betaFeedback

WorksComputer Science

Reflection Is Not Always Worth It: Budgeted Reflection Allocation for Recommendation Agents

Abstract

This work aims to make recommendation agents more efficient by reducing unnecessary LLM reflection while preserving full-catalog ranking and inference reranking. Since BRAVE changes only how training-time reflection is allocated, it does not introduce new user data requirements or additional LLM exposure beyond the base agent protocol. The main potential risk is that a poorly calibrated allocator may assign less reflection to users whose preferences are sparse, niche, or under-represented, which could weaken personalization for those users or indirectly affect item exposure. This risk can be mitigated through deployment-time auditing and conservative allocation design. In practice, reflection budgets should be evaluated not only by aggregate Recall or NDCG, but also across user-history lengths, item popularity groups, and application-relevant cohorts when such analysis is permitted. The allocator should rely on task-side ranking evidence rather than sensitive attributes, and fallback policies can reserve a small portion of the budget for random or fairness-aware coverage when subgroup robustness is important. Privacy risks can be controlled with the same protections required by the base recommendation agent: minimizing user text sent to the LLM controller, anonymizing user identifiers, retaining only necessary interaction evidence, and applying standard access-control and logging policies. Overall, BRAVE is designed as a cost-aware allocation layer, and its ethical deployment depends on validating that the saved computation does not come at the expense of vulnerable users or responsible exposure.

CCS Concepts

  • Information systems → Recommender systems;

Keywords

Recommender Systems, Large Language Models, Language Agents, Reflection Allocation, Budgeted Learning, Value Estimation

ACM Reference Format: Jinfeng Xu, Zheyu Chen, Ziyue Peng, Jianheng Tang, Zheng Lin, Wenhao Yuan, Jian Chen, Shujie Li, and Victor Leung. 2027. Reflection Is Not Always Worth It: Budgeted Reflection Allocation for Recommendation Agents. In Proceedings of the Twentieth ACM International Conference on Web Search and Data Mining (WSDM ’27), February 15–19, 2027, Hong Kong, Hong Kong. ACM, New York, NY, USA, 11 pages. https://doi.org/10.1145/3862322.3863445

Introduction

Large language models (LLMs) have become an appealing interface for recommender systems because they can reason over user histories, item text, and intermediate ranking evidence [1, 9, 12, 31, 36, 49, 50]. Yet direct LLM recommendation is still poorly matched to full-ranking scenarios: LLMs may hallucinate non-existent item IDs, struggle with large catalogs, and incur substantial latency when asked to reason over many candidates [13, 14]. A recent line of work therefore casts the LLM as an agent that orchestrates conventional recommenders as external modules [42, 43, 51], echoing the broader shift from passive generation to language agents that call external modules and APIs [24, 34, 44]. In this setting, recommender tools provide valid full-catalog rankings, and the LLM reasons over their exposed evidence.

This architecture raises a question that classical recommendation rarely needs to ask: how should reflection be allocated? Existing recommendation agents mainly study prompts, memories, and recommender-tool interfaces, and typically apply reflection to every sampled user once reflection is deemed useful. The policy is simple, but the value of reflection is not uniform. Recent LLM routing and

This work is licensed under a Creative Commons Attribution 4.0 International License. WSDM ’27, Hong Kong, Hong Kong. © 2027 Copyright held by the owner/author(s). ACM ISBN 979-8-4007-3117-4/2027/02 https://doi.org/10.1145/3862322.3863445 cascade studies make a similar point at the system level: expensive LLM computation should be invoked selectively when cost and quality must be balanced [3, 47]. In recommendation agents, some users have sharp and mutually consistent recommender-tool rankings, so the mixture over recommenders is nearly determined before LLM feedback. Others have flat score margins or conflicting recommender evidence, where a small change in personalized tool-mixture memory can move relevant items near the top of the list. Treating both groups identically conflates reflection capability with reflection allocation. The distinction matters because LLM judgments are useful but noisy: unnecessary updates can consume calls and perturb reliable mixture memories, especially when LLM evaluators exhibit order or self-evaluation biases [32, 41].

We study reflection allocation as a first-class problem for recommender agents. The goal is to separate two roles that are often entangled. The reflector defines how LLM feedback changes a user’s tool-mixture memory, following the broader literature on language-agent reflection and self-feedback [18, 28, 52]; the allocator decides which users should receive that feedback. Personalized recommender-tool learning agents mainly improve the reflector. This paper asks whether the allocator can be built from recommender evidence before any LLM call is made. The objective is not only to reduce cost, but also to improve agent training under a limited reflection budget.

