acceptodds
Under review as a conference paper at ICLR 2027

COIN: Collaborative Intent for Efficient Reasoning in Native Generative Recommendation

Abstract

Reasoning can improve generative recommendation (GenRec) by forming intermediate hypotheses that guide item prediction. While reasoning is natural for LLM-powered GenRec, introducing it into native recommenders is less straightforward due to the lack of a natural language interface. We introduce COIN (Collaborative Intent-aware Generative Recommendation), a lightweight plan-then-execute framework that grounds reasoning in a collaborative intent space. COIN represents an intent as a coarse collaborative region learned from interactions and expresses the plan through a single token, providing guidance complementary to item semantics. During training, a lightweight forecaster supplies plausible intents, and ambiguity-aware transfer trains the generator to predict the target item under planning uncertainty. COIN is trained solely through next-token prediction and does not rely on reinforcement learning, LLM supervision, or recurrent refinement. At inference, the generator predicts and executes its own plan with only one additional token. COIN accommodates different generator architectures and semantic-ID tokenizers, serving as a plug-in for existing native GenRec models. Experiments on four recommendation datasets demonstrate consistent ranking improvements with minimal training and inference overhead, establishing collaborative intent as an efficient interface for reasoning in native generative recommendation.

Primary area:
generative models
Keywords:
Recommender systems, generative recommendation, information retrieval

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_e1qhOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper addresses the reasoning problem in generative recommenders, with a particular focus on lightweight and efficient methods. The paper proposes COIN, which defines collaborative intent as the reasoning language for native generative recommenders and trains a lightweight causal Transformer to predict intents from interaction histories. It then amortizes the planning capability into the native recommender. Experiments on multiple datasets demonstrate the effectiveness and efficiency of this approach.

Critical strengths
  • The experiments are comprehensive, comparing the proposed method with multiple baselines across multiple datasets, thereby demonstrating its effectiveness and efficiency.
  • This paper provides clear theoretical justification and discussion that explain the rationale and principles behind the proposed method, thereby contributing to its interpretation and credibility.
Critical weaknesses
  • The design principles and novelty of some technologies require further clarification.
  • The conclusions of some experiments are insufficient to support the authors' claims.
Questions for authors
  1. COIN divides items into different clusters based on co-occurrence, then trains a model to predict collaborative intent based on user history. However, this process is essentially very similar to first predicting coarse-grained product categories and then predicting specific products. The paper does not adequately explain why this intermediate step can be defined as reasoning or how it differs from traditional intent prediction.
  2. Figure 5 shows that Native GR has learned some of the Teacher’s preferences, but the consistency is not high. More than 20% of the plans do not even appear in the Teacher’s top 10, so the authors need to explain this phenomenon.
  3. Even when there is a significant discrepancy between the plans predicted by Native GR and Teacher, Figure 5(b) shows that the recommendation results are virtually unaffected. Does this suggest that the choice of intent has virtually no effect on GR’s predictions?
  4. Furthermore, the author should show the model’s performance when the native plan is among the top 1, top 10, and outside the top 10 of the Teacher’s top 10 set.
  5. Native GR's planning capability stems entirely from the Teacher's ambiguity-aware transfer, but the authors did not report the accuracy of the Teacher's predicted plans relative to the true intents.
  6. The Introduction should include an example or figure illustrating the COIN design—for example, showing what collaborative intent is and how it relates to recommended items—which would make the motivation easier to understand.
11 comments
AuthorsOpenReview ↗

Response to Reviewer e1qh (1/5)

We sincerely thank the reviewer for taking the time to review our work. We appreciate the reviewer’s recognition of the broad effectiveness and theoretical justification of our work, as well as their thoughtful comments and insightful questions. We address each concern below.

Revised PDF. We have updated the paper with the analyses the reviewer asked for; all changes are in blue. App. B.1.1 reports the teacher's accuracy against the true intents (Table 4). App. B.2.3 adds a coarse-to-fine control that supervises the plan with the next item's cluster (Table 6). App. B.3.2–B.3.3 add an intervention on the executed plan (Fig. 13) and a breakdown of ranking by native–teacher agreement (Fig. 14). Several of these results are easier to read as figures, so we encourage the reviewer to look at the revised PDF. We have also tried to make this response self-contained, with the key numbers in the tables below.

