null typically fails to model local density or distance distributions, which risks systematic misclassification of inliers as anomalies and thereby undermines the claimed seed-guided expansion and noise isolation.","section":"Cluster definition and Perception algorithm (abstract and §3)"},{"comment":"The experimental claims of 'competitive and often very strong performance' with 10-30 seeds rest on the appropriateness of the null; without explicit validation (e.g., comparison of anomaly scores against data-generating process on the image/text embeddings) the low-tuning benchmarking protocol cannot be taken as supporting the central claim.","section":"Evaluation section (synthetic and real-world benchmarks)"}],"minor_comments":[{"comment":"The abstract states linear scalability 'with respect to both observations and dimensionality for a fixed number of seeded clusters and iterations'; this qualifier should appear in the main text when the complexity analysis is presented.","section":"Complexity analysis"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive comments on the statistical foundations of the Perception algorithm. We address each major point below and outline planned revisions.","responses":[{"response":"The a-contrario framework deliberately employs a simple uniform null to enable parameter-free detection via the E < 1 expectation threshold, which bounds false positives regardless of moderate model mismatch. While we recognize that high-dimensional embeddings may exhibit concentration effects not captured by the null, the seed-guided iterative expansion and empirical results on image/text embeddings demonstrate effective cluster formation and noise isolation. In revision we will augment §3 with explicit discussion of the null's limitations on manifolds and the robustness properties of the E < 1 criterion.","revision_made":"partial","referee_comment":"[Cluster definition and Perception algorithm (abstract and §3)] The uniform-randomness null underlying the Perception algorithm and the E < 1 threshold (central to the cluster definition and duality claim) is load-bearing for all performance results. In high-dimensional or manifold-structured embeddings the null typically fails to model local density or distance distributions, which risks systematic misclassification of inliers as anomalies and thereby undermines the claimed seed-guided expansion and noise isolation."},{"response":"On synthetic data the anomaly scores are directly evaluated against the known generative process, supporting the claimed behavior. For real embeddings we will add in the revised evaluation section a qualitative examination of flagged anomalies (e.g., their correspondence to known outliers or fringe points) together with quantitative checks on a subset of benchmarks. This will provide the requested validation while preserving the low-tuning protocol.","revision_made":"yes","referee_comment":"[Evaluation section (synthetic and real-world benchmarks)] The experimental claims of 'competitive and often very strong performance' with 10-30 seeds rest on the appropriateness of the null; without explicit validation (e.g., comparison of anomaly scores against data-generating process on the image/text embeddings) the low-tuning benchmarking protocol cannot be taken as supporting the central claim."}],"tokens_in":1430,"tokens_out":430,"duration_ms":18374,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The main point is a seed-guided method that treats clusters as maximal anomaly-free sets under uniform randomness, using an iterative Perception algorithm with an E < 1 threshold to expand from seeds while excluding noise. It reports competitive results on image and text embeddings with 10-30 seeds and linear scaling.\n\nWhat is new is the explicit duality between grouping and a-contrario anomaly detection, plus the clustering-by-exclusion loop that starts from user seeds and isolates fringe points without global parameters. The Gestalt proximity angle and the expectation-based threshold give it a distinct statistical flavor compared with standard partitioning or density approaches.\n\nIt handles the noise-over-assignment problem in a principled way and the low-tuning protocol is a practical plus. The scalability claim for fixed clusters and iterations is straightforward to see from the description.\n\nThe soft spot is the uniform-randomness null itself. Embeddings from images or text are rarely uniform; they live on manifolds or have local density structure, so the E < 1 rule can mislabel inliers as anomalies or miss real outliers. The abstract gives no derivations, error bars, or dataset specifics, so it is impossible to check whether the claimed performance survives that mismatch. The stress-test concern lands directly on the central construction.\n\nThis is for people working on robust semi-supervised clustering who want a statistically motivated alternative to k-means or DBSCAN variants. A reader already familiar with a-contrario methods would see the most value.