{"id":"d69464b3-c84d-4044-b48e-d3762b670581","arxiv_id":"2412.07573","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Clustering-based subtopic views plus per-epoch random view training improves document matching on CNSE, CNSS, LeCaRDv2, and COLIEE 2023.","lead":"This paper introduces a training framework, SST, that clusters sentences of a document pair into subtopics, samples several document views from those subtopics, and trains the matcher on a different sampled view each epoch. The method reports consistent gains on news duplication and legal case retrieval over strong baselines including fine-tuned LLMs, with possible use in legal search and news deduplication.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Temporal ordering is not isolated from multi-view sampling; without a random-per-batch control, the central temporal-aggregation claim is unproven.","rationale":"The reader identified the under-specified attention baseline as the weakest assumption, and I agree that this is a real flaw. However, I see a more direct problem for the central claim: the temporal schedule itself is never isolated from multi-view exposure. The attention-baseline issue is a symptom of the same under-testing—the paper wants to show that spatial aggregation fails while temporal aggregation succeeds, but it lacks controls that would establish either proposition. The framework's overall empirical results are plausible and the improvements over BERT/SAILER/Match-Ignition are substantial, so I would not reject the paper. The missing random-per-batch control and the under-specified attention baseline warrant a conditional verdict rather than full acceptance. Since the reader already reached CONDITIONAL, my read does not change the verdict; it sharpens the reason for it. The proposed test is deliberately narrow: it will settle whether the temporal ordering is doing any work beyond multi-view data augmentation. If the random-per-batch version matches, the title claim should be substantially weakened even though the empirical framework may still be useful.","tokens_in":1869,"tokens_out":1402,"duration_ms":72514,"concrete_test":"Run BERT+SST-as on CNSE and SAILER+SST-as on COLIEE with the same view pool, but replace the epoch-fixed view schedule of Eq. (10) with a view sampled uniformly at random from the pool for every mini-batch, keeping all other hyperparameters, inference aggregation, and seeds fixed. If random-per-batch training matches the epoch-scheduled accuracy/NDCG, the temporal ordering is not load-bearing and the method reduces to multi-view sampling; if it does not match, the temporal mechanism is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's core novelty is the temporal aggregation strategy: Eq. (10) in Sec. 3.4 trains with view t at epoch t from a pre-sampled view pool. The views are drawn i.i.d. from the sampling distribution, so there is no dependency or designed ordering between successive views that the model could exploit. The ablation 'w/o temporal aggregation' (Table 4) removes all multi-view exposure rather than only the temporal schedule, and the 'uniform/random sample' rows change the sampling distribution, not the schedule. Thus the reported gains cannot separate 'seeing many views over training' (standard multi-view augmentation) from 'gradually integrating views in a temporal order.' This matters because the paper claims temporal aggregation is preferable to spatial attention aggregation; without a control that holds multi-view exposure fixed and varies only the timing, the distinctive mechanism is not established. The attention baseline is also under-specified—no architecture, tuning, or variance is reported, and its collapse to 80.77 Acc on CNSE is asserted as evidence that spatial aggregation is unsuitable. That evidence is at best unverified, but the more fundamental gap is the missing random-per-batch multi-view control.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SST, a learning framework for long-form document matching. SST first discovers document-pair subtopics by clustering sentences (with a direct spectral-clustering variant and an adaptive, learned variant), then constructs multiple document views via uniform, hard, or soft sampling from the resulting subtopic clusters, and finally trains a matching model with a 'temporal aggregation' strategy: each training epoch uses a different view from a pre-sampled view pool, and at inference the model aggregates scores over a small view pool. Experiments on two news-duplication datasets (CNSE, CNSS) and two legal case retrieval datasets (LeCaRDv2, COLIEE 2023) show consistent improvements over several strong baselines, with the best variant (BERT+SST_as) reaching 87.65 Acc on CNSE and (SAILER+SST_as) 53.19 NDCG@10 on LeCaRDv2. The main positive result is that the subtopic-aware multi-view sampling strategy improves over content-selection baselines such as Match-Ignition and OPT-Match. The distinctive claim, however, is that 'temporal aggregation' is preferable to 'spatial aggregation' (attention); this claim is not supported by the current ablations.","tokens_in":18170,"tokens_out":5369,"duration_ms":55066,"significance":"If the reported gains are robust, the view-sampling component is a useful contribution: it is model-agnostic, consistently improves over content-selection baselines in two domains, and the paper ships an anonymous code release, reports hyperparameter analyses, evaluates against LLM baselines, and includes statistical significance tests for the top results. These strengths make the empirical contribution credible as a multi-view training framework. The significance of the specifically 'temporal' mechanism, however, is