Design principle and novelty. Our work asks one question: How can native generative recommenders (GRs) reason without natural language? LLM-based GRs reason in language: they first write a rationale about what the user's history suggests, then predict the next item. Native GRs are directly trained on interaction data and lack such ability. Prior methods adopt different ways to fill the gap: LLM-assisted methods compress LLM rationales into reason tokens or embeddings, so their reasoning still comes from natural language; latent-reasoning methods learn thinking tokens with no predefined meaning, typically using RL to activate their specific roles. They can be effective, but typically incur heavy costs and high system complexity stemming from specialized training and inference pipelines.

COIN instead explores a novel reasoning language, specialized for recommendation: the collaborative intent space, where each intent token favors a specific direction of item generation. The vocabulary gains its meaning through item co-occurrence, which provides heterogeneous task information that semantic IDs lack. Given a history, COIN first infers an intent token, a compact conclusion about which group of frequently co-engaged items the user is heading toward, and then generates the next item conditioned on it, just as an LLM writes a rationale before answering. COIN injects such ability through a lightweight intent teacher and trains the generator model through plain next-token prediction, with no LLM calls, RL rollouts, or extra model passes. At inference, COIN injects only ONE extra token before the generated item, adding negligible cost. We demonstrate the effectiveness of our recipe through broad evaluation with multiple tokenizers and architectures, showing significant ranking improvements with little computational overhead. This makes COIN a simple-yet-effective plug-in for the native GR family, and it opens a new direction for reasoning in recommendation beyond natural language.

W1 & Q1. Why the plan is reasoning, and how it differs from coarse-to-fine prediction (revised Sec. 3.2)

We thank the reviewer for this question. To answer it, we first revisit the two complementary roles of COIN, which we have made clearer in the revised Sec. 3.2.

  1. Collaborative representation learning. The plan token is trained to match the teacher's intent distribution, and Theorem 1 shows that the planning loss equals the teacher-to-student KL. Minimizing it makes the history representation retain the teacher's collaborative distinctions, as in knowledge distillation (KD).
  2. Plan-conditioned execution, which is the reasoning step. At inference, the model generates its own plan token and then conditions SID generation on it. The plan is an explicit intermediate output, which sets the direction of decoding.
AuthorsOpenReview ↗

Response to Reviewer e1qh (2/5)

W1 & Q1 (continued).

Why this is reasoning. Reasoning derives an intermediate conclusion from the available evidence and uses it to guide the answer: the conclusion is about the input, the model forms it itself, and it shapes what the model concludes next. Chain-of-thought [1] and plan-and-solve prompting [2] follow this pattern in LLMs. COIN's plan token plays exactly this role: it is formed from the history, about where the user's next interaction is heading; the model generates it at inference; and the item is then generated conditioned on it, so a different plan leads to a different recommendation. Prior work in reasoning-enhanced recommendation follows the same paradigm: OneRec-Think [3], GREAM [4], and SIDReasoner [5] first generate an explicit reasoning trace about the user's history and then generate the item.

How COIN differs from coarse-to-fine prediction. Coarse-to-fine prediction, such as predicting an item's category or a coarse prefix of its code before the item, also produces an intermediate token, but that token plays a different role. It is a property of the next item: it splits the prediction into stages rather than forming a conclusion about the user. COIN's plan is instead a function of the history. It summarizes the evidence in the history, expresses it with a token in the collaborative intent space, and keeps several plausible directions open when the history is ambiguous. During training, COIN learns the intermediate step from its teacher: the training targets represent the "rationale" predictable from the history, rather than part of the answer. The following experiment makes this difference concrete.

