acceptodds
Under review as a conference paper at ICLR 2027

The Next Recommendation Is a Direction: Discovering High-Value Semantic Opportunities Beyond User History

Abstract

Recommenders usually choose the next item, but interest exploration and generative recommendation raise an earlier question: where in the content space should the system go next? We formulate next semantic direction selection: given a short user history, find a continuous semantic location whose nearby, previously unobserved content contains a high-value realization. Our hypothesis is that population feedback contains cross-content behavioral regularities that transfer information from observed responses to semantically different content. We introduce NORTH, a population-anchored framework for continuous direction search. Shared feedback learns behavioral representations; the user's history turns population opportunity into a personalized direction score. A learned policy starts from promising population anchors, decides where to search and whether to stay or move, and proposes continuous displacements. Each proposal is grounded in a fixed-size core of real items and assessed by the same direction scorer. Best-Realization Value measures the best observed response within this fixed realization budget. On the full and fully-observed dataset variants, KuaiRec-Full and KuaiRec-FO, NORTH exceeds population-only bank selection in all four evaluated settings. It leads the compared OPEN generators on KuaiRec-Full and KuaiRec-FO. Ablations support the use of history-conditioned proposals and multimodal displacements.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
recommender systems, user interest exploration, next semantic direction selection, cross-content behavioral regularities, collaborative representation learning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 2 reviews, above 85% of the others.
Reviewer_2dHfOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper frames recommendation as the task of "next semantic direction selection", where given a user action sequence, the goal is to identify a continuous location within the content space such that the previously unobserved items in its admissible neighborhood contain high-value relaizations. The author propose NORTH, a framework that learns cross-content behavior representation, employs a history-conditioned scorer for personalization of the cluster-based oppurtunities, and utilizes a policy that navigates via "stay or move" decisions and multimodal displacements. Directions are evaluated using the Best Realized Value (BRV), where the optimal response observed within a fixed "relaization budget". Experiments on KuaiRec-Full and KuaiRec-FO datasets demonstrates that the NORHT variant outperforms baselines that rely on fixed candidate pools, and that the full NORTH framework surpasses independent open-ended generation models such as DreamRec or PreferDiff. Ablation studies validate the effecrtiveness of history-conditioned proposals against various displacement strategies.

Critical strengths
  • A principled reformulation of recommendation paradigm. This paper shifts the objective of recommendation from next item prediction to semantic direction prediction, formalizing this as the selection of the continous direction anchored to a fixed-size subset of real items. The reformulation is significant and establishes an intermediate objective that can be utilized by item retrieval or item generation paradigms.
  • Clear, principled derivation accompanied by a corresponding evaluation protocol. NORTH separates the learning of collective behavior, the scoring personalized directions and training and evaluation of history conditioned search. The BRV objective correctly targets the optimal outcome at the core level, which is critical: by using a fixed candidate pool alongside an independent open-ended generation or retrieval components, while employing a shared, frozen scorer, the approach decouples scoring accuracy from quality of the direction .
Critical weaknesses
  • A gap in presentation and reproducibility. Several key quantities in the methodology are left undefined or under specified. For example, the local semantic scale c_i in equation 14, the support region R used for grouding the direction, and the reference cited after "The representation used for short-history adaptation". The canddiate generation process where the predefined core is not fully discussed and motivated. This is significant because thse are essential components for reproducing NORTH and adapting it toward broader recommendation tasks. The core mechanism of the prosed method cannot be verified or replicated based on the main body of text.
  • The evaluation is concentrated in terms of both scenario and metrics. All evidence is derived from single source, KuaiRec, and reported exclusively using the new proposed BRV metric, with no comparisons against established metrics based on recall or ranking. This is critical since BRV is a custom, core-level maximum value metrc, which makes it difficult to determine whether the observed performance gain reflect geneuine improves in recommendation quality.
Questions for authors

How is the support region R defined and sized, and how sensitive are the result to its choice?

