acceptodds
Under review as a conference paper at ICLR 2027

FlySCoRe: Fly-Inspired Sparse Cohort Retrieval for Next-Basket Recommendation

Abstract

Next-basket recommendation ranks the products that a household is likely to purchase on its next shopping occasion. The task combines recurring patterns in the household's own history with purchase evidence shared across households. We introduce Fly-inspired Sparse Cohort Retrieval (FlySCoRe), a retrieval-based ranking model inspired by the sparse associative organization of the Drosophila mushroom body. For each history–candidate pair, a history-conditioned user signal summarizes purchasing behavior across time scales, while a sparse address selects memory locations containing weighted purchase and opportunity evidence from earlier observations. A support-aware readout combines the retrieved evidence with the user signal to rank the full product catalog. We instantiate this design in three ways. A central nervous system model (CNS) couples learned task interfaces to fixed measured connectivity; a probabilistic model (PROB) expresses memory selection through choice probabilities; and a neural model (NET) uses learned activation strengths. Across four repeat-purchase datasets, the highest-scoring implementation exceeds the strongest of eleven baselines in NDCG@20 by 4.81%-33.60%. We further construct an alternative memory interface from retained input activity and composed connections. Rebuilding historical memory at this interface preserves 97.0%-99.3% of complete CNS NDCG@20. These results support sparse cohort retrieval across measured, probabilistic, and neural implementations.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
next-basket recommendation, recommender systems, brain-inspired computing, hybrid AI systems, sparse random projections, associative memory

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_1Mb6OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes FlySCoRe, a fly-inspired sparse retrieval framework for next-basket recommendation. The key idea is to perform candidate-dependent sparse addressing over historical observations, retrieve relevant purchase evidence from memory, and combine it with a user-level signal for candidate scoring. The authors develop three variants based on connectome-constrained, probabilistic, and neural addressing, and evaluate them on four repeat-purchase datasets.

Critical strengths
  1. Interesting candidate-dependent retrieval formulation. The idea of retrieving different historical evidence for different candidate products is intuitive and potentially useful for next-basket recommendation.

  2. Strong empirical performance. The proposed variants achieve substantial improvements over the considered baselines across four datasets.

  3. Multiple implementations and analyses. The paper investigates the proposed retrieval mechanism through connectome-constrained, probabilistic, and neural variants, together with several ablation and reconstruction experiments.

Critical weaknesses
  1. Efficiency of candidate-dependent addressing. The candidate-dependent addressing mechanism may introduce non-trivial inference overhead, since the model needs to perform the addressing process for individual candidate products when ranking the catalog. The paper would benefit from reporting the time complexity and empirical inference latency, preferably compared with the baseline methods.
  2. The necessity of sparse selection is not sufficiently justified. The paper does not clearly establish the advantage of sparse addressing over dense addressing. In fact, the retrained dense-addressing results reported in the Appendix achieve better NDCG@20 on three of the four datasets. This raises the question of whether sparse selection provides a meaningful benefit beyond the biological motivation.
  3. Insufficient sensitivity analysis for the sparsity level. The effect of the sparse rate, or equivalently the number of activated memory units, is not sufficiently studied. Since this is a central design choice of the proposed method, a sensitivity analysis would help clarify the robustness of the method and the trade-off between sparsity and recommendation performance.
  4. No analysis of memory-unit utilization and load balance. The paper does not discuss whether memory units are utilized uniformly. Since the learned addressing repeatedly selects a limited subset of memory units, some units may receive substantially more support than others. This could potentially reduce the effective memory pool size and lead to repeated selection of the same memory units. An analysis of memory-unit utilization would help clarify whether such an imbalance occurs and whether it affects the reported results.
Questions for authors
  1. Efficiency: What is the computational complexity and empirical inference latency of the candidate-dependent addressing mechanism, especially when ranking the full catalog? How does it compare with the baseline methods?
  2. Sparsity: Have the authors systematically studied how the sparse rate (i.e., the number of activated memory units) affects recommendation performance? How sensitive are the results to this choice?
  3. Memory-unit utilization: Have the authors analyzed the distribution of memory-unit utilization/support? In particular, is there a potential rich-get-richer effect where a small subset of memory units is repeatedly selected, making the effective memory pool substantially smaller than the nominal pool size? If such an imbalance exists, how does it affect the learned memory representation and recommendation performance?
3 comments

Thank you for your review. Below, we clarify the computational scope, the exact empirical comparison between sparse and dense addressing, the fixed operating point, and memory utilization.