Experiment: supervising the plan with the next item's cluster (new, App. B.2.3, Table 6)

To test coarse-to-fine supervision directly, we supervise the plan token with the collaborative cluster c(i+)c(i^+) of the observed next item. At inference, the model generates the cluster token by beam search, like the subsequent SID tokens. As a diagnostic, we also feed the same checkpoints the true cluster c(i+)c(i^+) in place of their prediction; this leaks target information and is unavailable in practice. The following table compares these models with COIN by Recall@10 (%):

Plan supervisionBeautyToysSports
No plan token (baseline)6.245.843.42
Next-item cluster (predicted cluster)6.475.863.25
Next-item cluster (true cluster, diagnostic only)59.0356.6350.93
COIN (teacher distribution)7.237.134.07

Target-cluster supervision shifts the bottleneck instead of easing it. The table above shows two versions of target-cluster supervision: when fed the true next-item cluster at inference, the model achieves very high ranking accuracy, whereas with its own predicted cluster, it stays close to the baseline and well below COIN. This is because the target cluster is a function of the next item: although accurate by construction, it is hard to predict from the history. Trained with the true next-item cluster as the plan token, the decoder learns to rely on it: the model succeeds when it is given the correct plan, whereas an early error in the plan rules out the correct item. Thus, supervising the plan with the target cluster does not make next-item prediction easier; it only moves the difficulty to the plan token.

COIN avoids this bottleneck in two ways. Its plan distribution qτ(⋅∣h)q_\tau(\cdot\mid h) is a function of the observed history, so the planning head can learn to match it (Theorem 1). Its decoder is trained under alternative plans that share the same target, so recommendation does not hinge on a single "correct" plan (Theorem 3).

AuthorsOpenReview ↗

Response to Reviewer e1qh (3/5)

Q2. Why do many native plans fall outside the teacher's top 10?

We thank the reviewer for this question. The partial agreement between the teacher and student has two causes, both expected under COIN's design.

1. The two planners read the history through different inputs, by design. The teacher encodes the history with freely learned item-ID embeddings from interactions, while the student encodes it with semantic IDs built from item content. The teacher is useful only when it carries distinctions the student does not already represent, so we chose this heterogeneity deliberately (Sec. 3.1). App. B.1.2 compares two kinds of teachers: one with ID embeddings and one with semantic embeddings. To measure the recoverability of the teacher's forecasts from the student model, we fit a linear probe from the baseline generator to the teacher's outputs. The following table reports the normalized probe error (Fig. 8(a); higher means less recoverable):

Teacher inputBeautyToysSports
Item-ID embeddings (COIN)0.6330.6360.439
Semantic embeddings0.3250.3570.206

We then train a student with each teacher. The following table reports the student's Recall@10 (%) at matched entropy levels (η = 0.10 / 0.25 / 0.40):

Teacher inputBeautyToysSports
Item-ID embeddings (COIN)6.92 / 7.03 / 6.787.29 / 7.46 / 6.854.08 / 3.98 / 3.95
Semantic embeddings6.71 / 6.75 / 6.716.14 / 6.23 / 6.523.54 / 3.50 / 3.52

The ID-embedding teacher's forecasts are much harder to recover from the student's representation, and they also yield a better student at every matched entropy level. This suggests that the two planners process a history in genuinely different ways. This is what makes the teacher useful, and it also limits the recoverability of the teacher's plan from the student.

2. Some histories are inherently difficult. For a history with weak or mixed evidence, the teacher's forecast is diffuse, spreading its probability over many intents. Since COIN learns from plans sampled from this forecast, the student never sees a stable target for such histories. As a result, its prediction can land on many intent regions, even outside the teacher's top 10. Q3 & Q4 show that the teacher-student disagreements concentrate on such histories.

Reviewer_zo2AOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes COIN, a lightweight planning framework for native generative recommendation. It constructs collaborative intent clusters from interaction data and trains a teacher to predict an intent distribution from user history. The student learns to predict sampled intent tokens and generate the same target item under these alternative plans through next-token prediction. At inference, the student predicts and executes its own plan with one additional token. Experiments cover four datasets, two generator architectures, and three semantic-ID tokenizers.