We instantiate this idea as BRAVE, short for Budgeted Reflection Allocation via Value Estimation. Before invoking the LLM, BRAVE estimates which users are likely to benefit from reflection and allocates the limited budget to those users. The estimator is gated: margin-based ranking fragility is the primary signal, while cross-tool disagreement is admitted only when the current ranking is sufficiently fragile. This requires no additional model training and no extra LLM calls. BRAVE is not a new recommender backbone or reflection prompt; it decides when the existing reflection operation should be spent. The recommender tools still rank the whole catalog, and inference-time reranking remains unchanged, isolating the paper’s central claim that reflection should be allocated rather than automatically spent.

The design also clarifies a common intuition about uncertainty. Cross-tool disagreement sounds like a natural reason to reflect: if recommender tools disagree, the LLM should decide which recommender to trust. This intuition is related to classic query-by-committee active learning, where committee disagreement identifies informative examples [26]. In recommendation agents, however, disagreement is not always value. It may indicate a recoverable conflict where one recommender exposes the relevant neighborhood, or it may indicate that all recommenders are locally unreliable. In the latter case, reflection spends calls reallocating weight among weak alternatives. We therefore distinguish ranking fragility, which measures whether a memory update can change the decision, from cross-tool disagreement, which measures whether recommender tools produce different evidence. BRAVE operationalizes this distinction with a gate: margin uncertainty first identifies users near a ranking decision boundary [29], and disagreement is then used only as a conditional secondary signal.

We evaluate BRAVE on full-ranking recommendation benchmarks. Across datasets and reflection budgets, BRAVE preserves the accuracy benefits of uniform reflection while using fewer LLM calls, and its strongest budget setting improves top-ranked quality over a full-reflection agent baseline. A controlled branch comparison further shows that the allocation score is correlated with realized user-level reflection gain, rather than only improving aggregate metrics after the fact. We also find that disagreement is useful only when conditioned on ranking fragility. The broader message is that recommendation agents should optimize not only how they reflect, but also when reflection is warranted and which uncertainty signals should control that decision.

Related Work

LLMs for Recommendation

LLM-based recommendation has moved from semantic scoring toward reasoning over user histories, item text, and ranking evidence. Early approaches formulate recommendation as point-wise preference prediction, prompt-based scoring, or zero-shot reranking [1, 9, 12, 37, 38]. They exploit textual knowledge but often operate on restricted candidate sets and remain weakly coupled with full-catalog collaborative signals. A second line aligns collaborative representations with language-model spaces, enabling LLMs to consume user/item embeddings or generate recommendation outputs [14, 50]. More recent work further emphasizes preference reasoning beyond shallow semantic matching; for example, ThinkRec uses reasoning activation and personalized expert fusion to improve LLM-based recommendation [46].

Together, these studies show that LLMs can enrich recommendation with semantic and reasoning ability, but they also expose a deployment tension: direct LLM recommenders must generate valid item identifiers, rerank limited candidates, or pay substantial inference and training costs. BRAVE takes a complementary route. It keeps full-catalog scoring inside conventional recommenders and uses the LLM only for agentic reflection and final reranking. The key question therefore shifts from making the LLM recommend directly to deciding when its costly reflection is worth invoking during agent training.

Language Agents with External Modules

Language agents often extend LLMs beyond pure text generation by connecting them to external knowledge sources, symbolic modules, APIs, or specialist models [16, 27]. This line is commonly described as LLM tool use, but the “tools” in that literature range from search engines and calculators to code interpreters and domain models. Representative systems interleave reasoning and action [44], learn module invocation from self-supervised demonstrations [24], improve API calling with retrieval-aware training [22], and optimize module-use trajectories through contrastive or trajectory-level reasoning [34]. These works establish external-module access as a way to mitigate missing knowledge, limited exact computation, and unreliable direct generation.

Recommendation agents instantiate this paradigm where output validity and catalog scale are central. Rather than asking an LLM to generate item identifiers, module-based recommendation agents use conventional recommenders to produce valid full-catalog rankings and let the LLM reason over recommender evidence. ToolRec studies LLM-empowered selection among recommender modules [51], AgentDR uses LLM agents for dynamic recommendation [42], and PRTA learns personalized tool-mixture memories through local recommender evaluation, recommender comparison, and ranking comparison [43]. BRAVE studies a complementary dimension: given the same reflection operation, which users should receive LLM feedback under a budget? This separates the reflector from the allocator and treats user-level reflection as a selective training decision.