W1 / Q1: We construct memory offline and do not rescan histories during inference.

FlySCoRe pre-aggregates eligible historical support into S/NS/N outcome memory via Eqs. (2)--(3). Once memory is populated, candidates are encoded, sparsely addressed, and scored directly from stored memory statistics without rescanning all historical observations for every product in the catalog. Offline memory construction from training histories is thus separated from inference-time ranking. Our primary empirical claim focuses on quantifying the ranking-accuracy contribution of candidate-dependent evidence under a unified full-catalog protocol, whereas wall-clock latency benchmarks depend on hardware, batching, and low-level implementation choices and represent a separate systems evaluation.

W2: Sparse CNS outperforms retrained dense addressing on all four datasets.

We would like to clarify the exact numbers from Tables 2 and 5: sparse CNS actually outperforms retrained dense addressing in NDCG@20 on all four datasets, rather than trailing on three: CJ: 0.4841 (Sparse CNS) vs. 0.4731 (Retrained dense) TF: 0.4326 (Sparse CNS) vs. 0.4152 (Retrained dense) DC: 0.8435 (Sparse CNS) vs. 0.8412 (Retrained dense) IC: 0.6357 (Sparse CNS) vs. 0.6315 (Retrained dense)

As specified in Appendix E.5, this control "removes sparse selection and refits the encoder and signal strength under the original training protocol and matched initialization." These matched results directly validate the empirical advantage of sparse, selective addressing over a refitted dense alternative across every evaluated dataset, while the K bypass complements this comparison by measuring the joint contribution of sparse addressing and the personal signal.

W3 / Q2: We use the same fixed k=7k=7 across all four datasets.

In our experiments, we deliberately fix the number of active units to k=7k=7 across all four datasets without per-dataset tuning (Appendix B.3) and compare this single operating point against its matched retrained dense counterpart. The four-dataset evaluation therefore tests a unified sparse regime rather than dataset-specific capacity tuning. Because the same sparse addressing rule governs both memory writing and reading, kk defines the coupled memory interface rather than an inference-time threshold; changing it requires rebuilding and refitting both sides of that interface.

W4 / Q3: Address selection does not depend on accumulated memory support.

As defined in Eqs. (1), (4), and (5), memory routing is computed solely from current history-candidate features and learned parameters prior to reading accumulated S/NS/N counts. Address selection is therefore independent of how much support a memory unit already holds, while non-uniform occupancy across units naturally reflects the uneven prevalence of underlying shopping situations in the data. From a modeling standpoint, the relevant property is candidate discrimination: as shown in Section 3.3.1, if all candidates shared an identical address, they would retrieve identical totals and leave the personal-signal ranking unchanged. Both the matched M bypass and the sparse-versus-dense comparison directly verify that candidate-dependent memory access improves ranking performance.

We hope these clarifications address your concerns, and we would be happy to discuss any remaining points.

Thank you for clarifying several points that were unclear to me in the original paper. While I appreciate these explanations, some concerns including the lack of sparsity sensitivity analysis and efficiency evaluation remain insufficiently addressed. I will maintain my original score.

AuthorsOpenReview ↗

Additional Results on Sparsity and Full-Catalog Efficiency

Thank you for your follow-up. Standard benchmarks in the cited next-basket recommendation literature compare models using ranking accuracy (Recall@K and NDCG@K) under a common protocol, which is why our main experiments focus on these metrics. To address your questions directly, we also ran the requested analyses of sparsity sensitivity, memory coverage, and inference latency under the same full-catalog test setup.