Could the author presents experimental evidence that the direction search outperforms item based recommendation?

2 comments

Thank you for your careful reading and for recognizing both the upstream direction-selection problem and the separation between behavioral learning, personalized scoring, and search. Your questions identified implementation details that needed a precise account. We have added the local-scale explanation immediately after Eq. (14) and collected the geometric construction in Appendix G.1.

1. The local scale and support region are now specified

In Eq. (14), cjc_j is the anchor coordinate and ℓ(cj)\ell(c_j) is its local semantic scale. All geometric operations use frozen 16-dimensional PCA content coordinates. For item ii, let dk+(i)d_k^+(i) be the distance to its kk-th nearest distinct content location, excluding coincident locations and distances within the numerical tolerance. We use

s_i = \max\left\\{\varepsilon, \min\left[d_8^+(i), 3d_3^+(i)\right]\right\\}

\ell(c) = \max\left\\{\varepsilon, \operatorname{median}_{i\in N_3(c)}s_i\right\\}

where N3(c)N_3(c) contains the three nearest catalogue items. Consequently, the same normalized displacement produces a shorter move in a dense region and a longer move in a sparse region; the 3d3+(i)3d_3^+(i) term limits the influence of a distant eighth neighbor.

The support is

ρmax⁡=Quantile⁡0.95(si)\rho_{\max} = \operatorname{Quantile}_{0.95}\left(\\{s_i\\}\right)

R(c) = \left\\{i : \lVert x_i-c \rVert_2 \leq \min\left(\ell(c), \rho_{\max}\right)\right\\}

Here ρmax⁡=0.85487997\rho_{\max}=0.85487997. Grounding returns the four nearest distinct items in this support, resolving equal distances by ascending original item ID. Proposals with insufficient support are rejected; grounded cores must also be external to the target history and satisfy the training-donor support requirement. The formulas and operation order are in Eqs. (42)--(45).

2. The bank construction and its role are explicit

Each item contributes its own coordinate, two midpoints with its nearest two other items, and the mean coordinate of itself and its nearest three other items. These centers cover both item locations and intervening local positions while keeping every realization budget at four items.

The construction produces 12,852 center proposals. Deduplication by the generating unordered item pairs and quadruples gives 11,114 candidate records; common grounding retains 10,495 legal records; deduplication by the resulting four-item identities gives 6,434 bank cores. Search uses this bank to obtain the personalized incumbent and up to eight population-supported anchors, then evaluates 72 continuous proposal attempts. This makes the bank both a reproducible starting set and an identity-based reference for testing discovery beyond it (Appendix G.1).

3. The missing reference has been repaired

Section 4.1 now cites Bertinetto et al., "Meta-learning with differentiable closed-form solvers" for differentiation through a closed-form ridge solver to learn a representation that supports rapid adaptation. The cited arXiv version dates to 2018; the work was subsequently published at ICLR 2019. This reference supports the solver mechanism used in Eq. (4), while the population-centered core objective and direction-search formulation are specified in our own method.

Thank you for pointing out these presentation gaps. The additions now connect the main-paper equations to the concrete geometry and bank-building procedure.

Thank you also for asking how the support-region choice affects the results and whether the direction benefits actual item recommendation. We added experiments for both questions.

4. Support-radius sensitivity

Appendix G.6 and Table 14 evaluate radius multipliers 0.8, 1.0, and 1.2 across five additional search seeds. Expanding the default support from 1.0 to 1.2 preserves the original bank and anchors and produces the following paired changes:

SettingBRV change, 1.2 − 1.0 [95% CI]
Full, W=20W=20+0.0011 [0, 0.0029]
Full, W=50W=50+0.0017 [−0.0013, 0.0060]
FO, W=20W=20+0.0003 [0, 0.0010]
FO, W=50W=50−0.0001 [−0.0003, 0]

The expansion adds 5,574 legal proposal occurrences but changes only 26 of 5,640 final selections (0.46%). Thus the default results are stable to this support expansion even when it creates additional geometric opportunities.