Critical strengths
  1. The method is simple and derives reasoning supervision from collaborative intents in interaction data, without natural-language reasoning supervision. Its single-token planning interface also supports integration with different native generative recommendation architectures.
  2. The experiments show consistent gains across multiple datasets, architectures, and SID tokenizers, with broad comparisons against different recommendation approaches.
Critical weaknesses
  1. The necessity of teacher-distribution supervision is not sufficiently established. See Question 1.
  2. How shared-target training balances useful planning with cross-plan consistency remains unclear. Supervising different plans with the same target may also reduce the model’s reliance on the plan. See Question 2.
Questions for authors
  1. Are there ablations supporting the advantage of teacher-distribution supervision over: (1) directly supervising the plan token with the target item’s collaborative cluster label, without a teacher; and (2) sampling training plans from the student’s current plan-token distribution rather than the teacher’s?
  2. Please clarify the rationale for pairing different sampled plans (r) with the same target (Y). Could this introduce plans that are inconsistent with the target, or encourage the model to ignore the planning token?
6 comments
AuthorsOpenReview ↗

Response to Reviewer zo2A (1/5)

We sincerely thank the reviewer for their positive feedback on our work. We appreciate the reviewer’s recognition of the simple and effective design of our work, as well as their thoughtful comments and insightful questions. We address each concern below.

Revised PDF. We have updated the paper with the controls and analyses the reviewer asked for; all changes are in blue. App. B.2.3 adds both requested supervision controls: target-cluster labels and self-sampled plans (Table 6). Sec. 3.2 adds a new Theorem 2, which states when plan conditioning improves prediction under shared targets. App. B.3.1 visualizes how each plan shifts item likelihoods (Figs. 10–12), and App. B.2.2 adds a temperature sweep that measures how much the model relies on the plan (Fig. 9). Several of these results are easier to read as figures, so we encourage the reviewer to look at the revised PDF. We have also tried to make this response self-contained, with the key numbers in the tables below.

W1 & Q1. Why teacher-distribution supervision (new, App. B.2.3, Table 6)

We thank the reviewer for the insightful question. Here, we aim to answer three sub-questions: why the method needs a teacher at all, why the teacher's forecast rather than the next item's cluster, and why a distribution.

1. Why a teacher at all? To learn to plan, a model needs examples of intermediate steps that guide the answer. LLMs learn such steps from reasoning traces, either annotated by humans or distilled from a stronger teacher, such as DeepSeek-R1 [1]. Interaction logs contain no such steps: they record which item a user interacted with next, but not the intent that guided that choice. A teacher is needed to fill this gap: in COIN, the teacher reads the history through learned item-ID embeddings and forecasts a distribution qτ(⋅∣h)q_\tau(\cdot\mid h) over collaborative intents, which serves as the target of the plan token.

2. Why not the next item's cluster? An intent should be inferred from the history rather than read off the target. The cluster label c(i+)c(i^+) is a property of the next item, while the teacher's forecast is a function of the history. It represents what the history makes predictable about the next item. This makes it a better source for learning to reason, because it provides the "rationale" before the answer.

3. Why a distribution? A distribution over intents conveys more than a single label: it tells the student how plausible each intent is relative to the others. As in knowledge distillation (KD) [2], matching these relative probabilities teaches the student finer distinctions that are not present in its own representation. Theorem 1 makes this precise.

AuthorsOpenReview ↗

Response to Reviewer zo2A (2/5)

W1 & Q1 (continued).

Experiment: alternative supervisions. In this part, we train COIN with alternative supervision sources: The target-cluster model supervises the plan token with c(i+)c(i^+). The self-sampled model samples a plan from the current student's plan distribution and pairs it with the observed target. At inference, the target-cluster model generates its cluster token by beam search, like the subsequent SID tokens, and the self-sampled model executes its own predicted plan. As a diagnostic, we also feed the target-cluster model the true cluster c(i+)c(i^+) in place of its prediction; this leaks target information and is unavailable in practice. The following table compares them with COIN by Recall@10 / NDCG@10 (%), averaged over three seeds:

Plan supervisionBeautyToysSports
No plan token (baseline)6.24 / 3.455.84 / 3.273.42 / 1.88
Target cluster (predicted cluster)6.47 / 3.625.86 / 3.193.25 / 1.71
Target cluster (true cluster, diagnostic only)59.03 / 33.9156.63 / 32.4350.93 / 28.95
Self-sampled plans6.29 / 3.546.07 / 3.353.23 / 1.72
COIN (teacher distribution)7.23 / 4.157.13 / 3.894.07 / 2.26