Cost-Aware Reasoning and Uncertainty-Based Selection

BRAVE is also related to cost-aware reasoning and selective computation. LLM routing and cascade methods invoke expensive reasoning only when cheaper models or intermediate states are insufficient [3, 47]. Adaptive computation, early exiting, and selective prediction similarly allocate computation or coverage according to confidence [7, 8, 10, 35]. Their shared principle is that expensive computation should follow expected value rather than a uniform schedule. Reflection allocation differs from model routing or abstention because the decision is not which model answers a query, but whether an LLM should update a user’s personalized tool-mixture memory during training. Its value is measured by downstream top-KK ranking after the memory update, making recommender-side evidence especially useful before any LLM call is made. This connects BRAVE to active learning: uncertainty sampling and margin selection prioritize examples near decision boundaries [17, 25, 29], while query-by-committee uses disagreement as an informativeness signal [26]. In recommendation agents, however, disagreement can indicate either complementary evidence or uniformly weak recommenders, so BRAVE gates disagreement by ranking fragility instead of treating it as an unconditional objective.

Problem Setup

We consider an agentic full-ranking setting in which conventional recommenders produce valid catalog-wide scores and an LLM agent maintains personalized tool-mixture memories. The key resource is training-time reflection: it can improve the mixture memory, but each call is expensive. We therefore formulate reflection as a budgeted allocation problem that BRAVE must approximate before making LLM calls. Throughout this paper, unless we explicitly discuss external-module language-agent literature, a tool refers to a pretrained full-catalog recommender. A tool pool is a set of such recommenders, and a tool-mixture memory denotes the user-specific weights used to combine their scores.

Agentic Full-Ranking Recommendation

Let U\mathcal{U} and I\mathcal{I} denote users and items. For each user uu, a recommender must rank all items in I\mathcal{I} and is evaluated on held-out future interactions. We assume a set of pretrained recommender tools T={1,…,M}\mathcal{T}=\{1,\ldots,M\}. Each recommender tool mm returns a full score vector sum∈R∣I∣\mathbf{s}^{m}_{u}\in\mathbb{R}^{|\mathcal{I}|} over the complete item catalog. The formulation is agnostic to the concrete recommender family: the pool may include graph-based, sequential, collaborative-filtering, diffusion-based, or other full-catalog recommenders, as long as they expose scores or top-ranked lists.

The agent maintains a personalized tool-mixture memory wu∈RM\mathbf{w}_{u}\in\mathbb{R}^{M} for each user and aggregates recommender-tool scores as:

su=∑m=1Mwu,msum.(1)\mathbf{s}_{u}=\sum_{m=1}^{M}w_{u,m}\mathbf{s}^{m}_{u}. \tag*{(1)}

The LLM therefore does not generate item IDs or rank the full catalog directly. Its role is to update the tool-mixture memory by judging textual evidence from user profiles, recent interactions, and top-ranked candidate lists.

Reflection as Costly Agentic Computation

Such an agent can update memory with three complementary signals: LLM-based local recommender evaluation, LLM-based global recommender comparison, and a non-LLM ranking comparison [43]. Local evaluation judges each recommender tool independently, while global comparison chooses the best recommender among candidate lists. With MM recommender tools, nn users, and tt epochs, a uniform full-reflection policy uses nn profile calls, (M+1)nt(M+1)nt training reflection calls, and nn inference reranking calls, for (M+1)nt+2n(M+1)nt+2n total LLM calls. We define a reflection policy π(u)∈{0,1}\pi(u)\in\{0,1\} that decides whether user uu receives training-time LLM reflection. Given a budget B∈(0,1]B\in(0,1], the policy must satisfy

∑u∈U0π(u)≤B∣U0∣,(2)\sum_{u\in\mathcal{U}_0}\pi(u)\leq B|\mathcal{U}_0|, \tag*{(2)}

where U0⊆U\mathcal{U}_0\subseteq\mathcal{U} is the sampled-agent window used by the LLM-agent stage. Following PRTA [43], this window controls the number of user-level LLM calls; it does not restrict recommender training or the item catalog. The learning objective is to maximize ranking utility after the memory update under this budget:

max⁡π R(Wπ+)−λC(π),(3)\max_{\pi}\ \mathcal{R}(\mathbf{W}^{+}_{\pi})-\lambda\mathcal{C}(\pi), \tag*{(3)}

where R\mathcal{R} is a top-KK ranking metric, Wπ+\mathbf{W}^{+}_{\pi} denotes the post-update tool-mixture memories induced by policy π\pi, C(π)\mathcal{C}(\pi) is the realized LLM-call cost, and λ\lambda controls the cost-accuracy trade-off. Directly optimizing (3) is unacceptable before making LLM calls, so the central challenge is to build a cheap pre-reflection surrogate for the value of reflection.