The tighter 0.8 setting retains 1,187 of the 6,434 original bank representatives and therefore also changes the available starting points. The paired changes on Full are −0.0174 and −0.0358 at W=20W=20 and W=50W=50. Reporting both tests makes the distinction explicit: expanding an already supported neighborhood leaves the decisions largely unchanged, whereas substantially restricting support removes useful starting opportunities.

5. Direct evidence with item-based recommenders

We added a single-item experiment using BPR, SASRec, and LightGCN as frozen downstream rankers. Each ranker selects one item either directly from the catalogue or from NORTH's four-item core. We then measure the observed response of that chosen item. This evaluates an item actually chosen by a ranker, not an oracle maximum over four observed responses.

Table 15 reports the paired gains on common measurable users:

SettingBPRSASRecLightGCN
Full, W=20W=20+0.0995+0.0644+0.0675
Full, W=50W=50+0.1621+0.0650+0.0373
FO, W=20W=20+0.0404+0.0775+0.0192
FO, W=50W=50+0.0444+0.0186+0.0345

All twelve paired differences are positive. This directly supports the proposed intermediate role: selecting the direction first supplies useful opportunities to several established item-ranking architectures. The experiment is summarized in Section 5.4 and detailed in Appendix G.7.

Regarding your request for established recommendation metrics, the added evaluation targets the underlying question directly through observed selected-item response, alongside BRV. Recall would require a specified relevant-item set; our watch-based feedback is graded, and the selected-item outcome retains that information.

6. Why the dataset is particularly suited to this task

KuaiRec combines near-complete target feedback with population logs and content information. Its 99.6%-dense small matrix (Gao et al., 2022) makes it possible to measure the complete four-item realization set using observed outcomes. Sparse logs generally leave some selected members unobserved, requiring an additional missing-outcome assumption or a change to the evaluated set. KuaiRec lets us preserve the actual selected core and the fixed budget.

The two settings also test different population-learning conditions: Full trains on 5,765 users outside the FO subset, whereas FO uses 903 training users from that subset. Both evaluate on held-out FO users, and the Full-trained model transfers without further population training (Section 5.1). Coverage and common-user comparisons are now reported explicitly in Appendices G.2–G.4.

We sincerely appreciate your suggestions. They have helped us make both the implementation and the connection to realized item recommendation much clearer, and we would be grateful for your reassessment of the revised evidence.

References

Gao, C., Li, S., Lei, W., Chen, J., Li, B., Jiang, P., He, X., Mao, J., and Chua, T.-S. (2022). KuaiRec: A Fully-observed Dataset and Insights for Evaluating Recommender Systems. Proceedings of the 31st ACM International Conference on Information & Knowledge Management (CIKM 2022).

Reviewer_LL8QOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes NORTH, a population-anchored framework for discovering high-value semantic directions beyond users’ interaction histories. It combines population-level behavioral learning, personalized direction scoring, and continuous search to identify promising content opportunities. Experiments on KuaiRec-Full and KuaiRec-FO demonstrate improvements over population-only selection and competitive continuous proposal methods.

Critical strengths

1.The paper extends the recommendation objective from predicting the next item to discovering high-value semantic directions beyond user history, offering a new perspective on interest exploration.

2.The framework combines population-level behavioral learning, personalized direction scoring, and continuous direction search, evaluating potential value through fixed-size candidate sets composed of real items.

Critical weaknesses
  1. The method mainly combines existing techniques for population-level behavioral learning, personalized scoring, and continuous search, without sufficiently validating the proposed search strategy against simpler alternatives.

2.BRV measures the best observed feedback within a fixed-size candidate set, but does not directly reflect the actual performance of retrieval or generation tasks, nor does a higher predicted score guarantee greater real-world user value.

