{"id":"17a24534-4a86-46de-beb3-7e08f37dc569","arxiv_id":"2608.09580","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"CoRCi reconstructs a mixed-domain user sequence from separate domain sequences via cross-attention and a domain-agnostic FocalNCE loss, outperforming prior cross-domain sequential recommenders.","lead":"CoRCi is a new recommendation model that combines user behavior from two shopping domains into one coherent sequence to better predict what the user buys next in either domain. The paper reports consistent accuracy gains over prior cross-domain recommendation methods across four real-world datasets, with code released.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"FocalNCE's claimed domain-bias mechanism is not evidenced; the loss in Eq. 14 is a scalar focal weighting that never reads domain labels, and its measured gains are small and depend on a per-dataset tuned alpha.","rationale":"The reader's conditional verdict rests on the same weakest assumption I identify, so no verdict change is needed. I considered the evaluation-protocol concern (999 sampled negatives and paired t-tests over five seeds) as an alternative; it is real but secondary, because the raw margins in Table 3 are large and the code is available for reproduction. The FocalNCE mechanism is more load-bearing because it is the paper's stated reason for why the mixed-domain loss preserves coherent domain-invariant interests. The loss in Eq. 14 is a scalar focal weighting of the positive's probability; it does not read domain labels, and the paper never shows that same-domain negatives are the high-probability negatives. The ablations in Table 4 and Figure 5 show the contribution is modest and that alpha is tuned per dataset, so the mechanism claim is not yet evidenced. A direct measurement of per-negative probabilities by domain would settle whether the concern lands.","tokens_in":21375,"tokens_out":16583,"duration_ms":162675,"concrete_test":"Instrument the released code on AFK (alpha=4): for each mixed-domain query in a held-out training batch, record the softmax probability p_i of every sampled negative together with whether that negative is same-domain or cross-domain relative to the query, then aggregate mean and top-decile p_i by domain. If same-domain negatives do not show systematically higher p_i than cross-domain negatives, the mechanism stated in Section 4.6 is not what FocalNCE is doing, and the claimed domain-bias suppression is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.6 (Equation 14) defines FocalNCE as -(1-p)^alpha log p, where p is the softmax probability of the positive item. This is a scalar function of p only; the loss never uses the domain of any negative. Its gradient with respect to a negative logit is proportional to that negative's softmax probability times a common positive factor, so high-probability negatives are upweighted regardless of which domain they come from. The paper's claim that same-domain negatives are 'naturally difficult' is therefore an empirical hypothesis, not a property of the loss. No measurement is reported showing that same-domain negatives actually receive higher p_i or larger gradient norms than cross-domain negatives. The ablations in Table 4 show the FocalNCE-versus-InfoNCE gap is small and statistically insignificant on the hard domains of ABE and AMB (e.g., Book MRR 0.1571 vs 0.1566; Electronics 0.1331 vs 0.1321), and the focusing parameter is tuned per dataset (alpha=4 for AFK, alpha=2 for ABE and AMB). Consequently, the contribution that FocalNCE 'suppresses intra-domain bias and strengthens domain-invariant alignment' is not established: the gains could be a tuned loss-shaping effect rather than the proposed mechanism. This is load-bearing for the paper's claimed novelty, although the raw ranking improvements over baselines in Table 3 may still reproduce.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CoRCi, a dual-target cross-domain sequential recommendation (CDSR) model. Specific-domain sequences are encoded by two separate self-attention encoders; a cross-reconstruction module (CRM) produces mixed-domain representations by applying cross-attention over the sum of the encoded specific-domain representations, using mixed-domain positional embeddings as queries. Cross-attention transferors (CAT) then fuse mixed-domain knowledge back into the specific-domain representations. The mixed-domain objective is a focal-weighted InfoNCE loss called FocalNCE, and the total loss combines domain-specific InfoNCE losses with a weighted mixed-domain loss. Experiments on three Amazon domain pairs and one Douban three-domain dataset compare CoRCi against single-target, dual-target, and mixed-target baselines, reporting consistent improvements in HR, NDCG, and MRR with claimed statistical significance at p < 0.01.","tokens_in":21660,"tokens_out":4370,"duration_ms":40426,"significance":"If the results are robust, CoRCi would be a strong new state of the art for dual-target CDSR. The paper is commendable for shipping code, using five seeds, reporting paired t-tests, and running extensive ablations. The CRM architecture and the stop-gradient design are reasonably well supported by the ablation table. However, the two contributions are not equally validated: the claimed mechanism of FocalNCE is not directly evidenced, and its benefit is entangled with per-dataset tuning of the focusing parameter