Budgeted Reflection Allocation

This section presents BRAVE2 as a budgeted reflection allocator on top of a fixed recommendation agent. BRAVE keeps the recommender tools, reflection prompts, and inference reranking unchanged, and estimates before training which users are worth LLM reflection. The method consists of a reflection-value surrogate, a gated uncertainty estimator, and a budgeted memory-update policy.

Reflection Value Surrogate

BRAVE treats the base agent as a fixed reflector plus a budgeted allocator. The reflector defines the memory update once LLM feedback is available; the allocator decides whether that feedback is worth its cost. For a user uu, let Wu,LLM+\mathbf{W}^{+}_{u,\mathrm{LLM}} denote post-update mixture memories with LLM local evaluation and global comparison, and let $\mathbf{W}^{+}_{u,\mathrm{rank}} denote mixture memories with only the non-LLM rank-grounded update. We define the latent value of reflection as:

Figure 1. The overall structure of BRAVE.

V(u)=E[R(Wu,LLM+)−R(Wu,rank+)∣Eu],(4)V(u) = \mathbb{E}\left[\mathcal{R}(W^{+}_{u,\mathrm{LLM}})-\mathcal{R}(W^{+}_{u,\mathrm{rank}})\mid\mathcal{E}_u\right], \tag*{(4)}

where R(⋅)\mathcal{R}(\cdot) is a top-KK ranking utility such as Recall or NDCG, and Eu\mathcal{E}_u denotes pre-reflection evidence, including normalized recommender-tool scores, top-ranked lists, and the current tool-mixture memory. Since V(u)V(u) depends on unobserved LLM feedback and cannot be computed before the call, BRAVE estimates it with a recommender-side score computed from saved recommender outputs.

The core design choice is that not all uncertainty should control reflection equally. BRAVE uses two signals with different roles. Ranking fragility measures whether a bounded tool-mixture memory update can change top-ranked items. Cross-tool disagreement measures whether different recommender tools expose different candidate neighborhoods. The former is the primary signal; the latter is used only when the current ranking is fragile enough for reflection to plausibly matter.

Ranking Fragility

For each recommender tool mm, let s~um\tilde{s}^{m}_u be the score vector after per-user min-max normalization over the inspected candidate union, so that scores from different recommenders are comparable. Let s~u,(r)m\tilde{s}^{m}_{u,(r)} denote the rr-th largest entry in s~um\tilde{s}^{m}_u. We compute a top-margin uncertainty: Δum=s~u,(1)m−s~u,(2)m\Delta^{m}_u = \tilde{s}^{m}_{u,(1)}-\tilde{s}^{m}_{u,(2)}, where a small Δum\Delta^{m}_u means that a recommender tool is not confident about its preferred item. A memory update can therefore more easily change the final top-ranked set. We aggregate this signal as:

gfrag(u)=−1M∑m=1MΔum.(5)g_{\mathrm{frag}}(u) = -\frac{1}{M}\sum_{m=1}^{M}\Delta^{m}_u. \tag*{(5)}

A larger gfrag(u)g_{\mathrm{frag}}(u) indicates that the current ranking is more fragile, since the top-ranked items have smaller average score gaps within the recommender pool. This signal serves as a low-cost proxy for aggregate rank-boundary sensitivity: it can be computed before reflection from cached recommender scores, without simulating LLM feedback or enumerating possible mixture-memory updates.

updates. With cached top-KK outputs, its cost is O(M∣U0∣K)O(M|\mathcal{U}_0|K); sorting is required only when candidate lists are not already cached. Section 5.1 motivates why low-margin users offer greater rank-flip opportunity, while Sections 6.3 and 6.4 compare this proxy with aggregate-margin alternatives.