3.The experiments are conducted primarily on the KuaiRec dataset, and BRV is calculated only when all items in a candidate set have observable feedback. Since historical exposure mechanisms affect feedback observability and training sample selection, the results may be subject to selection bias. The effectiveness of the method in broader real-world recommendation scenarios remains to be validated.

Questions for authors

Please refer to Weaknesses

3 comments

Thank you for your thoughtful review and for recognizing the value of selecting a semantic direction before recommending its realization. Your request for comparisons with simpler search strategies was particularly helpful: it led us to add controls that directly test the learned policy while holding its starting opportunities and scoring interface fixed.

1. Technical contribution: a different decision problem creates different learning and search challenges

NORTH asks which content opportunity to explore next, before choosing a particular item. Its target is the expected best response within a fixed-size, content-grounded set:

E ⁣[max⁡i∈CYui∣Hu],∣C∣=4.\mathbb{E}\!\left[\max_{i\in C}Y_{ui}\mid H_u\right],\qquad |C|=4.

This differs from taking the largest predicted item mean, max⁡i∈CE[Yui∣Hu]\max_{i\in C}\mathbb{E}[Y_{ui}\mid H_u]. It creates three connected challenges: transferring a short history to semantically different content, estimating opportunity at the set level, and discovering additional realizable sets beyond an enumerated bank.

NORTH addresses these challenges through a coordinated design. The population-trained behavioral encoder learns how responses transfer across content. The direction scorer combines population opportunity with a centered, history-dependent correction,

Q(u,C)=AC+α(χu−χˉC)⊤wC,Q(u,C)=A_C+\alpha(\chi_u-\bar\chi_C)^\top w_C,

and is trained against observed core-level value. Its member-based representation evaluates newly grounded cores without fitting new core-specific parameters. The search policy then learns where to allocate proposals and whether and how to move from population-supported anchors. The same frozen scorer evaluates the resulting cores without accessing the target user's held-out outcomes. The contribution is this task-aligned learning and search framework for a distinct upstream decision, with each component addressing one of its information or realization constraints (Sections 3–4).

2. Simpler alternatives now directly test the learned search

We added two controls with the same bank incumbent, initial anchors, grounding, frozen scorer, and 72-attempt budget. Uniform-OPEN replaces the learned policy with uniform anchor selection and random local proposals. Matched RR additionally preserves NORTH's anchor and stay-or-move decision on every draw, replacing only its learned displacement.

Table 13 reports paired BRV differences on common measurable users:

SettingNORTH − Uniform-OPEN [95% CI]NORTH − Matched RR [95% CI]
Full, W=20W=20+0.0065 [0.0033, 0.0129]+0.0043 [−0.0027, 0.0131]
Full, W=50W=50+0.0072 [0.0009, 0.0151]+0.0059 [0.0031, 0.0110]
FO, W=20W=20+0.0121 [0.0084, 0.0192]+0.0130 [0.0039, 0.0227]
FO, W=50W=50+0.0150 [0.0096, 0.0213]+0.0118 [−0.0065, 0.0318]

All four Uniform-OPEN intervals are above zero. The route-matched control also gives intervals above zero on Full W=50W=50 and FO W=20W=20, identifying a contribution from learned displacement beyond the routing decisions in those settings.

We also added the personalized bank incumbent alone under the same OPEN protocol. Across five additional search seeds, NORTH's mean paired improvements are +0.0029, +0.0052, +0.0204, and +0.0362 in the order above; the last three 95% intervals exclude zero (Table 12). This directly separates the additional search benefit from personalized selection within the bank.

These additions appear in Section 5.4 and Appendices G.3–G.5. We appreciate your suggestion: it has made the evidence for the search design substantially more direct.

Thank you also for highlighting the connection between BRV, realized recommendation quality, and feedback observability. We have added an actual single-item selection experiment and explicit coverage accounting to address these points.