not established: the paper's central narrative that temporal aggregation is preferable to spatial attention aggregation rests on ablations that do not isolate the temporal schedule, and on an under-specified attention baseline. The paper's value therefore currently lies in subtopic-aware view sampling plus inference-time aggregation, not in a demonstrated temporal mechanism.","major_comments":[{"comment":"The experimental design does not isolate the 'temporal' component from the multi-view exposure. In Eq. (10), the model is trained on view v^t at epoch t, but the views are drawn from a pre-sampled pool without any ordering or dependency between successive views; the ablation 'w/o temporal aggregation' in Table 4 removes all multi-view training, so its drop measures the value of seeing many views over training, not the value of a temporal schedule. To support the central claim that temporal aggregation is preferable, the authors should add a control that holds the number of views per training run fixed and varies only the schedule, e.g., randomly resampling a view each batch or each epoch without a fixed sequence, and compare against the fixed per-epoch schedule. Without such a control, the reported gains are equally explained by standard multi-view data augmentation.","section":"§3.4, Eq. (10); Table 4"},{"comment":"The attention aggregation baseline is under-specified. The paper reports only that attention is used 'as [43]' and the row collapses to 80.77 Acc on CNSE, worse than plain BERT, but provides no architecture description, training setup, hyperparameter tuning, or variance over runs. Because the title and abstract claim that temporal aggregation is preferable to spatial aggregation, this comparison is load-bearing; if the attention baseline were implemented or tuned more carefully, the collapse might not occur. The authors should report the full configuration, tuning procedure, and multiple random seeds, or temper the claim.","section":"§5.2, Table 4 (row 'w attention')"},{"comment":"The adaptive clustering component is not unsupervised as stated. L_p in Eq. (4) uses the JS divergence between the cluster distributions of positive document pairs, and L_n in Eq. (5) is applied to negative pairs, so the clustering head is trained with the same document-pair relevance labels used for the matching model. This is a form of supervised view construction that should be acknowledged and discussed; without an analysis of how much of the gain comes from label-informed clustering, the comparison to content selection baselines that do not use labels is not apples-to-apples.","section":"§3.2, Eqs. (4)-(6)"}],"minor_comments":[{"comment":"The notation is unclear: 'Max' is described as max-pooling, but the expression 'Max(B_i)' suggests a scalar per sentence, and the displayed sums over P_cq and P_cd do not make the cluster dimension explicit; the description of the JS-divergence loss would be clearer with dimensions spelled out.","section":"§3.2, Eq. (4)"},{"comment":"The grouping of ablation rows is confusing: the header 'w/o aggregation' is followed by 'w/o L_n' and 'w/o L_p', which appear to refer to removing the adaptive clustering losses, not to removing aggregation; renaming the groups (e.g., 'w/o adaptive clustering losses') would prevent misreading.","section":"Table 4"},{"comment":"The training view pool size is not specified: the paper says 'receptive field of models is 40 sentences' and later 'view pool size as 3' for inference, but the number of views used during training (the pool V(q,d) in Eq. (10)) is never explicitly given, making the training-time exposure unclear.","section":"Appendix B"},{"comment":"The y-axis labels are duplicated on both sides ('F-1 on CNSS' appears on the left and right), and the sentence 'we sort the samples in descending order according to their similarity... Generally, the more similar a view to the whole document, the more primary information it contains' would be clearer if the direction of the sort were tied to the x-axis explicitly.","section":"Figure 7"},{"comment":"The phrase 'temporal aggregation' is introduced as a mechanism that 'effectively integrates different views gradually as the training progresses,' but the method in §3.4 only uses a different pre-sampled view per epoch without any gradual or adaptive ordering; the wording should be revised to match the actual mechanism or the mechanism should be extended.","section":"Abstract and §1"}],"recommendation":"major_revision","confidential_remarks":"The core issue is that the paper's headline novelty, temporal aggregation, is not actually isolated from standard multi-view training: the views are i.i.d. samples from a pool, and the only 'temporal' element is the epoch index. I would require the random-per-batch control experiment before accepting the central claim. Also, the adaptive clustering is supervised by the same labels as the matching model, which is not disclosed clearly and should be handled carefully in revision. The view-sampling contribution itself appears solid and worth publishing if the temporal claims are toned down or the missing controls are added."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know: this paper reports steady improvements on news duplication and legal case retrieval with a new view-sampling scheme, and the gains look credible. But the headline mechanism—temporal aggregation—is not actually isolated in the experiments, and the paper overstates it.