Cross-Tool Disagreement

Ranking fragility alone may miss cases where each recommender tool is individually confident but recommenders expose different item neighborhoods. BRAVE computes a disagreement score over the top-KK set Lum⊆I\mathcal{L}^{m}_u \subseteq\mathcal{I} returned by each recommender tool mm:

gconf(u)=1−2M(M−1)∑a<b∣Lua∩Lub∣∣Lua∪Lub∣.(6)g_{\mathrm{conf}}(u) = 1-\frac{2}{M(M-1)}\sum_{a<b}\frac{\left|\mathcal{L}^{a}_u\cap\mathcal{L}^{b}_u\right|}{\left|\mathcal{L}^{a}_u\cup\mathcal{L}^{b}_u\right|}. \tag*{(6)}

High cross-tool disagreement indicates that personalized recommender weighting may matter because different recommenders expose different candidate regions. However, disagreement is not the same as value: it can arise because one recommender captures the relevant neighborhood, but it can also arise because all recommenders are unreliable for the current user. This motivates using disagreement as a conditional signal rather than as an unconditional additive objective.

Gated Value Estimator

BRAVE combines the two signals with a gated estimator:

gBRAVE(u)=gfrag(u)+η gconf(u)⋅I[gfrag(u)≥qτ],(7)g_{\mathrm{BRAVE}}(u)=g_{\mathrm{frag}}(u)+\eta\,g_{\mathrm{conf}}(u)\cdot\mathbb{I}\left[g_{\mathrm{frag}}(u)\ge q_{\tau}\right], \tag*{(7)}

where qτq_{\tau} is the τ\tau-quantile of {gfrag(u):u∈U0}\{g_{\mathrm{frag}}(u):u\in\mathcal{U}_0\} over the sampled-agent window, η≥0\eta\ge0 controls the strength of the conditional disagreement term, and I[⋅]\mathbb{I}[\cdot] is the indicator function. The gate first identifies users whose rankings are near a decision boundary, then uses cross-tool disagreement to refine their priority. Disagreement can be useful when reflection can plausibly change the learned recommender mixture, but noisy when the ranking is already stable. In the experiments, margin-only and ungated additive scores are allocator ablations of this estimator.

Algorithm 1 BRAVE training-time reflection allocation

Require: User window U0\mathcal{U}_0, recommender tools T\mathcal{T}, budget BB, epochs tt, disagreement weight η\eta, gate quantile τ\tau

  1. Obtain normalized recommender-tool scores {s~um}u,m\{\tilde{s}^{m}_{u}\}_{u,m} and top lists {Lum}u,m\{L^{m}_{u}\}_{u,m}

  2. for each user u∈U0u \in\mathcal{U}_0 do

  3. Compute ranking fragility gfrag(u)g_{\mathrm{frag}}(u) from top-score margins by (5)

  4. Compute cross-tool disagreement gconf(u)g_{\mathrm{conf}}(u) from top-list overlap by (6)

  5. end for

  6. Compute the gate threshold qτq_{\tau} from {gfrag(u):u∈U0}\{g_{\mathrm{frag}}(u):u \in\mathcal{U}_0\}

  7. for each user u∈U0u \in\mathcal{U}_0 do

  8. Compute gBRAVE(u)g_{\mathrm{BRAVE}}(u) by (7)

  9. end for

  10. Select UB=Top⁡⌈B∣U0∣⌉{gBRAVE(u)}\mathcal{U}_B = \operatorname{Top}_{\lceil B|\mathcal{U}_0|\rceil}\{g_{\mathrm{BRAVE}}(u)\}

  11. for e=1,…,te = 1,\ldots,t do

  12. for each minibatch of users do

  13. Apply non-LLM rank-grounded update to all users

  14. Query LLM local evaluation and global comparison only for users in UB\mathcal{U}_B

  15. Update tool-mixture memories with (9)

  16. end for

  17. end for

  18. Keep the original full-catalog scoring and inference-time reranking protocol unchanged

Budgeted Reflection Policy

Given budget BB, BRAVE selects the users with the largest estimated reflection value:

UB=Top⁡⌈B∣U0⌉{gBRAVE(u):u∈U0}.(8)\mathcal{U}_B = \operatorname{Top}_{\lceil B|\mathcal{U}_0\rceil}\{g_{\mathrm{BRAVE}}(u):u \in\mathcal{U}_0\}. \tag*{(8)}