Target-cluster supervision shifts the bottleneck instead of easing it. The table above shows two versions of target-cluster supervision: when fed the true next-item cluster at inference, the model achieves very high ranking accuracy, whereas with its own predicted cluster, it stays close to the baseline and well below COIN. This is because the target cluster is a function of the next item: although accurate by construction, it is hard to predict from the history. Trained with the true next-item cluster as the plan token, the decoder learns to rely on it: the model succeeds when it is given the correct plan, whereas an early error in the plan rules out the correct item. Thus, supervising the plan with the target cluster does not make next-item prediction easier; it only moves the difficulty to the plan token.

Self-sampled plans provide no new information. Self-sampled plans are not tied to the target and avoid this bottleneck, but they stay close to the baseline and well below COIN on every dataset. A plan sampled from the student's own prediction depends on the history only through the student's representation ZZ. It carries no information beyond ZZ, and the counterpart of the representation term Rτ=Iτ(R;h∣Z)\mathcal R_\tau=I_\tau(R;h\mid Z) in Theorem 1 is zero, so learning to predict such plans adds nothing to the representation.

COIN avoids both limitations. Its teacher's forecast qτ(⋅∣h)q_\tau(\cdot\mid h) is a function of the observed history, so the planning head can learn to match it, and the teacher distribution conveys collaborative distinctions that the student does not already represent (Theorem 1).

AuthorsOpenReview ↗

Response to Reviewer zo2A (3/5)

W2 & Q2. Why pair different plans with the same target, and does it weaken the plan? (revised Sec. 3.2)

We thank the reviewer for this insightful question. Pairing sampled plans with the same target is how COIN prepares execution for an imperfect plan.

Rationale. At inference, the native planner predicts its own plan with no teacher to correct it, and it cannot always generate the single "correct" plan, for two reasons.

1. The user's intent is inherently uncertain and multimodal. A history rarely determines a single next intent: it can support several plausible continuations [3, 4], and the observed next item reveals only one of them. Even a perfect planner cannot know in advance which one will be realized, which is why the teacher forecasts a distribution rather than a single label.

2. The student cannot exactly copy the teacher. The two planners read the history through different inputs: learned item-ID embeddings for the teacher and content-based semantic IDs for the student. This is a deliberate design, as the teacher is useful only when it carries distinctions the student does not already represent (Sec. 3.1 and 3.2). It also means that some of the teacher's distinctions are not recoverable from the student's representation. To measure the recoverability of the teacher's forecasts from the student model, we fit a linear probe from the baseline generator to the teacher's outputs (App. B.1.2, Fig. 8(a)). The normalized probe error is 0.633 / 0.636 / 0.439 on Beauty / Toys / Sports, compared with 0.325 / 0.357 / 0.206 for a teacher built on semantic embeddings (higher means less recoverable). From the results, the ID-embedding teacher's forecasts are harder to recover from the student's representation.

Both reasons lead to the same consequence: the student often generates a plan other than the "correct" one, so its item prediction must not rely on that single plan. Plan sampling during training ensures this. Each time a history is presented, a plan is drawn from the teacher's distribution and paired with the observed target, so the decoder learns to reach the target under every plan the history supports, not only under the most likely one. Theorem 3 formalizes the result: the ranking regret of executing the native plan k^\hat k is bounded by 2ετ(h)/qk^\sqrt{2\varepsilon_\tau(h)/q_{\hat k}} (plus a beam-search term), where ετ(h)\varepsilon_\tau(h) is the execution error and qk^q_{\hat k} is the teacher's support for that plan. With sampled plans, any plan with sufficient support yields an accurate ranking.

Reviewer_7zByOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

COIN introduces a collaborative-intent token before SID generation in native generative recommenders. A lightweight teacher supplies intent distributions, and ambiguity-aware next-token training learns planning and execution under alternative plans. Experiments show consistent gains across datasets, architectures, and tokenizers.