alpha. The three-domain experiment is also narrower than the two-domain comparison. These gaps do not invalidate the raw ranking improvements, but they do mean the paper's central mechanistic claims need additional support before the contribution is fully established.","major_comments":[{"comment":"FocalNCE is defined as -(1-p)^alpha log p, where p is the softmax probability of the positive item. This is a scalar function of p only; the loss never reads the domain labels of the negatives. The assertion that FocalNCE 'assigns higher penalties to negatives drawn from the same domain as the query' is therefore not a property of the loss. Its gradient with respect to a negative logit is proportional to that negative's softmax probability times a common positive factor, so any upweighting of high-probability negatives is domain-agnostic. To support the stated mechanism, the authors should report direct evidence, such as average softmax probabilities or gradient norms for same-domain versus cross-domain negatives, or ablate by manipulating the domain composition of the negative set. As written, FocalNCE is a tuned loss-shaping term whose claimed intra-domain-bias suppression is unverified.","section":"Section 4.6, Eq. (14)"},{"comment":"The focusing parameter alpha is selected per dataset (alpha = 4 for AFK, alpha = 2 for ABE and AMB), and the RQ3 ablation shows that replacing FocalNCE with plain InfoNCE yields differences on the hard domains of ABE and AMB that the authors themselves describe as statistically insignificant fluctuations (e.g., Book MRR 0.1571 for Vii vs 0.1566 for CoRCi; Electronics 0.1331 for Vii vs 0.1331 for CoRCi). This undercuts the load-bearing claim that FocalNCE's mechanism drives the gains; the improvement could be an artifact of per-dataset loss-weight tuning. The authors should report significance tests for the RQ3 variants and demonstrate that alpha values in a neighborhood of the optimum preserve the advantage over InfoNCE.","section":"Section 5.6, Figure 5, Table 4"},{"comment":"The three-domain evaluation on DMBM compares CoRCi only against ABXI, with no additional DT-CDSR baseline, and the preprocessing differs from the two-domain setup (minimum item frequency 5 instead of 10, maximum sequence length 100 instead of 50, day-granularity timestamps with a fixed tie-breaking order). The generalizability claim for three domains is therefore weaker than the two-domain comparison. At least one additional DT-CDSR baseline, plus a description of how ABXI was configured under this protocol, is needed to support the RQ5 conclusion.","section":"Section 5.8, Table 5"}],"minor_comments":[{"comment":"The symbol h^F_{m,j} in Eq. (15) is not defined in the text; the preceding paragraph introduces h_m as the mixed-domain query representation. Please define h^F consistently.","section":"Section 4.6, Eq. (15)"},{"comment":"The significance claims are reported only as 'p <= 0.01'. Please provide exact p-values or at least the test statistic and clarify whether the paired t-test is over the five seeds or over a per-user/per-item evaluation sample.","section":"Section 5.3, Table 3"},{"comment":"The variant name 'Vw/o-LM: cancels mixed-domain task on Vw/o-sg' is confusing; it should be clarified whether this is Vw/o-sg with the mixed-domain loss removed or a separate variant, and how it differs from Vw/o-Lm used elsewhere in the table.","section":"Section 5.5, Table 4"},{"comment":"The FLOPs comparison states that CoRCi uses '5 SA plus 5 FFN' but the preceding sentence says 'each SA and FFN processes only one sequence'; the accounting for the cross-attention modules and projection layers should be spelled out so the 1.06x figure is reproducible.","section":"Section 4.7"},{"comment":"The preliminary motivating experiment with SASRec and BERT4Rec variants is reported only as a figure. A small table with the underlying MRR values and standard deviations would make the motivation more checkable.","section":"Section 1, Figure 2"}],"recommendation":"major_revision","confidential_remarks":"The main risk is that FocalNCE, which is one of the two headline contributions, may be a tuned loss-weighting effect rather than the domain-bias mechanism described in the paper. The architectural results appear solid, and the raw improvements are likely reproducible, but the mechanistic claim needs direct evidence. I would like to see the authors add negative-level analysis (e.g., same-domain vs cross-domain negative probabilities or gradient norms) and alpha-robustness checks before acceptance; if these cannot be provided, the FocalNCE contribution should be reframed as an empirical loss heuristic."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"CoRCi is a solid, well-run empirical paper in dual-target cross-domain sequential recommendation. The headline is that it consistently beats ABXI and other baselines on four datasets with five-seed runs and paired t-tests; the architecture is the main contribution, and the FocalNCE loss is a smaller, somewhat overclaimed addition.\n\nWhat's actually new: the cross-reconstruction module (Eq. 10) builds mixed-domain representations by cross-attending to the sum of pre-encoded domain-specific sequences, rather than training a separate mixed-sequence encoder. That, plus the stop-gradient transferors, is a real design change and the ablations show it carries most of the gain. The paper is honest about experimental detail: five seeds, per-dataset alpha/beta tuning, sensitivity curves, a three-domain generalization test. Code is promised. For a subfield that mostly stacks encoders and aggregates per-domain losses, this is a useful, reproducible step.