3. From opportunity value to the response of an actually selected item

BRV evaluates whether a direction contains a valuable realization. To test whether a downstream recommender can use that opportunity, we now freeze an item ranker and compare selecting one item directly from the catalogue with selecting one item from NORTH's four-item core. Both items are selected by the ranker's scores; evaluation uses the observed response of the selected item, not the maximum outcome in the core or the predicted direction score.

Table 15 gives the paired improvement in selected-item response on users measurable for both procedures:

SettingBPRSASRecLightGCN
Full, W=20W=20+0.0995+0.0644+0.0675
Full, W=50W=50+0.1621+0.0650+0.0373
FO, W=20W=20+0.0404+0.0775+0.0192
FO, W=50W=50+0.0444+0.0186+0.0345

All twelve paired differences favor using the NORTH core. This supplies downstream evidence beyond BRV: three different frozen rankers obtain higher observed selected-item response after direction selection. Section 5.4 and Appendix G.7 now describe this experiment and report its coverage.

4. Why KuaiRec matches the measurement requirements

The dataset choice follows from the task. Evaluating the best outcome of a selected four-item core requires observed responses for all four members. KuaiRec's distinctive combination is near-complete feedback for a shared user–item subset, broader population interaction logs, and content information. Its released small matrix has 99.6% density (Gao et al., 2022). This supports measuring selected cores from real feedback while using content to define their geometry.

In a sparse interaction log, the response to one or more selected members is frequently unknown. Treating those responses as negative, imputing them, or deleting members would change either the evaluation target or the fixed realization budget. Dense target outcomes let us preserve both. We use the broader logs for the Full population-training setting and the dense subset for the FO population-training setting; both are evaluated on held-out FO users. Full training uses 5,765 users outside the FO subset and transfers to the target evaluation without additional population-model training (Section 5.1 and Appendix A.1).

5. Exposure and measurability are handled explicitly

We distinguish population-training support from target-user outcome measurement. Training-donor eligibility uses training feedback only. At evaluation, the core is selected before held-out outcomes are consulted; missing responses remain missing, and an unmeasurable core is not replaced using those outcomes (Appendices A.4–A.5).

The revision now reports each method's measurable-user count, returned-output count, and bank-external fraction separately (Table 7), and compares methods on their pairwise common measurable users with paired intervals (Tables 8 and 12). At seed 42, NORTH is measurable for 269/282 and 263/282 Full-setting users, and 282/282 and 281/282 FO-setting users. Thus the comparisons no longer rely on subtracting means from different user cohorts.

This protocol addresses the concrete observability issue while keeping the training source and the measured target population explicit. Together, the controlled search comparisons and the selected-item experiment strengthen the evidence for the practical value of the proposed direction decision. We are grateful for the guidance that prompted these additions and would greatly appreciate your reassessment in light of the new evidence.

References

Gao, C., Li, S., Lei, W., Chen, J., Li, B., Jiang, P., He, X., Mao, J., and Chua, T.-S. (2022). KuaiRec: A Fully-observed Dataset and Insights for Evaluating Recommender Systems. Proceedings of the 31st ACM International Conference on Information & Knowledge Management (CIKM 2022).

Dear Reviewer,

Thank you again for your thoughtful and constructive feedback. We have posted our detailed point-by-point responses and the corresponding additional experimental results in the discussion above.

To make them easier to locate, we are adding this note directly under your review. We would be very grateful if you could consider these clarifications and new results when reassessing our revised manuscript.

Thank you for your time and consideration.

Comments on the paper

3 comments

Thank you for the precise and constructive review, and for recognizing the distinction between expected best-set value and the largest predicted item mean. Your questions identified exactly which comparisons were needed to separate personalized bank selection from continuous search. We have added these comparisons under the same OPEN protocol, together with output-identity statistics and five additional search seeds.

1. Continuous search versus its own personalized bank incumbent

The new Bank comparator returns the personalized bank incumbent used by NORTH in the OPEN run. It therefore preserves the starting opportunity and personalized scorer while removing continuous proposals. This is a within-protocol comparison. The new results are below. Each BRV difference uses the users measurable for both NORTH and Bank. The five-seed column equally averages the per-seed paired differences for seeds 48, 166, 250, 497, and 909.