\n\nWhat is new: clustering sentences from both documents into subtopics (spectral clustering or a GAT-based adaptive variant), then sampling multiple views from those clusters under three distributions (uniform, hard, soft). That combination is not in Match-Ignition, OPT-Match, or CIG, which select a single salient summary. The adaptive clustering with the L_p and L_n losses is also a reasonable idea, though it uses document-pair labels, so calling it 'unsupervised' is misleading.\n\nWhat the paper does well: four datasets, strong baselines including fine-tuned 7B LLMs, and consistent gains (BERT+SST-as 87.65 Acc vs 83.97 BERT on CNSE; SAILER+SST-as 53.19 NDCG@10 vs 50.87 on LeCaRDv2). The analysis of receptive field and cluster size is useful, and the code is available under an anonymous link.\n\nThe soft spots: the temporal-aggregation claim is not supported. Views are drawn i.i.d. from a pre-sampled pool; there is no designed ordering for the model to exploit. The ablation 'w/o temporal aggregation' removes all multi-view exposure, not just the schedule, so it cannot separate 'seeing many views' from 'seeing them in a temporal order.' The 'w attention' baseline is under-specified—no architecture or tuning details—and its collapse below BERT suggests poor hyper-parameter choice rather than a fundamental limitation of spatial aggregation. No error bars are reported, though they do run pairwise t-tests.\n\nAlso, the adaptive clustering loss uses positive/negative pair labels to shape the cluster assignments. That is a two-stage supervised pipeline, not an unsupervised one. It may be fine, but it should be labeled honestly.\n\nIf the authors add a random-per-batch control—same number of views, but sampled randomly each batch—they could actually test the temporal claim. My guess is that the order does not matter and the gains come from multi-view exposure. That would still be a useful result, but it changes the story from 'temporal aggregation' to 'multi-view augmentation with subtopic-aware sampling.'\n\nThis deserves peer review, not desk rejection. Send it to a serious referee and ask them to push on the random-schedule control and the attention baseline. The core sampling idea is worth engaging with.","headline":"Consistent gains from subtopic-aware view sampling, but the 'temporal aggregation' framing needs a control that isolates schedule from multi-view exposure.","tokens_in":18688,"tokens_out":2663,"would_cite":true,"duration_ms":25427,"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":"The SST framework claims that sampling multiple subtopic views and training on a fresh view each epoch improves long-form document matching beyond single-view content selection.","keywords":["long-form document matching","subtopic analysis","view sampling","temporal aggregation","news duplication","legal case retrieval","document clustering","content selection"],"falsifier":"Re-run the ablation on CNSE and COLIEE 2023 with a tuned attention aggregator over the same view pool—reporting its architecture, heads, and hyperparameters—and compare against temporal aggregation. If tuned attention reaches or exceeds 87.65 accuracy on CNSE or 38.57 NDCG@10 on COLIEE, the paper's central claim that temporal aggregation is needed would be falsified.","tokens_in":17767,"feed_emoji":"📄","tokens_out":5949,"duration_ms":53378,"temperature":0.7,"pith_summary":"Long documents contain multiple subtopics, and the paper argues that existing content-selection methods, which extract only the most similar aligned sentences, discard complementary signals and can mislead the matching model. To fix this, SST first clusters the sentences of the query–candidate pair into subtopics, then samples several partial-document views that mix aligned and complementary subtopics, and finally trains the matcher by feeding a different view at each training epoch—a 'temporal aggregation' that gradually ensembles the views instead of attending to all of them at once. The paper shows that this framework improves accuracy on news-duplication datasets CNSE and CNSS, and NDCG on legal-case-retrieval datasets LeCaRDv2 and COLIEE 2023, over strong baselines such as BERT, Match-Ignition, and SAILER. If correct, it offers a training-time-only way to get finer-grained document understanding without changing the underlying encoder.","feed_headline":"New view each epoch lifts long-document matching","feed_subtitle":"SST samples subtopic views and aggregates them over training, beating BERT and SAILER on news and legal benchmarks.","key_machinery":"The machinery is a two-part training procedure. First, the combined sentence set of the query and candidate is clustered into 'subtopics'—semantic fragments—using either spectral clustering on a TextRank-style similarity matrix or an adaptive neural clustering head trained with a Jensen–Shannon alignment loss $\\mathcal{L}_p$ and an orthogonality loss $\\mathcal{L}_n$. Second, three sampling strategies (uniform, hard, and soft) construct document views—partial documents of $k$ sampled sentences—that balance aligned and complementary subtopics. Third, temporal aggregation feeds a different sampled view at each training epoch, so the model integrates views gradually over training, and at inference combines scores over a small view pool (max for classification, mean for ranking). The temporal step is inspired by Dropout's random sub-network sampling, treating each view as a sampled sub-network to be ensembled during training.","core_discovery":"The central discovery is that representative document views for matching must combine aligned subtopics with complementary subtopics, and that such views are best integrated over training time rather than by spatial attention. Concretely, with adaptive clustering and soft sampling (SST-as), BERT reaches 87.65 accuracy on CNSE and 93.45 on CNSS, versus 83.97 and 89.63 for plain BERT and 86.32 and 91.28 