\n\nThe soft spot is FocalNCE. Equation 14 is a scalar focal weighting of the positive's softmax probability; it never reads domain labels. The claim that it 'penalizes same-domain negatives more' is an empirical hypothesis about those negatives having higher softmax scores, not a property of the loss. The paper presents no per-domain negative-score or gradient analysis to support it. The ablation gap over plain InfoNCE is small and statistically insignificant on the hard domains of ABE and AMB (and slightly negative on Book). Alpha is tuned per dataset. So the mechanism story is not established; the loss is a reasonable heuristic that may help a little, but the paper should either provide evidence for the domain-bias mechanism or soften the claim.\n\nMinor: the 999-sampled-negative evaluation is standard in this area, so the 'statistically significant' phrasing should be read as significant under that protocol, not full-ranking significance. The three-domain experiment compares only against ABXI, which is fine as a sanity check but doesn't tell us how CoRCi stacks against other multi-domain methods.\n\nOverall: this is a competent systems paper with a plausible main contribution and an overclaimed secondary one. It deserves a serious referee. I'd send it out, with instructions to ask for either gradient-level evidence for FocalNCE's domain-bias story or a rewrite that presents FocalNCE as a tuned loss-shaping heuristic. The architecture and empirical work justify the review effort.","headline":"Cross-reconstruction is a genuine architectural step forward; FocalNCE is a small, overclaimed heuristic that needs either gradient-level evidence or softer claims.","tokens_in":22228,"tokens_out":3884,"would_cite":true,"duration_ms":34406,"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":"CoRCi: mixed-domain sequences rebuilt by cross-attention from per-domain encoders and trained with a single focal-contrastive loss preserve shared interests and beat prior cross-domain recommenders on four real-world datasets.","keywords":["cross-domain sequential recommendation","sequential recommendation","cross reconstruction","FocalNCE","contrastive learning","cross-attention","domain-invariant interests"],"falsifier":"Compute the mean cosine similarity between query representations and same-domain negatives versus cross-domain negatives in the mixed-domain space, and run CoRCi with FocalNCE's per-negative weights replaced by random domain labels; if same-domain negatives are not systematically closer, or if randomly reweighted negatives reproduce the gains, the intra-domain-bias mechanism is refuted.","tokens_in":21150,"feed_emoji":"🛒","tokens_out":14571,"duration_ms":120174,"temperature":0.7,"pith_summary":"CoRCi tackles the problem of keeping a user's shared interests coherent when training a recommender on two domains at once. The paper argues that prior methods harm this coherence by training a separate encoder on the chronologically merged cross-domain sequence and by aggregating losses separately per domain. Its proposal is to build the mixed-domain representation directly from the two per-domain encoders using cross-attention ($H_M = \\mathrm{ATTN}(E_M, H_A+H_B, H_A+H_B)$) and to train that representation with a single, domain-agnostic contrastive loss. To stop same-domain items from dominating that loss, it adds FocalNCE, which multiplies the standard InfoNCE term by $(1-p)^\\alpha$ so negatives from the query's own domain are penalized harder. If the paper is right, this combination is a straightforward recipe that improves next-item recommendation on both domains and scales beyond two domains.","feed_headline":"CoRCi beats cross-domain baselines by rebuilding mixed sequences","feed_subtitle":"Cross-attention reconstruction plus a focal loss keeps shared interests coherent and lifts hits on Amazon and Douban.","key_machinery":"The central object is the Cross-Reconstruction module (CRM): a cross-attention layer that turns the two per-domain self-attention encodings $H_A$ and $H_B$ into the mixed-domain representation $H_M$ by using $H_A+H_B$ as keys and values and the mixed-domain embedding $E_M$ as queries. The companion mechanism is FocalNCE, a modification of InfoNCE in which the per-sample loss $-(1-p)^\\alpha \\log p$ reweights the positive softmax probability $p$ so that negatives sampled from the same domain as the query receive larger penalties and intra-domain bias is suppressed. A stop-gradient on $H_M$ in the transfer cross-attention modules keeps the mixed-domain space from being distorted by per-domain supervision. These mechanisms together are what the paper claims preserve coherent, domain-invariant interests.","core_discovery":"The central claim is that domain-invariant interest coherence is what makes dual-target cross-domain sequential recommendation work, and that coherence is best achieved by cross-reconstruction rather than by learning a mixed-domain encoder from scratch. CoRCi encodes each domain separately with a self-attention encoder, then produces the mixed-domain representation with one cross-attention layer whose queries are the mixed-domain positional embeddings and whose keys and values are