SettingSeed-42 difference [95% CI]Five-seed difference [95% CI]Seed-42 bank-external outputs
Full, W=20W=20+0.0035 [−0.0041, 0.0136]+0.0029 [−0.0014, 0.0081]7/282 (2.48%)
Full, W=50W=50+0.0044 [−0.0032, 0.0142]+0.0052 [0.0009, 0.0105]8/282 (2.84%)
FO, W=20W=20+0.0103 [0.0020, 0.0209]+0.0204 [0.0062, 0.0354]92/282 (32.62%)
FO, W=50W=50+0.0249 [0.0087, 0.0465]+0.0362 [0.0226, 0.0504]59/282 (20.92%)

The five-seed mean increment is positive in every setting, and its interval excludes zero on Full W=50W=50 and both FO settings. All five individual FO Bank contrasts are positive at both history budgets (Table 10). These results directly establish the additional measured value of search beyond personalized bank selection in those three aggregate comparisons.

Bank-external means that the sorted four-item identity is absent from the original bank. A nonzero coordinate displacement alone does not count. This statistic therefore measures the discovery of new realizable content sets rather than movement that returns an existing set. Tables 7, 8, and 10–12 provide the output fractions, common-cohort sizes, and paired effects.

2. Generators with the same incumbent and initial anchors

We now report both independent and shared-initialization versions of DreamRec-HAdapt + Q and PreferDiff-HAdapt + Q. The shared variants receive NORTH's personalized bank incumbent and initial anchors, generate their own proposals, and use the same grounding and frozen scorer. They do not receive NORTH's newly visited cores (Appendix G.3).

The five-seed comparisons favor NORTH over both shared generators in Full W=50W=50 and both FO settings, with intervals above zero. For example, against shared DreamRec the effects are +0.0058 [0.0011, 0.0113] on Full W=50W=50, +0.0204 [0.0062, 0.0354] on FO W=20W=20, and +0.0362 [0.0226, 0.0504] on FO W=50W=50. The Full W=20W=20 point estimates are also positive, with intervals containing zero (Table 12).

These matched-start comparisons address the concern that the original advantage could come only from privileged initialization. We are very grateful for this suggestion: the revision now separates the value of the bank incumbent, the additional continuous proposals, and identity-based discovery explicitly.

Thank you also for requesting outcome coverage, common-user comparisons, and uncertainty across search seeds. We have added each of these, with the original seed-42 results kept separate from the five additional search seeds.

3. Coverage and paired evidence for the original OPEN comparison

There are 282 nominal test users in each setting. At seed 42, the measurable-user counts are:

SettingNORTHDreamRec-HAdapt + QPreferDiff-HAdapt + Q
Full, W=20269266262
Full, W=50263268273
FO, W=20282282281
FO, W=50281282282

Table 7 reports these counts for every method, including shared variants. It distinguishes outcome coverage from the number of returned outputs. For example, PreferDiff on Full W=50 returns 279 outputs, has 273 measurable users, and has 8 bank-external outputs. Its coverage is 273/282, while its external fraction is 8/279.

For comparator bb, we define the pairwise common measurable-user set as

U(s,N,b)=U(s,N)∩U(s,b)\mathcal{U}(s,N,b)=\mathcal{U}(s,N)\cap\mathcal{U}(s,b)