During agent training, LLM local-evaluation and global-comparison updates are applied only to u∈UBu \in\mathcal{U}_B. Users outside UB\mathcal{U}_B still receive the rank-grounded comparison update, so every user can improve its tool-mixture memory without incurring LLM reflection. At epoch ee, the update rule can be written as:

wue+1=wue+γrankrue+πB(u)(αlue+βcue),(9)\mathbf{w}^{e+1}_{u} = \mathbf{w}^{e}_{u} + \gamma_{\mathrm{rank}}\mathbf{r}^{e}_{u} + \pi_B(u)(\alpha\mathbf{l}^{e}_{u}+\beta\mathbf{c}^{e}_{u}), \tag*{(9)}

where wue∈RM\mathbf{w}^{e}_{u} \in\mathbb{R}^{M} is the tool-mixture memory before epoch ee, rue∈RM\mathbf{r}^{e}_{u} \in\mathbb{R}^{M} is the non-LLM rank-comparison signal, and lue,cue∈RM\mathbf{l}^{e}_{u},\mathbf{c}^{e}_{u} \in\mathbb{R}^{M} are the LLM local-evaluation and global-comparison signals. The coefficients γrank\gamma_{\mathrm{rank}}, α\alpha, and β\beta are learning rates for the three update sources. The budget indicator πB(u)=I[u∈UB]\pi_B(u)=\mathbb{I}[u \in\mathcal{U}_B] is fixed after allocation, so LLM reflection affects only selected users while the rank-grounded update remains available to all users.

Cost and Design Properties

Let n=∣U0∣n=|\mathcal{U}_0| be the number of sampled agents, tt be the number of agent-training epochs, and MM be the number of recommender tools. Under a full-reflection agent setting, each user requires one profile-generation call, each reflected user requires M+1M+1 training-time reflection calls per epoch (MM local recommender evaluations and one global comparison), and each user receives one inference-time reranking call. BRAVE therefore uses:

C(B,t,M)=n+(M+1)Bnt+n=(M+1)Bnt+2n.(10)C(B,t,M)=n+(M+1)Bnt+n=(M+1)Bnt+2n. \tag*{(10)}

LLM calls when profile generation and inference reranking are kept for all users. Compared with the uniform full-reflection policy Cfull(t,M)=(M+1)nt+2nC_{\mathrm{full}}(t,M)=(M+1)nt+2n, the relative call ratio is:

C(B,t,M)Cfull(t,M)=(M+1)Bt+2(M+1)t+2.(11)\frac{C(B,t,M)}{C_{\mathrm{full}}(t,M)}=\frac{(M+1)Bt+2}{(M+1)t+2}. \tag*{(11)}

For the three-tool Ultra pool, this reduces to (4Bt+2)/(4t+2)(4Bt+2)/(4t+2); for a four-recommender pool, the reflected training term becomes 5Bnt5Bnt. Thus BRAVE reduces only repeated training reflection while preserving inference capability. It is tool-agnostic, prompt-preserving, and fail-soft: when B=1B=1, it recovers uniform full reflection.

Theoretical Analysis

This section analyzes why allocation can help even when the reflector is unchanged: when a memory update can affect the final ranking, and why uncertainty signals differ as value proxies.

Rank-Flip Opportunity

Let zu(i)=[s~u1(i),…,s~uM(i)]\mathbf{z}_{u}(i)=[\tilde{s}^{1}_{u}(i),\ldots,\tilde{s}^{M}_{u}(i)] stack the normalized scores of item ii across recommender tools, and let:

Su(i)=wuTzu(i),(12)S_{u}(i)=\mathbf{w}_{u}^{\mathsf{T}}\mathbf{z}_{u}(i), \tag*{(12)}

be the aggregated score before reflection. Suppose a reflection step changes the tool-mixture memory by δwu\delta\mathbf{w}_{u}, yielding Su+(i)=(wu+δwu)Tzu(i)S^{+}_{u}(i)=(\mathbf{w}_{u}+\delta\mathbf{w}_{u})^{\mathsf{T}}\mathbf{z}_{u}(i). If normalized scores lie in [0,1][0,1] and ∥δwu∥1≤ρ\lVert\delta\mathbf{w}_{u}\rVert_{1}\leq\rho, then for any two items ii and jj,

∣(Su+(i)−Su+(j))−(Su(i)−Su(j))∣≤2ρ.(13)\left|(S^{+}_{u}(i)-S^{+}_{u}(j))-(S_{u}(i)-S_{u}(j))\right|\leq2\rho. \tag*{(13)}

Therefore, if the current top item i(1)i_{(1)} satisfies

Su(i(1))−Su(j)>2ρ,∀j≠i(1),(14)S_{u}(i_{(1)})-S_{u}(j)>2\rho,\qquad\forall j\neq i_{(1)}, \tag*{(14)}

then no reflection update with ℓ1\ell_{1} size at most ρ\rho can change the top-ranked item. The same perturbation argument applies to the top-KK boundary: if every item inside the current top-KK list is separated from every item outside the list by more than 2ρ2\rho, then a bounded memory update cannot change the top-KK. Thus the relevant opportunity for Recall@K and NDCG@K is not uncertainty in isolation, but uncertainty near a ranking boundary.