the sum of the two specific-domain encodings. That reconstructed representation is trained with a single sequence-level FocalNCE loss, so no per-domain loss aggregation fragments it; a stop-gradient cross-attention transfer then carries the mixed knowledge back into each domain. The paper reports that this design outperforms prior cross-domain sequential recommenders, including ABXI, across all metrics with paired t-test significance at $p \\le 0.01$ on Amazon Food-Kitchen, Beauty-Electronics, Movie-Book, and the three-domain Douban data, with larger gains on hard domains.","pith_inferences":["Beyond the paper: the FocalNCE reweighting scheme could be applied to any contrastive learning setup where negatives come from identifiable subgroups, not just recommendation.","Beyond the paper: an adaptive version of the focusing parameter, estimated from validation similarity, could remove the per-dataset grid search.","Beyond the paper: if the mechanism is real, the advantage of FocalNCE over plain InfoNCE should shrink on datasets where same-domain negatives are not measurably closer than cross-domain ones.","Beyond the paper: the cross-reconstruction pattern could transfer to other multi-source sequence tasks, though CoRCi only demonstrates it for two- and three-domain recommendation."],"forward_implications":["Prior per-domain loss aggregation in mixed-domain CDSR is the main thing CoRCi argues against; the correct recipe is a single sequence-level, domain-agnostic loss on the reconstructed representation.","The architecture generalizes to more than two domains: on the three-domain Douban data, CoRCi beats the strongest baseline in all nine domain-metric combinations.","The focusing parameter $\\alpha$ must be tuned per dataset; the paper finds $\\alpha=4$ works for Food-Kitchen, $\\alpha=2$ for Beauty-Electronics and Movie-Book, and large values can hurt datasets whose interests are mostly domain-specific.","The stop-gradient on the mixed representation keeps the shared space stable; removing it degrades performance, as does removing the mixed-domain task entirely.","The reported gains over the best baseline are statistically significant at $p \\le 0.01$ on every metric, so the observed improvements are unlikely to be random-seed noise."],"supporting_citations":[{"why":"Defines the self-attentive sequence encoder used as CoRCi's per-domain SAE and as a baseline.","marker":"[25]"},{"why":"ABXI, the strongest prior DT-CDSR baseline and the source of the experimental setup CoRCi follows.","marker":"[5]"},{"why":"C2DSR, a prior mixed-domain CDSR baseline representing the contrastive per-domain-loss approach CoRCi argues against.","marker":"[8]"},{"why":"DREAM, a prior mixed-domain CDSR baseline with separate encoders and per-domain loss aggregation.","marker":"[59]"},{"why":"Supplies the focal-loss weighting that FocalNCE embeds into the InfoNCE objective.","marker":"[32]"},{"why":"Defines InfoNCE, the base contrastive objective used for specific-domain losses and modified by FocalNCE.","marker":"[38]"},{"why":"Source of the Amazon review data used for the AFK, ABE, and AMB datasets.","marker":"[37]"},{"why":"BERT4Rec, a baseline used in the motivating comparison of single-, dual-, and mixed-target training.","marker":"[45]"},{"why":"Source of the Douban dataset used for the three-domain non-overlapping-users experiment.","marker":"[71]"}],"fun_headline_variants":["CoRCi: cross-reconstruction for coherent cross-domain interests","CoRCi uses cross-attention and focal loss to keep interests coherent","CoRCi cross-reconstructs mixed sequences for better CDSR","One focal loss, cross-reconstruction: CoRCi beats CDSR baselines","CoRCi's cross-reconstruction and focal loss yield coherent CDSR"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that negatives from the query's own domain are harder negatives, so penalizing them extra is what drives the improvement; if that premise is false, the FocalNCE gains could be an artifact of tuning the focal strength $\\alpha$ per dataset.","fun_headline_variants_meta":{"raw":{"variants":["CoRCi: cross-reconstruction for coherent cross-domain interests","CoRCi uses cross-attention and focal loss to keep interests coherent","CoRCi cross-reconstructs mixed sequences for better CDSR","One focal loss, cross-reconstruction: CoRCi beats CDSR baselines","CoRCi's cross-reconstruction and focal loss yield coherent CDSR"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001042,"raw_usage":{"total_tokens":4417,"prompt_tokens":1017,"completion_tokens":3400,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":633,"completion_tokens_details":{"reasoning_tokens":3305}},"tokens_in":633,"tokens_out":3400,"duration_ms":22109,"temperature":1.0,"reasoning_tokens":3305,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T14:40:15.706357+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the mean cosine similarity between query representations and same-domain negatives versus cross-domain negatives in the mixed-domain space, and run CoRCi with FocalNCE's per-negative weights replaced by random domain labels; if same-domain negatives are not systematically closer, or if randomly reweighted negatives reproduce the gains, the intra-domain-bias mechanism is refuted.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"ABXI, the strongest prior DT-CDSR baseline and the source of the experimental setup CoRCi follows."}],"review_version":1}