and compute the paired BRV difference as

\operatorname{Mean}\left\\{ V(u,\widehat{C}(u,s,N))- V(u,\widehat{C}(u,s,b)) : u\in\mathcal{U}(s,N,b) \right\\}$$ Here, $\widehat{C}(u,s,m)$ denotes the core selected by method $m$ for user $u$ under search seed $s$, and $V(u,C)$ denotes the observed BRV of core $C$ for user $u$. Both methods are therefore compared on exactly the same measurable users. Table 8 reports: | Setting | NORTH − DreamRec [95% CI]; common n | NORTH − PreferDiff [95% CI]; common n | |---|---|---| | Full, W=20 | +0.0175 [0.0026, 0.0373]; 259 | +0.0079 [0.0004, 0.0165]; 253 | | Full, W=50 | +0.0418 [0.0046, 0.0800]; 254 | +0.0778 [0.0228, 0.1347]; 255 | | FO, W=20 | +0.0103 [0.0020, 0.0209]; 282 | +0.0111 [0.0008, 0.0240]; 281 | | FO, W=50 | +0.0249 [0.0087, 0.0465]; 281 | +0.0085 [0.0012, 0.0163]; 281 | **All eight intervals have positive lower bounds.** These are paired effects computed on common measurable users, rather than differences between method-specific means with different outcome coverage. ### 4. Five additional search seeds Tables 9–12 report seeds **48, 166, 250, 497, and 909**, including NORTH's per-seed absolute BRV and coverage, per-seed paired comparisons, and equally weighted aggregate effects with 95% confidence intervals. | Setting | NORTH − independent DreamRec [95% CI] | NORTH − independent PreferDiff [95% CI] | |---|---|---| | Full, W=20 | +0.0484 [0.0029, 0.0949] | +0.0531 [0.0028, 0.1047] | | Full, W=50 | +0.0540 [0.0208, 0.0872] | +0.0753 [0.0263, 0.1274] | | FO, W=20 | +0.0102 [−0.0039, 0.0253] | +0.0092 [−0.0071, 0.0259] | | FO, W=50 | +0.0362 [0.0226, 0.0504] | +0.0251 [0.0095, 0.0425] | **NORTH achieves a positive paired point estimate in all 40 comparisons across five seeds, four settings, and two independent generators** (Tables 10–11). All eight aggregate effects are positive, with six 95% confidence intervals entirely above zero. The two intervals for FO W=20 include zero. These results establish the consistent direction of the observed improvements across additional search runs and provide paired uncertainty estimates beyond the original single-seed comparison. The revised manuscript summarizes these additions in Section 5.3, and Appendices G.2–G.4 give the coverage definitions and full tables. The underlying problem formulation and NORTH method remain unchanged. We sincerely appreciate your thoughtful and constructive feedback. Your questions have helped us make the empirical evidence substantially stronger and more transparent. We hope the additional experiments and detailed revisions have satisfactorily addressed your concerns and clarified the value of our contribution. We would be truly grateful if you would consider revisiting your assessment and score in light of this new evidence.

We sincerely thank all reviewers for their thoughtful and constructive feedback. Our central contribution is to formulate next semantic direction selection as an upstream recommendation problem: deciding where to explore before selecting or generating specific content. Unlike next-item prediction, this requires transferring behavioral evidence across semantic locations, estimating opportunity at the realization-set level, and discovering valuable directions beyond a predefined candidate bank. NORTH addresses these challenges through population-level behavioral learning, personalized direction scoring, and learned continuous search.

Summary of the most important updates and additional experiments incorporated into the revised manuscript.

We have substantially strengthened the empirical evidence while preserving the original problem formulation and method. Specifically, we added:

  1. Continuous search and paired evaluation: Personalized bank-incumbent and shared-initialization comparisons, bank-external discovery rates, and coverage-aware paired evaluations across five additional search seeds (Section 5.3; Appendices G.2–G.4).

  2. Search controls and downstream evaluation: Simpler random-search controls and downstream single-item evaluations with three frozen rankers (Section 5.4; Appendices G.5 and G.7).

  3. Implementation details and sensitivity analyses: Detailed geometric grounding, bank construction, and support-radius sensitivity analyses, together with clearer implementation specifications (Sections 4.1, 4.3, and 5.1; Appendices G.1 and G.6).

These additions directly address the reviewers' concerns and provide stronger evidence for the effectiveness and practical value of NORTH.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.