Implication. (14) shows that large-margin users are insensitive to bounded memory updates: even good reflection has little room to change their ranking. Low-margin users are closer to a decision boundary, so modest memory correction can change which recommender neighborhood dominates. The argument concerns aggregate score margin under wu\mathbf{w}_{u}; BRAVE uses a cheaper per-recommender margin as a pre-reflection proxy. This proxy is exact only when mixture memories are roughly balanced and recommender-level boundaries align with aggregate boundary, so we compare it with aggregate-margin and top-KK boundary baselines in Section 6.3 and validate it against realized gain in Section 6.4.

Budgeted Selection as Value Approximation

Let V(u)V(u) denote the unobserved reflection value in (4), and suppose the allocator observes a noisy surrogate

g(u)=V(u)+ϵu,(15)g(u)=V(u)+\epsilon_{u}, \tag*{(15)}

where ϵu\epsilon_u is surrogate error. Under a budget BB, the optimal allocation selects the top k=⌈B∣U0∣⌉k = \lceil B|U_0| \rceil users by V(u)V(u), while BRAVE selects the top kk users by g(u)g(u). If Δk\Delta_k is the value gap between the kk-th and (k+1)(k+1)-th users under the true values, then every user whose surrogate error satisfies ∣ϵu∣<Δk/2|\epsilon_u| < \Delta_k/2 preserves its side of the budget boundary. Thus the allocation is stable when the value signal has a clear separation near the budget cutoff.

Implication. The surrogate need not rank all users perfectly; it must avoid replacing high-value boundary users with low-value ones near the budget cutoff. An added uncertainty signal can therefore hurt if it changes cutoff users without improving true reflection value. We examine this boundary-swap effect in Section 6.6.

Why Disagreement Requires Calibration

Cross-tool disagreement is an intuitive uncertainty signal, but it is not identical to reflection value. High disagreement means that recommender tools retrieve different candidate neighborhoods:

gconf(u)↑⟹{Lum}m=1M have low overlap.(16)g_{\mathrm{conf}}(u) \uparrow\quad\Longrightarrow\quad\{L_u^m\}_{m=1}^{M}\ \text{have low overlap}. \tag*{(16)}

This can be beneficial when one recommender exposes the relevant neighborhood and reflection can learn to upweight it. It can be harmful when all recommenders are unreliable, because reflection then reallocates weight among weak alternatives. In the latter case, disagreement measures conflict rather than opportunity.

Calibration principle. Disagreement should therefore be conditional: ranking fragility measures whether memory updates can change the ranking, while cross-tool disagreement measures whether recommender choice may matter. This motivates the gated estimator in (7), whose behavior is tested in Section 6.6.

Experiments

Experimental Setup

Datasets. We evaluate on three full-ranking recommendation benchmarks: Amazon [20], Yelp [45], and Goodreads [30]. They cover product, local-business, and book recommendation scenarios, with different sparsity levels and textual-description regimes. For each dataset, interactions are ordered by timestamp to construct historical user sequences. Following prior sequential recommendation protocols [15, 43], we use leave-one-out evaluation. Table 1 summarizes the dataset statistics and agent-evaluation windows.

DatasetDomainUsersItemsInteractionsWindow size
AmazonProduct1,0541,83935,492160
YelpBusiness4,3033,429258,486160
GoodreadsBook3,2042,167246,034160

Table 1. Statistics of all datasets.

Baselines. We compare BRAVE against four baseline families. The first comprises conventional full-catalog recommenders: LightGCN [11], SASRec [15], SimpleX [19], ENMF [2], DiffRec [33], LRURec [48], FEARec [6], WGCL [5], and NLGC L [39]. The second contains language-based rankers, BM25 [23] and LLMRank [12]. The third contains RAG-style rerankers, which retrieve candidates from each recommender tool and then use an LLM for reranking. The fourth contains agentic recommender-tool learning baseline PRTA [43]. PRTA [43] performs full LLM reflection for every user, whereas BRAVE applies the same reflection operation only to budget-selected users. For both PRTA and BRAVE, Simple denotes the PRTA tool set, which combines a graph collaborative filter, a sequential recommender, and a matching-based collaborative model: LightGCN [11] + SASRec [15] + SimpleX [19]. Ultra denotes the strongest validation-selected tool combination, whose selection process is analyzed in Section 6.8. We use GPT-5.6 Luna [21] as the default LLM controller and study both closed-source controllers (GPT-4o, GPT-5.6 Luna, GPT-5.6 Terra) and open-source controllers (Phi-4-mini, Qwen3-8B, GPT-OSS-20B) in Section 6.9.