1. Sensitivity to the sparsity level kk

We varied the number of activated memory units using k=3,5,7,9,11k=3, 5, 7, 9, 11 and evaluated full-catalog ranking performance:

DatasetMetrick=3k=3k=5k=5k=7k=7 (selected)k=9k=9k=11k=11
CJNDCG@200.47610.48520.48410.48970.4902
CJRecall@200.86200.86200.86200.86200.8620
TFNDCG@200.42130.41240.43260.43280.4323
TFRecall@200.83090.81620.86030.84560.8456
DCNDCG@200.84430.84250.84350.83860.8390
DCRecall@201.00001.00001.00001.00001.0000
ICNDCG@200.63380.63580.63570.63500.6353
ICRecall@200.87590.87800.87620.87720.8776

Performance stays stable across all tested values of kk. The selected setting (k=7k=7) is within 0.0009 of the best tested NDCG@20 on TF, DC, and IC, and within 0.0061 on CJ, showing that it works reliably across datasets without per-dataset tuning.

2. Effective memory coverage at k=7k=7

Across the constructed outcome memory, 1,620 of the 4,064 memory units receive nonzero support on CJ, 1,062 on TF, 1,028 on DC, and 1,824 on IC, corresponding to 25.3%-44.9% of the full memory pool. Activating k=7k=7 units for each pair therefore controls per-candidate sparsity, while different history-candidate pairs populate a broad set of memory units rather than a fixed seven-unit subset.

3. Full-catalog inference latency

We measured full-catalog inference latency per query on the same machine using the selected checkpoints, reporting the median over three test-set runs:

MethodAverage latency (ms/query)
FlySCoRe-PROB7.40
FlySCoRe-CNS17.31
Triple2Vec9.65
Sets2Sets10.44
DREAM17.27
TIDE72.08
DNNTSP72.50
SFCNTSP72.79

With connectome-constrained addressing, FlySCoRe-CNS stays under 45 ms/query on every dataset and averages 17.31 ms/query, more than four times faster than the temporal-set models TIDE, DNNTSP, and SFCNTSP. Its probabilistic variant, FlySCoRe-PROB, further reduces average latency by 57.3% to 7.40 ms/query, making it faster than all listed learned history and sequence baselines.

We hope these additional results address your remaining concerns and support reconsideration of your score.

Reviewer_4VKXOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper works on next-basket recommendation problem with a model based on the fruit fly's mushroom body, which is a collaborative filtering like method. The paper works on old datasets with old AI solutions, which appears to be far behind what modern recommender systems is built based on deep learning. Its value is unclear to the broader scope of the community.

Critical strengths

None

Critical weaknesses

The paper works on old datasets with old AI solutions. Its value may be demonstrated in more SOTA recommender systems with more updated datasets.

Questions for authors

How can this solution be integrated into modern recommender systems which are done at a much larger scale with deep learning models?

3 comments

We appreciate the opportunity to clarify the methods, benchmarks, and neural integration evaluated in our work.

W1: We evaluate FlySCoRe against modern neural baselines on four standard datasets.

Our work directly incorporates contemporary deep learning architectures both within the proposed framework and across the evaluated baselines. Specifically, FlySCoRe-NET learns its neural expansion and sparse address mapping via Eq. (5), and the joint encoder in Eq. (8) models nonlinear history-product interactions via bilinear pooling (Section 3.3.2; Appendix B.3). Across identical queries and complete product catalogs, we evaluate against eleven baselines spanning graph, recurrent, and neural temporal-set models---including DNNTSP, SFCNTSP, and TIDE (SIGIR 2026) (Section 4.2)---following their cited papers and released implementations, adapted only to match the unified data interface and full-catalog evaluation protocol.

