{"id":"305551a4-4333-4d5c-90bf-48cc03cee6aa","arxiv_id":"2607.19234","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"MASHT, which feeds MultiRocket+Hydra features into the pretrained TabPFN-3 model, matches HIVE-COTE 2.0 on univariate time-series classification and is competitive on multivariate benchmarks.","lead":"This paper tests whether a pretrained 'tabular' AI model can classify time series after they are converted into feature tables by two standard convolutional transforms. The resulting pipeline, MASHT, matches the best existing time-series classifier on 112 univariate datasets while skipping task-specific training.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Feature-budget mismatch confounds the MASHT-vs-MR-Hydra comparison, so the TabPFN-3 attribution is unproven; the 'matches HC2' pipeline claim is unaffected.","rationale":"The reader's weakest assumption is exactly the protocol comparability of the MR-Hydra baseline: MASHT uses a dataset-dependent feature budget while the published baseline likely uses full default features. This is the most load-bearing concern because the paper's motivating question — whether TabPFN-3 improves over a linear classifier on the same random-convolutional features — cannot be answered from Table 2 alone. The pipeline-level claim of matching HC2 is independent of this confound and is supported by the reported numbers, honest statistical reporting (non-significant HC2 difference), and the multivariate results that do not overclaim. The paper explicitly acknowledges limitations and an uncontrolled runtime comparison, so the weakness is an omission of a control rather than a misrepresentation. A same-feature ridge baseline is a straightforward, decisive experiment that the authors could run. Because the central 'matches SOTA' claim is likely to survive a controlled retest, the verdict should remain CONDITIONAL rather than move to ACCEPT or REJECT. The reader already identified this issue; my stress-test does not change the verdict.","tokens_in":9565,"tokens_out":8406,"duration_ms":89209,"concrete_test":"Run the default MR-Hydra classifier (Ridge) on the exact MASHT feature tables — same dataset-dependent budgets, same 30 resamples, same preprocessing — for all 112 UTF-112 datasets. Compare its mean accuracy and per-dataset ranks to MASHT (0.892) and to the published MR-Hydra (0.884). If the same-feature ridge baseline reaches 0.890 or higher, the 0.008 gain attributed to TabPFN-3 collapses; if it stays near 0.884, the classifier attribution is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central research question is whether a pretrained tabular foundation model (TabPFN-3) more effectively harnesses MultiRocket–Hydra features than a linear classifier. The evidence for this is the comparison between MASHT and MR-Hydra in Table 2: a 0.008 mean accuracy gain (p=0.03 after Holm) with the same transform family. However, the comparison is confounded by a feature-budget mismatch. Section 4 specifies a dataset-dependent total feature budget of 10,000/2,000/200 features, split equally between Hydra and MultiRocket (i.e., 5,000/1,000/100 each). The MR-Hydra baseline is taken from Middlehurst et al. (2024), which uses the default full feature sets of both transforms. For medium and large datasets, the budgeted dimensionality is far below defaults (e.g., 200 total features for >100k instances). Thus the MASHT-vs-MR-Hydra delta changes both the classifier (Ridge -> TabPFN-3) and the representation (full -> budgeted). If reducing the feature count itself improves generalization (or if the budget selection interacts with dataset size), the 0.008 gain could be partly or wholly an artifact of the representation change, not the classifier. The paper's conclusion in Section 5 — 'TabPFN-3 is a promising classifier for the MultiRocket–Hydra features' — therefore lacks a controlled comparison. This does not invalidate the pipeline-level claim that MASHT matches HC2 (0.892 vs 0.891, rank 3.27 vs 3.41, not significantly different), which is a separate, well-supported empirical statement.