\n\nIt deserves a serious referee because the framing is distinct and the low-seed, low-tuning angle is worth checking in detail, even if the null hypothesis needs work.","headline":"The paper frames semi-supervised clustering as anomaly detection under a uniform null via a-contrario reasoning and claims low-seed performance, but that null looks mismatched to structured embeddings.","tokens_in":2370,"tokens_out":406,"would_cite":false,"duration_ms":19364,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"A semi-supervised clustering method grows seed groups by excluding anomalies defined against uniform randomness, reaching competitive results with 10-30 labels per cluster.","keywords":["semi-supervised clustering","a-contrario anomaly detection","seed-guided clustering","Perception algorithm","clustering by exclusion","anomaly detection","Gestalt proximity"],"falsifier":"On a synthetic dataset whose outliers are drawn from a clearly non-uniform distribution while the seeded clusters remain compact, measure whether the E < 1 rule still excludes the correct points or begins to assign them incorrectly.","tokens_in":2656,"feed_emoji":"🧩","tokens_out":686,"duration_ms":17843,"temperature":0.7,"pith_summary":"The paper establishes a clustering framework that treats grouping as the dual of anomaly detection, defining a valid cluster as any maximal subset containing no points that count as anomalies under a uniform-randomness null model. It introduces the Perception algorithm, which applies an expectation threshold E less than 1 to flag outliers in a parameter-light way, then uses this rule inside an iterative seed-guided expansion process that adds only non-anomalous points. The approach is shown to isolate noise and unknown clusters automatically while delivering competitive accuracy on image and text benchmarks under a low-tuning protocol. Linear scaling with both the number of points and their dimension is reported once the number of seeded clusters and iterations is fixed.","feed_headline":"10-30 seeds suffice for competitive clustering via anomaly exclusion","feed_subtitle":"Iterative exclusion under uniform-randomness model grows seeded groups while isolating noise without extra tuning.","key_machinery":"The Perception algorithm, which computes an expectation-based threshold E < 1 drawn from a-contrario reasoning to decide whether a point is an outlier relative to a uniform-randomness null hypothesis.","core_discovery":"Clustering is performed by iteratively admitting points that satisfy the a-contrario condition of not being anomalous relative to uniform randomness; the Perception algorithm supplies the E < 1 decision rule that requires no global density or distance parameters, allowing a small set of user seeds to initialise medians that are then expanded into robust clusters while fringe and noise points are left unassigned.","pith_inferences":["The uniform-randomness null may need replacement by a data-specific background model when clusters sit in highly structured manifolds.","Because expansion decisions are local and expectation-based, the procedure could be adapted to an online setting where new points arrive sequentially.","The duality between clustering and anomaly detection suggests the same machinery could be run in reverse to surface anomalies once clusters are known."],"forward_implications":["With 10-30 seeds per cluster the method produces competitive or superior accuracy on standard image and text benchmarks under a low-tuning protocol.","Runtime remains linear in both the number of observations and the feature dimension once the number of seeded clusters and iterations is held fixed.","Fringe points, isolated noise, and points belonging to unseeded clusters are left unassigned by construction.","The same pipeline applies without modification to raw data, linearly reduced embeddings, and neighbourhood-preserving embeddings."],"fun_headline_variants":["Seed-guided a-contrario clustering by anomaly exclusion","Iterative exclusion expands seeded clusters without tuning","A-contrario method isolates noise while forming seeded clusters","Clustering defined as maximal non-anomalous seed subsets","Clustering as dual of a-contrario anomaly detection"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"The null hypothesis of uniform randomness is an appropriate model for defining anomalies within the data distribution being clustered.","fun_headline_variants_meta":{"raw":{"variants":["Seed-guided a-contrario clustering by anomaly exclusion","Iterative exclusion expands seeded clusters without tuning","A-contrario method isolates noise while forming seeded clusters","Clustering defined as maximal non-anomalous seed subsets","Clustering as dual of a-contrario anomaly detection"]},"model":"grok-4.3","cost_usd":0.007435,"raw_usage":{"total_tokens":3429,"prompt_tokens":695,"num_sources_used":0,"completion_tokens":62,"cost_in_usd_ticks":74349500,"prompt_tokens_details":{"text_tokens":695,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":2672,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":695,"tokens_out":62,"duration_ms":19912,"temperature":1.0,"reasoning_tokens":2672,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-26T21:38:23.748284+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"On a synthetic dataset whose outliers are drawn from a clearly non-uniform distribution while the seeded clusters remain compact, measure whether the E < 1 rule still excludes the correct points or begins to assign them incorrectly.","supporting_citations":[],"review_version":1}