Critical strengths
  1. The framework introduces collaborative guidance through a simple, self-contained inference pipeline.
  2. Evaluation across four datasets, two architectures, and three tokenizers supports its broad applicability.
  3. Detailed ablations distinguish representation learning from plan execution and examine robustness and efficiency.
Critical weaknesses
  1. The distinction between reasoning and collaborative knowledge distillation remains insufficiently established.
  2. Training different plans toward the same target may weaken plan specificity; the theory does not establish when plan conditioning improves recommendation.
  3. Comparisons lack simpler intermediate-supervision controls, such as hard cluster labels and direct soft-label distillation.
Questions for authors
  1. How does transfer temperature affect both ranking quality and dependence on the plan token?
  2. Which histories benefit most from plan execution beyond representation-only training?
  3. How does COIN compare with reasoning-enhanced generative recommenders (e.g., CARE [1]) in recommendation accuracy and inference cost under matched backbones and tokenizers?
  4. How does COIN compare with using the target item’s CF cluster or the teacher’s top-one intent as supervision?

[1] Lin et al. Bringing Reasoning to Generative Recommendation Through the Lens of Cascaded Ranking. WWW 2026.

11 comments
AuthorsOpenReview ↗

Response to Reviewer 7zBy (1/5)

We sincerely thank the reviewer for taking the time to review our work. We appreciate the reviewer’s recognition of the simple and effective design of our work, as well as their thoughtful comments and insightful questions. We address each concern below.

Revised PDF. We have updated the paper with the theory and experiments the reviewer asked for; all changes are in blue. Sec. 3.2 adds a new Theorem 2, which states when plan conditioning improves prediction. App. B.2–B.3 add the new controls and analyses: direct soft-label distillation, target-cluster supervision, teacher top-one supervision, a temperature sweep with plan-dependence measurements, and execution gains by history type. Several of these results are easier to read as figures, so we encourage the reviewer to look at the revised PDF. We have also tried to make this response self-contained, with the key numbers in the tables below.

W1. Reasoning vs. collaborative knowledge distillation (revised Sec. 3.2)

We thank the reviewer for raising this point. Sec. 3.2 of the submitted paper discusses the two complementary roles of COIN, and we have revised the section to make the distinction clearer.

  1. Collaborative representation learning, which is closer to distillation. The plan token is trained to match the teacher's intent distribution, and Theorem 1 shows that this planning loss equals the teacher-to-student KL. Minimizing it makes the history representation retain the teacher's collaborative distinctions, as in knowledge distillation (KD).
  2. Plan-conditioned execution, which is the reasoning step. At inference, the model generates its own plan token and then conditions SID generation on it. The plan is an explicit intermediate output: it selects one decoder from a family of plan-specific decoders, so changing the plan changes the recommendation. KD has no such step.

The same distinction appears in LLM reasoning. Small reasoning models are often obtained by fine-tuning on reasoning traces generated by a stronger teacher, as with the distilled models of DeepSeek-R1 [1]. This training is distillation in form, yet the result is a reasoning model, because at inference it generates intermediate steps and conditions its answer on them, as in chain-of-thought [2] or plan-and-solve prompting [3]. Distilling only the teacher's answer distribution would be ordinary KD. COIN follows this plan-then-execute pattern with a deliberately minimal step, a single plan token, which keeps reasoning affordable for native GRs.

Experiment: separating the two mechanisms (App. B.2.1, Table 5; KD row new)

To test whether COIN's gain goes beyond distillation, we compare it with two variants that learn from the same teacher distribution qτ(⋅∣h)q_\tau(\cdot\mid h) but never execute a plan. Standard soft-label distillation (KD) adds an extra loss λKL(qτ(⋅∣h)∥aKD(⋅∣Z))\lambda\mathrm{KL}(q_\tau(\cdot\mid h)\Vert a_{\rm KD}(\cdot\mid Z)) alongside the SID loss. Representation-only COIN keeps COIN's planning objective but generates SIDs without feeding the plan to the decoder. Full COIN conditions SID generation on its own predicted plan. The following table compares the four models by Recall@10 / NDCG@10 (%):

VariantBeautyToysSports
Baseline (no plan token)6.37 / 3.535.87 / 3.253.52 / 1.91
KD (same teacher, no plan token)6.87 / 3.856.71 / 3.943.81 / 2.11
COIN, representation only (plan predicted, not executed)6.98 / 4.046.90 / 3.943.78 / 2.11
COIN, full (plan predicted and executed)7.25 / 4.217.32 / 4.084.10 / 2.25

Both variants without plan execution improve over the baseline, consistent with collaborative representation learning. Conditioning SID generation on the plan improves further on every dataset, which isolates the contribution of plan-conditioned execution. W2 below formalizes when this gain arises (Theorem 2).