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MASHT, a two-stage time-series classification pipeline that concatenates MultiRocket and Hydra random-convolutional features and feeds the resulting table to the pretrained TabPFN-3 model for in-context classification. The central empirical claim is that this training-free pipeline matches state-of-the-art TSC ensembles on univariate benchmarks: on UTF-112 it reports mean accuracy 0.892 (rank 3.27) vs. HIVE-COTE 2.0's 0.891 (rank 3.41), and it is significantly better than MR-Hydra (0.884, p=0.03 after Holm correction) while being statistically tied with HC2. On the 71-dataset Multiverse benchmark, MASHT is competitive but not leading (0.795 accuracy, rank 7.65 vs HC2 0.805/6.51). The paper also reports runtime and limitation analyses, and makes code available.","tokens_in":9810,"tokens_out":7150,"duration_ms":89152,"significance":"If the comparison to MR-Hydra were fully controlled, this would be an important result: it would show that a frozen in-context tabular model can replace a trained linear head on random-convolutional TSC features, with accuracy competitive with a much larger ensemble. The paper's statistical practice is careful: Wilcoxon signed-rank tests with Holm correction, critical-difference diagrams, and per-dataset scatterplots, and the limitations are honestly described. However, the attribution to TabPFN-3 is currently underdetermined because the key comparison changes both the classifier and the feature budget, and the provenance of some baseline metrics is not documented.","major_comments":[{"comment":"MASHT uses a dataset-dependent feature budget (10,000/2,000/200 features, split equally between Hydra and MultiRocket), but the MR-Hydra baseline is imported from Middlehurst et al. (2024) and the manuscript never states that the baseline uses the same feature budget, the same realized feature dimensionality, or the same feature-extraction configuration. The text in §4 says 'obtain all MASHT results under the same dataset protocols', but this statement refers to datasets/resamples and does not control the feature budget. Since the comparison changes both the classifier (Ridge → TabPFN-3) and the representation (full/default → budgeted), the significant 0.008 accuracy gain and the associated Wilcoxon p=0.03 cannot be unambiguously attributed to TabPFN-3. This is load-bearing for the paper's research question and for the §5 statement that 'TabPFN-3 is a promising classifier for the MultiRo","section":"§4 Experiment Setup and Table 2"},{"comment":"Table 1 reports balanced accuracy and AUROC for all baselines, but the manuscript only says baseline results are 'from Middlehurst, Schäfer, and Bagnall (2024)'. It does not state whether those AUROC/balanced-accuracy values were taken directly from that publication or recomputed, and if recomputed, from which probability outputs and by which protocol. The AUROC gap between MASHT and MR-Hydra (0.970 vs. 0.913) is very large despite the same transform family, which raises the possibility of a protocol mismatch in how AUROC is averaged or which class is treated as positive. Please clarify the source and exact computation, and if the values were recomputed, make the procedure explicit for every baseline.","section":"§4 Table 1"}],"minor_comments":[{"comment":"Typo: 'provide a provide broad temporal features' should be 'provide broad temporal features'.","section":"§1 Introduction"},{"comment":"The formula for the number of Hydra groups, g = floor(B/(2kd)), uses d = (log2(L-1))/(9-1). The derivation is terse; please define all variables and explain why the 9-1 term appears (Hydra's kernel length). Also, the realized dimensionality may be 'slightly below' the target; for reproducibility, a table of achieved feature counts across dataset sizes would help.","section":"§3 Feature Extraction"},{"comment":"The runtime table shows mean total runtime lower than the median for both benchmarks (e.g., 43.954 s mean vs. 59.743 s median on UTF-112). Given positive runtimes, this is unusual and suggests a reporting inconsistency. Please verify the numbers or explain the distribution.","section":"§4 Table 5"},{"comment":"The footnote containing the data link has a line break and the word 'Direktlink' inserted mid-sentence. Please clean up the formatting.","section":"§4 Evaluation"},{"comment":"Please specify the exact TabPFN-3 version/commit, the seed handling, and whether 'automatic estimator scaling' and 'automatic inference precision' produce deterministic results across runs. Also, since the feature-budget thresholds are fixed but ad hoc, a short sensitivity analysis (e.g., how accuracy varies with the budget) would strengthen the paper.","section":"Appendix A"},{"comment":"The runtime comparison is appropriately caveated as uncontrolled. However, the conclusion section says 'These results show that tabular foundation models are a promising classifier for random convolutional time-series features' — this should be tempered or explicitly conditioned on the feature-budget caveat.","section":"§5 Runtime"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is honest and well executed on the surface, but the central attribution of the improvement over MR-Hydra to TabPFN-3 is confounded by the feature-budget difference. The good news is that the fix is straightforward: run a same-budget ridge baseline on the exact same transformed tables, or an unbudgeted TabPFN-3 variant, and report both. The AUROC provenance issue should also be resolved. If those are addressed, the paper would be a solid contribution. No concerns about citation practice or novelty disclosure."