Furthermore, these four datasets remain standard benchmarks for controlled repeat-purchase modeling; for example, the recent SIGIR 2026 CASE study likewise evaluates on Ta-Feng, Dunnhumby-Carbo, and Instacart (Cao et al., Section 3.1). More fundamentally, as documented in Section 4.1, "observed history-product pairs have a median purchase count of one on every dataset, and 56.2%--84.8% appear only once." Because product-level evidence within an individual household is inherently sparse regardless of dataset age, different candidate items require distinct supporting evidence from the same household history.

FlySCoRe addresses this core modeling problem by making evidence selection explicit: candidate-dependent addresses select relevant historical outcomes, while storing separate purchase and opportunity counts distinguishes a raw purchase rate from its empirical support strength (Section 3.2; Eqs. (2)--(4)). Within this framework, CNS examines fixed measured connectivity, PROB formalizes address overlap as a co-selection probability, and NET learns neural addresses end-to-end. Each branch outperforms the strongest NDCG@20 baseline on three datasets (Table 2), and Section 4.5 further shows that reconstructing an alternative read/write interface from retained input activity and composed connections preserves 97.0%--99.3% of CNS NDCG@20. Together, these results demonstrate that the shared computational design is effective across both measured biological connectivity and learned neural abstractions.

Q1: FlySCoRe can be integrated with deep recommenders through a modular memory interface.

FlySCoRe can be integrated with modern deep recommenders: a backbone neural model supplies learned history and candidate representations to the address encoder, while its personal-history score anchors the support-aware readout. During training or indexing, historical prefixes write purchase and opportunity statistics through support addresses; at inference, candidate items retrieve relevant evidence through query addresses. NET provides a neural realization of this interface (Eqs. (5) and (8)). Because support observations are pre-aggregated into S/NS/N statistics (Eqs. (2)--(3)), inference avoids rescanning raw historical records per candidate. This modular interface enables contemporary deep recommenders to retain their learned representations while adding historical evidence together with its support.

We hope these clarifications address your concerns.

Reviewer_4VKXOpenReview ↗

Thanks

My score remains as no new info added but rephrasing what it had. We thank authors follow up.

AuthorsOpenReview ↗

Clarification of Evidence Already in the Submission

We note the reviewer's response. Our clarification cited the submitted manuscript because the review's sole criticism---that the work relies on "old AI solutions" on older datasets without modern deep learning---is contradicted by methods and experiments already present in the paper.

Specifically, the submission already includes: (1) a learned neural implementation (FlySCoRe-NET, Eqs. (5) and (8)) and a modular memory interface for deep neural backbones (Section 3.3.2); (2) full-catalog comparisons under an identical protocol against eleven baselines covering graph, recurrent, and neural temporal-set models, including SFCNTSP and TIDE (Section 4.2); and (3) evaluation on four repeat-purchase benchmarks, three of which are also used in the recent CASE study.

Because these neural models, baseline comparisons, and benchmarks are already evaluated in the manuscript, the premise of the critique does not apply to the submitted work.

Reviewer_8yhJOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes a next-basket recommendation model that combines a customer's own purchasing history with evidence retrieved from other historical shopping situations. Each history-candidate pair selects sparse memory locations containing purchase and opportunity counts. A support-aware readout combines this evidence with a learned personal-history signal. The authors implement the design using measured fly connectivity (CNS), probabilistic addressing (PROB), and learned neural addressing (NET), and evaluate the implementations on four grocery datasets.

I find the candidate-dependent retrieval idea relevant to recommendation: the same customer history may require different supporting evidence when predicting milk purchases versus detergent purchases. My main question is how much of the reported improvement comes from the proposed memory mechanism beyond the strong personal-history predictor already present in the system.

Critical strengths
  1. A relevant modeling problem and an interpretable design. Pooling evidence at the history-candidate level is a reasonable way to address sparse product-level observations for established customers. The distinction between a purchase rate and the amount of evidence supporting that rate is useful. The memory equations make the contribution of historical observations explicit, and aggregated counts give the retrieval mechanism a concrete implementation. (Sections 2-3.)