AuthorsOpenReview ↗

Response to Reviewer 7zBy (2/5)

W2. Plan specificity under shared targets, and when plan conditioning helps (new Theorem 2, Sec. 3.2; proof in App. D.3)

We thank the reviewer for this insightful point. COIN's training loss splits into a planning error and an execution error (Sec. 3.2). The submitted theory explained the planning error (Theorem 1) and what a small execution error guarantees at inference (formerly Theorem 2, now Theorem 3). The revision adds a new Theorem 2, which shows when conditioning on a plan can make the execution error smaller than any plan-free decoder can achieve. We restate it below.

Setup.

  • A history hh is followed by a next item YY with true distribution πh=P(Y∣h)\pi_h=P(Y\mid h). The student encodes hh as a representation ZZ.
  • The teacher gives a distribution qτ(k∣h)q_\tau(k\mid h) over KK plans. Each time (h,Y)(h,Y) is presented in training, COIN samples a plan k∼qτ(⋅∣h)k\sim q_\tau(\cdot\mid h) and trains on the sequence (plan kk, then YY). All plans sampled for hh therefore share the target YY.
  • Conditioning on plan kk yields a plan-specific decoder fk(⋅∣Z)f_k(\cdot\mid Z), a distribution over next items, from a decoder class D\mathcal D (what the generator can express). A decoder's error at hh is KL(πh∥f(⋅∣Z))\mathrm{KL}(\pi_h\Vert f(\cdot\mid Z)), its excess log-loss.
  • The execution error is Eτ=Eh[ετ(h)]\mathcal E_\tau=\mathbb E_h[\varepsilon_\tau(h)], where ετ(h)=∑kqτ(k∣h)KL(πh∥fk(⋅∣Z))\varepsilon_\tau(h)=\sum_k q_\tau(k\mid h)\mathrm{KL}(\pi_h\Vert f_k(\cdot\mid Z)). Grouping its terms by plan gives Eτ=∑kwkDk(fk)\mathcal E_\tau=\sum_k w_kD_k(f_k). Here wk=Eh[qτ(k∣h)]w_k=\mathbb E_h[q_\tau(k\mid h)] is plan kk's average training weight, Pk(h)=qτ(k∣h)P(h)/wkP_k(h)=q_\tau(k\mid h)P(h)/w_k is its training distribution over histories, and Dk(f)=Eh∼PkKL(πh∥f(⋅∣Z))D_k(f)=\mathbb E_{h\sim P_k}\mathrm{KL}(\pi_h\Vert f(\cdot\mid Z)) is a decoder's error on that distribution. Each plan's decoder is thus trained on its own distribution of histories.

Theorem 2 (intent specialization through plan conditioning). For a fixed decoder class D\mathcal D, let e0=inf⁡f∈DEhKL(πh∥f(⋅∣Z))e_0=\inf_{f\in\mathcal D}\mathbb E_h\mathrm{KL}(\pi_h\Vert f(\cdot\mid Z)) and esep=∑kwkinf⁡f∈DDk(f)e_{\rm sep}=\sum_k w_k\inf_{f\in\mathcal D}D_k(f). These are the smallest execution errors attainable when all plans share one decoder and when each plan has its own, respectively. Then ΔD=e0−esep≥0\Delta_{\mathcal D}=e_0-e_{\rm sep}\ge 0. When these infima are attained, ΔD>0\Delta_{\mathcal D}>0 if and only if no single decoder in D\mathcal D minimizes every DkD_k. Moreover, ΔD=0\Delta_{\mathcal D}=0 if qτ(k∣h)q_\tau(k\mid h) does not depend on hh, or if some f∈Df\in\mathcal D satisfies f(⋅∣Z)=πhf(\cdot\mid Z)=\pi_h for almost every hh.