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The headline result holds up: MASHT, a no-training pipeline that feeds MultiRocket+Hydra features into frozen TabPFN-3, statistically ties HIVE-COTE 2.0 on UTF-112 (0.892 vs 0.891; rank 3.27 vs 3.41; Wilcoxon–Holm not significant). That is a genuinely interesting operating point. The paper is also refreshingly honest: it openly reports the HC2 tie, the multivariate non-results, and the uncontrolled runtime comparison.\n\nWhat is new: pairing random convolutional TSC features with a pretrained tabular foundation model, evaluated on 183 datasets across two established benchmarks. The statistical methodology is appropriate — Holm-corrected Wilcoxon, critical-difference diagrams, same resamples for MASHT. The multivariate evaluation is also fairly reported: competitive, but not leading, and explicitly not a new state of the art.\n\nSoft spot, and it is the main one: the claim that TabPFN-3 improves over ridge on MR-Hydra features is confounded by a feature-budget mismatch. MASHT uses a dataset-dependent budget of 10,000/2,000/200 features split equally between Hydra and MultiRocket. The MR-Hydra baseline from Middlehurst et al. uses the full default feature sets. So the 0.008 mean accuracy gain (p=0.03) could come partly from the reduced representation rather than from TabPFN-3. The paper needs a control: run ridge on the same budgeted features. This is a missing baseline, not a fatal flaw. The pipeline-level claim (MASHT matches HC2) is unaffected.\n\nMinor issues: the abstract's 'lower average rank' overstates a statistically null accuracy difference; Figure 2 labels use the obsolete acronym 'ICTS'; the runtime comparison is correctly labeled uncontrolled; and I could not verify the GitHub repo. None of these change the main empirical finding.\n\nWho this is for: anyone working on efficient TSC or on applying tabular foundation models outside their original domain. The paper deserves a serious referee; the central finding is solid and the confound is fixable with one additional experiment. I would send it to review.","headline":"A clean, honest empirical study showing a frozen tabular foundation model on random convolutional features can match HIVE-COTE 2.0 on univariate TSC; the causal claim about TabPFN-3 is less clean.","tokens_in":10442,"tokens_out":1189,"would_cite":true,"duration_ms":14747,"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":"This paper claims that a pretrained tabular foundation model can replace the linear classifier in random-convolutional time series classification, matching the strongest published ensemble on univariate benchmarks without any task-specific","keywords":["time series classification","random convolutional features","tabular foundation model","in-context learning","MultiRocket","Hydra","no-training classification","benchmark evaluation"],"falsifier":"Run MR-Hydra under exactly MASHT's feature-budget protocol (equal Hydra/MultiRocket split, 10,000/2,000/200 features by dataset size) on UTF-112 and check whether its 0.008 accuracy gap to MASHT persists; if it vanishes, the gain attributed to the tabular foundation model is actually a feature-budget effect.","tokens_in":9302,"feed_emoji":"📈","tokens_out":3246,"duration_ms":31130,"temperature":0.7,"pith_summary":"This paper asks whether a pretrained tabular foundation model can replace the simple linear classifiers normally paired with random convolutional time-series features. It proposes MASHT, which extracts MultiRocket and Hydra features and feeds them directly to TabPFN-3, requiring no task-specific training. On 112 univariate datasets MASHT matches the accuracy of HIVE-COTE 2.0, the strongest published ensemble, and significantly beats MR-Hydra, the same feature family with a ridge classifier. On 71 multivariate datasets it is competitive with leading methods but does not lead. The result matters because it suggests a modular, training-free route to accurate time series classification.","feed_headline":"No-training time series classifier ties top ensemble","feed_subtitle":"Random convolutional features plus a pretrained tabular model match HIVE-COTE 2.0 on 112 univariate benchmarks.","key_machinery":"The pipeline's load-bearing components