  2. Encouraging results under a shared full-catalog protocol. All methods rank the same complete catalog for the same evaluation queries. The strongest FlySCoRe branch exceeds the best listed baseline in NDCG@20 on each dataset, with the reported relative improvements ranging from 4.81% to 33.60%. The best branch varies by dataset, which the paper acknowledges. The comparison includes frequency, neighborhood, latent-factor, and sequential methods. (Sections 4.2-4.4 and Table 2.)

  3. Useful supplementary analysis. The paper includes history-based slices, recommendation-distribution analysis, household-bootstrap intervals, and retrained addressing/readout controls. These are useful additions beyond the main accuracy table. The learned alternatives to measured connectivity also help frame the contribution as a computational retrieval design. (Appendices D and E.)

Critical weaknesses
  1. The incremental contribution of outcome retrieval is not established clearly enough. The readout-bypass model ranks directly by the unchanged personal-history signal and omits outcome-memory writing and retrieval (Appendix E.4). It achieves NDCG@20 of 0.4689 versus 0.4841 for full CNS on Complete Journey, and 0.6306 versus 0.6357 on Instacart (Table 2). These results show that much of the advantage over the listed baselines is already present without outcome retrieval. The remaining improvement may still be useful, but it should be the central quantity analyzed. The large drop in the addressing ablation cannot isolate sparsity because that intervention also removes the personal-history signal, as the authors acknowledge. The reconstruction result needs the same context: on Instacart, reconstructed memory scores 0.6307, almost identical to the signal-only score of 0.6306 (Table 5). Retaining nearly all CNS performance therefore provides limited evidence for the reconstructed memory's incremental value. Matched signal-only and simpler retrieval comparisons are needed for PROB and NET as well; the CNS result alone does not establish what happens in those separately fitted branches.

  2. The evaluation population and split construction need a reproducible specification. Table 1 contains only 32 test queries for Complete Journey and 68 for Ta-Feng. A focus on established repeat-purchase customers is legitimate, but Appendix C.1 does not state the exact eligibility rule, thresholds, filtering retention rates, or split-construction procedure. Without these details, it is difficult to understand why these queries remain and how broadly the conclusions apply within the stated recurring-demand setting. The supplementary history slices and bootstrap intervals are useful, but they do not resolve how the population was selected. The temporal protocol should also clarify support eligibility across households, particularly for Instacart, whose timestamps are household-relative (Appendix B.3). This is a request for a complete protocol, not evidence that leakage occurred.

  3. The strong learned initialization is insufficiently specified. Appendix B.3 states that the final adaptation inherits feature encoders and the user-signal router from earlier complete-input models, uses Triple2Vec representations, and follows an earlier ranking stage. I could not reconstruct the full training procedure for those inherited components from the description. This matters because the personal-history signal is already strong and validation selects the initial checkpoint in 51 of 108 adaptation runs (Appendix E.8). The three-seed study measures additional adaptation from a common learned initialization, rather than variability from training the complete system independently. A complete specification of the starting models, their training data and objectives, and the information and tuning budgets available to the baselines is necessary to interpret the comparison.

Questions for authors
  1. What exact history-based eligibility and splitting rules produce Table 1? Please give retained household/query counts and explain temporal support eligibility across households.

  2. How much does outcome retrieval improve over the same personal-history signal in each of CNS, PROB, and NET? Can the principal deltas be compared with a simple matched retrieval or readout alternative, with uncertainty estimates?

  3. As a secondary clarification, what are the full-catalog scoring and memory-rebuild costs relative to the signal-only model? This would help interpret the accuracy gained from the additional mechanism.

1 comment

Thank you for your constructive review and detailed questions regarding our experimental controls, data splits, initialization protocol, and computational scope. We address each point below.

W1 / Q2: We use a matched signal-only control to isolate the gain from outcome memory.