Discussion. By conditioning the decoding function on the plan, the model obtains a family of plan-specific decoders, each able to focus on the histories associated with its intent. With limited capacity, a single decoder may need to compromise when fitting the transition patterns associated with different intents. Plan-specific decoders can reduce this compromise, with ΔD\Delta_{\mathcal D} measuring the potential reduction in prediction error.

AuthorsOpenReview ↗

Response to Reviewer 7zBy (3/5)

W2 (continued): what Theorem 2 implies for the reviewer's two concerns.

(a) Do shared targets weaken plan specificity? No. A target is shared only among plans sampled for the same history. Across histories, history-dependent plan probabilities route different histories to different plans, so each plan's decoder is fit to its own distribution PkP_k and learns the transition patterns of its own histories. The specificity remains as long as PkP_k weighs histories differently. Within a history, plausible plans are trained toward the same πh\pi_h, which softly couples plans and item predictions, so their recommendations overlap. This is intended, because a plan is a hypothesis that steers generation rather than a filter.

Theorem 2 concerns the best decoders, and the trained models show the same behavior:

  • Fig. 6(a): forcing different plans on the same history, nearby plans share most of their top-10 lists (median 8–9 of 10), while the most distant plans share about 3.
  • Figs. 10–12 (new, App. B.3.1): with ZZ fixed, each plan raises the likelihood of a compact item region around its own collaborative cluster. We encourage the reviewer to view these maps in the PDF, as they clearly demonstrate how plan specificity and overlap can coexist.

(b) When does plan conditioning help? When no single decoder in D\mathcal D is best for every plan's histories, that is, when one decoder would have to compromise between the transition patterns of different intents. This holds in our setting, where a compact decoder must serve a large item catalog and diverse user interests [4, 5]. Plan conditioning then lets each plan's decoder specialize, reducing the execution error by ΔD>0\Delta_{\mathcal D}>0. This matches our results: executing the plan improves on representation-only training (W1) and on an averaged plan embedding with the same representation (Q2).

W3 & Q4. Simpler intermediate supervision (new, App. B.2.3, Table 6)

We added all of the requested controls, each changing only how the plan token is supervised. Direct soft-label distillation is reported under W1, and teacher top-one supervision under Q1. Here we test whether a teacher is needed at all, using the most direct form of collaborative supervision: the plan token is trained on the CF cluster c(i+)c(i^+) of the observed next item i+i^+ as a hard label. At inference, the model generates the cluster token by beam search, like the subsequent SID tokens. As a diagnostic, we also feed the same checkpoints the true cluster c(i+)c(i^+) in place of their prediction; this leaks target information unavailable in practice. The following table compares these forms of plan supervision with COIN by Recall@10 (%):

Plan supervisionBeautyToysSports
No plan token (baseline)6.245.843.42
Target cluster (predicted cluster)6.475.863.25
Target cluster (true cluster, diagnostic only)59.0356.6350.93
COIN (teacher distribution)7.237.134.07

Target-cluster supervision shifts the bottleneck instead of easing it. The table above shows two versions of target-cluster supervision: when fed the true next-item cluster at inference, the model achieves very high ranking accuracy, whereas with its own predicted cluster, it stays close to the baseline and well below COIN. This is because the target cluster is a function of the next item: although accurate by construction, it is hard to predict from the history. Trained with the true next-item cluster as the plan token, the decoder learns to rely on that token: the model succeeds when it is given the correct plan, whereas an early error in the plan rules out the correct item. Thus, supervising the plan with the target cluster does not make next-item prediction easier; it only moves the difficulty to the plan token.

COIN avoids this bottleneck in two ways. Its plan distribution qτ(⋅∣h)q_\tau(\cdot\mid h) is a function of the observed history rather than of the unobserved next item, so the planning head can learn to match it (Theorem 1). Its decoder is trained under alternative plans that share the same target, so recommendation does not hinge on a single correct plan (Theorem 3).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.