for Match-Ignition. On legal retrieval, SAILER+SST-as reaches 53.19 NDCG@10 on LeCaRDv2 versus 50.87 for the two-stage SAILER baseline, and 38.57 NDCG@10 on COLIEE 2023 versus 36.51 for SAILER. The ablation attributes the gains to the temporal aggregation step: removing it drops CNSE accuracy to 85.37, and replacing it with an attention aggregator drops it to 80.77, below plain BERT.","pith_inferences":["Inference: A testable extension the paper leaves implicit is whether ordering the view pool by salience or difficulty, rather than using a fixed random order, sharpens the curriculum effect of temporal aggregation; the paper does not explore view ordering.","Inference: Because temporal aggregation is model-independent and only changes training, it should combine with long-context encoders or LLM-based rankers; the paper tests only BERT and SAILER, so this is an extrapolation.","Inference: The strong collapse of the attention baseline might reflect overfitting to noisy views rather than a general failure of spatial aggregation; a matched-capacity comparison with a properly tuned attention aggregator would clarify the mechanism."],"forward_implications":["On news duplication, BERT+SST-as improves accuracy over plain BERT by about 3.7 points on CNSE and 3.8 points on CNSS, and it also beats the previous strong content-selection baseline Match-Ignition.","On legal case retrieval, SAILER+SST-as raises NDCG@10 on LeCaRDv2 from 50.87 to 53.19 and improves COLIEE 2023 NDCG@10 from 36.51 to 38.57.","Removing temporal aggregation or replacing it with attention aggregation degrades results, with the attention variant falling below plain BERT on CNSE, which supports the paper's claim that spatial aggregation is poorly suited to heterogeneous views.","Soft sampling generally outperforms hard sampling, suggesting that views with moderate heterogeneity combined with aligned salient content are the most representative.","The framework helps on both long and short document pairs, indicating that the gain comes from view diversity and training dynamics rather than simply extending the input window."],"supporting_citations":[{"why":"Supplies the CNSE and CNSS news-duplication datasets and the CIG baseline that SST must beat.","marker":"[22]"},{"why":"Supplies the Match-Ignition content-selection baseline and the PageRank-based sentence-selection idea SST extends.","marker":"[27]"},{"why":"Supplies the OPT-Match sentence-selection baseline, another content-selection method SST is compared against.","marker":"[46]"},{"why":"Spectral clustering algorithm used in the direct subtopic-discovery variant.","marker":"[31]"},{"why":"DMoN graph-clustering model that the adaptive clustering head is built on.","marker":"[35]"},{"why":"TextRank similarity formula used to build the sentence-similarity matrix for direct clustering.","marker":"[25]"},{"why":"Dropout, whose random sub-network sampling inspires the temporal-aggregation training strategy.","marker":"[33]"},{"why":"SAILER, the legal-domain pre-trained encoder used as backbone and baseline for legal case retrieval.","marker":"[19]"},{"why":"LeCaRDv2 dataset used for legal case retrieval evaluation.","marker":"[20]"}],"fun_headline_variants":["Temporal view aggregation boosts long-document matching","Subtopic-aware views, integrated over training, beat BERT","Train document views over time, not with attention","SST samples subtopics and fuses them gradually for better match","Gradual view fusion lifts long-doc relevance"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's case that temporal aggregation is the right way to combine views rests on a single, under-described attention baseline in the ablation study; if that baseline were properly implemented and tuned, temporal aggregation might show no advantage over standard multi-view training.","fun_headline_variants_meta":{"raw":{"variants":["Temporal view aggregation boosts long-document matching","Subtopic-aware views, integrated over training, beat BERT","Train document views over time, not with attention","SST samples subtopics and fuses them gradually for better match","Gradual view fusion lifts long-doc relevance"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000559,"raw_usage":{"total_tokens":2657,"prompt_tokens":946,"completion_tokens":1711,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":562,"completion_tokens_details":{"reasoning_tokens":1632}},"tokens_in":562,"tokens_out":1711,"duration_ms":11283,"temperature":1.0,"reasoning_tokens":1632,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T18:42:30.664889+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the ablation on CNSE and COLIEE 2023 with a tuned attention aggregator over the same view pool—reporting its architecture, heads, and hyperparameters—and compare against temporal aggregation. If tuned attention reaches or exceeds 87.65 accuracy on CNSE or 38.57 NDCG@10 on COLIEE, the paper's central claim that temporal aggregation is needed would be falsified.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the CNSE and CNSS news-duplication datasets and the CIG baseline that SST must beat."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the OPT-Match sentence-selection baseline, another content-selection method SST is compared against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Spectral clustering algorithm used in the direct subtopic-discovery variant."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"TextRank similarity formula used to build the sentence-similarity matrix for direct clustering."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"LeCaRDv2 dataset used for legal case retrieval evaluation."}],"review_version":1}