Table 2 includes a matched signal-only control for CNS (w/o Readout). As specified in Appendix E.4, this ablation "bypasses M, preserves A and K, and ranks products directly by the user signal supplied by K," omitting S/NS/N writing, retrieval, and fusion. Full CNS improves NDCG@20 over this control by +0.0152, +0.0053, +0.0197, and +0.0051 on CJ, TF, DC, and IC, respectively. Furthermore, Table 5 evaluates validation-calibrated constant fusion and a compact trained residual head. Together, these controls isolate the contribution of explicit outcome memory from simpler readout alternatives, whereas the K bypass and interface reconstruction address the distinct questions of joint module removal and transfer to a different memory interface. Accordingly, CNS provides the matched estimate of the incremental outcome-memory gain, while PROB and NET test the same retrieval design under independently fitted probabilistic and neural implementations.

W2 / Q1: We use the same cleaned queries and chronological splits for all methods.

Before query construction, routine data cleaning removes baskets with insufficient coverage in the retained product records, using the same 34% threshold across all splits. As stated in Section 2 and Section 4.1, query eligibility is then determined solely from recurring-purchase evidence in the household history observed prior to the prediction cutoff. All methods are evaluated on identical queries and complete product catalogs. Data splits are strictly chronological: earlier eligible cutoffs form the training set, the penultimate retained basket serves as validation, and the final retained basket serves as test. Table 1 reports the retained query counts for each split. The corresponding unique household counts for train / validation / test are 138 / 25 / 32 for CJ, 72 / 23 / 68 for TF, 144 / 156 / 153 for DC, and 5,121 / 1,314 / 1,151 for IC. Validation and test each contain one retained query per household, whereas a household may contribute multiple training queries.

As detailed in Appendix B.3, memory construction uses up to 3,000 household-balanced training prefixes augmented with scheduled training queries. Historical support requires the support outcome day to precede the query day---using calendar days on CJ, TF, and DC, and cumulative within-household elapsed days across households on IC---thereby excluding same-day, future, and test outcomes.

W3: We follow the cited implementations and select checkpoints on validation data.

Model initialization follows the cited methods and their released implementations. Specifically, original training histories supply the complete-input feature representations, training baskets supply Triple2Vec relational embeddings, and LambdaRank provides ranking initialization. Model checkpoints are selected exclusively via validation performance, while test labels remain strictly held out for final evaluation. During adaptation, PROB and NET update the joint feature mixer and address map, whereas CNS updates the history and product projections into ALPN activity; inherited feature encoders and user-signal components remain fixed. All baselines follow their original papers and implementations on identical splits, queries, and full catalogs, using validation NDCG@20 (with Recall@20 as tie-breaker) for hyperparameter and checkpoint selection (Appendix C.4). Appendix E.8 further reports multi-seed stability starting from these learned initializations.

Q3: We construct memory offline and do not rescan histories during inference.

As formulated in Eqs. (2)--(3), FlySCoRe pre-aggregates eligible historical support into purchase (SS) and opportunity (NN) statistics. Consequently, full-catalog inference encodes, addresses, and scores candidates directly from stored memory statistics rather than rescanning all historical observations for every candidate item, while offline memory construction is decoupled from ranking. Our empirical evaluation is designed to isolate the accuracy contribution of candidate-dependent evidence under a standardized full-catalog protocol; wall-clock runtime comparisons depend heavily on hardware, batching strategies, and engineering optimizations, making them a separate systems-level evaluation beyond the scope of this study.

Thank you again for your careful reading. Please let us know if any further clarification would be helpful.

Reviewer_zZoCOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes FlySCoRe, which transfers the sparse associative organization of the Drosophila mushroom body to candidate-specific retrieval in next-basket recommendation, using purchase and opportunity counts to retrieve historical evidence with a signal-anchored readout. The authors implement this design in three branches and evaluate CNS, PROB, and NET on four datasets.