are two random convolutional transforms — MultiRocket, which pools kernel responses from raw and differenced series, and Hydra, which records competing-kernel extreme-response counts — concatenated into a fixed feature table, and TabPFN-3, a pretrained transformer that performs in-context classification on that table without fitting task-specific weights. The feature table turns temporal structure into a standard tabular problem; the foundation model supplies the learned decision rule.","core_discovery":"The central claim is that the pairing of random convolutional feature transforms with an in-context tabular foundation model is itself a competitive time series classifier. Concretely, on the UTF-112 benchmark MASHT reaches a mean accuracy of 0.892 and a mean rank of 3.27, slightly above HIVE-COTE 2.0 (0.891, rank 3.41) and significantly better than MR-Hydra (0.884, rank 3.88) after Holm-corrected Wilcoxon testing. The authors present this as evidence that the tabular model captures nonlinear structure in the transformed features that a linear classifier misses, while the univariate/multivariate contrast suggests the representation is most effective for single-channel series.","pith_inferences":["If the feature-budget confound is resolved and the gain holds, the same recipe could be applied to other feature transforms and other pretrained tabular models, making 'transform + foundation model' a general template for time series classification.","The multivariate shortfall points to a testable extension: adding cross-channel convolution or a second-stage feature interaction layer before the tabular model might close the gap to the leading ensemble on the Multiverse benchmark.","The dataset-dependent feature budget (200 features for large data) implies accuracy may degrade at scale; investigating where that boundary bites would clarify the method's practical limits.","A controlled speed comparison on identical hardware and datasets would be the natural next evaluation; the paper leaves this open."],"forward_implications":["On univariate benchmarks, MASHT matches the accuracy of HIVE-COTE 2.0 while requiring no task-specific training, only feature extraction and pretrained-model inference.","Because the classifier is pretrained, new datasets can be classified immediately at inference time once features are computed, which changes the deployment cost profile for time series classification.","The significant gain over MR-Hydra indicates that the choice of downstream classifier matters for random convolutional features, not just the transform.","On multivariate benchmarks, the pipeline remains competitive but not leading, so the tabular representation does not yet capture cross-channel structure as well as ensemble methods.","The reported runtimes suggest a favorable accuracy–runtime trade-off, though the paper notes the comparison is not controlled."],"fun_headline_variants":["Random features plus tabular model match top ensemble","No-training pipeline ties state-of-the-art time series classifier","Zero-training approach rivals HIVE-COTE 2.0","Pretrained tabular model lifts random features to top rank","Simple recipe: random convolution + foundation model for TS"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The central premise is that the published baseline results were produced under evaluation protocols and feature budgets comparable to MASHT's, so the measured accuracy differences reflect the classifier rather than the amount of features.","fun_headline_variants_meta":{"raw":{"variants":["Random features plus tabular model match top ensemble","No-training pipeline ties state-of-the-art time series classifier","Zero-training approach rivals HIVE-COTE 2.0","Pretrained tabular model lifts random features to top rank","Simple recipe: random convolution + foundation model for TS"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0003,"raw_usage":{"total_tokens":1545,"prompt_tokens":698,"completion_tokens":847,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":442,"completion_tokens_details":{"reasoning_tokens":768}},"tokens_in":442,"tokens_out":847,"duration_ms":9521,"temperature":1.0,"reasoning_tokens":768,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T13:01:00.577793+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run MR-Hydra under exactly MASHT's feature-budget protocol (equal Hydra/MultiRocket split, 10,000/2,000/200 features by dataset size) on UTF-112 and check whether its 0.008 accuracy gap to MASHT persists; if it vanishes, the gain attributed to the tabular foundation model is actually a feature-budget effect.","supporting_citations":[],"review_version":1}