Critical strengths
  1. The paper transfers the fly’s sparse associative organization to next-basket recommendation through adapted input and readout interfaces, making the biological inspiration computationally consequential and constituting a genuinely novel idea.
  2. The evaluation covers four datasets and includes module ablations, reconstruction of memory access at another interface, and a spike-count control, providing thorough empirical support.
Critical weaknesses
  1. The ablation confounds sparse selection with the personal-history signal, and the readout bypass control remains close to full CNS, so the retrieval mechanism’s contribution is not isolated.
  2. The reconstruction retains 97.0% to 99.3% of CNS NDCG@20 but performs comparably to the readout bypass, so the result does not establish a benefit from reconstruction beyond that control.
Questions for authors
  1. The K bypass removes both sparse selection and the personal-history signal. Why is the performance degradation specifically attributed to sparse retrieval rather than the personal history signals?
  2. Since reconstruction performs comparably to the readout bypass control, what does retaining 97.0% to 99.3% of the CNS NDCG@20 actually show?
  3. Given that PROB and NET do not rely on measured connectivity yet remain competitive, what independent contribution does the connectome constraint provide beyond the shared retrieval design?
  4. Which implementation should be taken as the basis for the paper’s main claim when no single branch leads across all four datasets?
2 comments

Thank you for your thoughtful feedback and for recognizing both the cross-domain computational transfer and the breadth of our evaluation. Below, we clarify what each of our matched interventions, interface reconstructions, and complete-model comparisons establishes.

W1 / Q1: We use matched controls to isolate the contribution of outcome memory.

The K bypass removes sparse memory selection and the personal-history signal simultaneously. As noted in Section 4.4, "this intervention therefore measures their joint contribution." The matched control that isolates the retrieval mechanism is the M bypass (w/o Readout). As detailed in Appendix E.4, this variant "bypasses M, preserves A and K, and ranks products directly by the user signal supplied by K," removing S/NS/N writing, retrieval, and fusion without retraining. Full CNS improves NDCG@20 over this signal-only control by +0.0152, +0.0053, +0.0197, and +0.0051 on Complete Journey (CJ), Ta-Feng (TF), Dunnhumby-Carbo (DC), and Instacart (IC), respectively (Table 2). These increments directly quantify the contribution of outcome memory beyond the identical personal signal. In addition, Table 5 compares this support-aware readout against validation-selected constant fusion and a trained 98-parameter residual head.

W2 / Q2: The reconstructed interface retains 97.0%--99.3% of CNS performance.

Section 4.5 characterizes retrieval reconstruction as "an alternative memory path from retained model activity." In this experiment, memory writing and retrieval are shifted from KC coordinates to retained ALPN input activity, and the S/NS/N memory statistics are rebuilt from historical support while keeping the learned input projections and measured connectome matrices fixed. The reported 97.0%--99.3% figures measure performance retention relative to complete CNS under this change of memory interface, whereas the matched M bypass separately isolates the incremental gain of the outcome-memory path itself.

Q3: We use measured connectivity for both addressing and memory access.

Measured connectivity governs both sparse address formation and connection-weighted memory writing and retrieval (Eq. (1); Appendix B.2, Eqs. (6)--(7)). As described in Section 3.1, these anatomical weights remain fixed while learned interfaces map shopping features into the circuit. CNS thus evaluates candidate-dependent retrieval under a fixed structural prior and provides the connection path used in the reconstruction experiments. PROB and NET also show that the same selective retrieval organization remains effective under probabilistic and learned neural realizations.

Q4: Each branch outperforms the strongest baseline on three datasets.

Table 2 shows that CNS leads on CJ, PROB leads on TF, and NET leads on DC and IC, while Section 4.4 further notes that each complete branch outperforms the strongest NDCG@20 baseline on three of the four datasets. Together, these branch-level results show that the underlying retrieval organization is independently effective across measured, probabilistic, and neural implementations. Within this unified framework, CNS anchors our controlled module interventions and structural reconstruction experiments, and reporting all three branches shows their complementary strengths across datasets.

We hope these clarifications address your questions, and we would be happy to discuss any remaining points.

Thanks, i would like to keep my score.

Comments on the paper

1 comment

Dear reviewers,

The rebuttal period is ending soon. Please read the authors rebuttal and check if they addressed your concerns. Please acknowledge the rebuttal and adjust your scores accordingly.

Thanks.

-AC

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.