Metrics. We follow standard top-KK evaluation practice in recommendation [4, 40]. The best checkpoint is selected by validation Recall@20, and test performance is reported with Recall@10, Recall@20, NDCG@10, and NDCG@20. Reported table values are means over five matched sampled-agent windows.

Implementation details. For fairness, BRAVE follows the PRTA recommender-training and LLM-agent protocol. All recommender tools are trained with a fixed seed on the full training set and rank the entire catalog. The LLM-agent stage is evaluated on five matched 160-user windows per dataset; these windows are shared by PRTA, BRAVE, and allocator variants and affect only per-user memory optimization/evaluation. Simple uses LightGCN [11] + SASRec [15] + SimpleX [19]; Ultra uses the validation-selected NLGC L [39] + WGCL [5] + FEARec [6] pool, analyzed in Section 6.8. We set α=β=0.1\alpha= \beta= 0.1, decay both by 0.50.5 after each epoch, and use keval=krank=20k_{\mathrm{eval}} = k_{\mathrm{rank}} = 20.

GPT-5.6 Luna is the default LLM controller, with temperature 0. Prompts, parsing rules, profile generation, learning rates, and inference reranking are identical between PRTA and BRAVE; the main comparison changes only training-time reflection allocation. Given budget BB, BRAVE selects ⌈Bn⌉\lceil Bn\rceil users before training and keeps the set fixed; Random follows the same fixed-set protocol unless explicitly labeled dynamic. Users outside the set still receive rank-grounded updates. Because metrics are averaged over five 160-user windows, Recall values are reported on the corresponding 1/8001/800 grid. Tool-pool and checkpoint selection use validation interactions only, test items are held out until final evaluation, and qτq_\tau is computed from pre-reflection recommender-tool scores without labels. Unless otherwise specified, B=0.6B = 0.6, η=0.25\eta= 0.25, and τ=0.5\tau= 0.5.

Overall Performance

Table 2 reports the performance of all compared methods on Amazon, Yelp, and Goodreads. We summarize three key observations.

MethodAmazonYelpGoodreads
R@10N@10R@20N@20R@10N@10R@20N@20R@10N@10R@20N@20
LightGCN [11]0.02500.01110.03750.01430.04380.02160.07500.02930.06250.02440.08750.0309
SASRec [15]0.04380.02060.08130.03010.06880.04130.10630.05080.10000.05390.16880.0714
SimpleX [19]0.01880.01470.03130.01790.05000.02320.07500.02990.04380.02270.05000.0243
ENMF [2]0.03750.02610.05000.02920.03130.01850.03130.01850.06250.03190.06880.0336
DiffRec [33]0.04380.01990.07250.02670.08130.04800.11250.05870.06880.03310.10000.0408
LRURec [48]0.04630.02150.10130.03040.06500.02830.10250.03160.05500.02710.11380.0415
FEARec [6]0.05000.01890.10630.03310.06880.03050.10630.03390.05630.02830.11880.0439
WGCL [5]0.05250.02370.10880.03490.08500.04450.11750.05480.06500.02990.10380.0420
NLGCL [39]0.05750.02590.11130.03670.08750.04750.12000.05740.06880.03170.10750.0435
BM25 [23]0.00630.00310.00630.00310.00000.00000.00000.00000.00000.00000.00630.0014
LLMRank [12]0.01250.00890.01250.00940.00000.00000.00630.00160.01250.00440.01250.0063
LightGCNRAG0.02500.01410.04380.01890.03750.01850.07500.03470.01880.01200.02500.0126
SASRecRAG0.06250.03220.07500.03440.06250.02820.07500.03770.01880.01080.01880.0113
SimpleXRAG0.04380.02020.05630.02470.05000.02840.10000.04370.00630.00630.01880.0077
ENMFRAG0.04630.02870.06250.03310.03000.01720.03250.01880.02380.01190.03000.0125
DiffRecRAG0.05250.02310.08130.02970.07630.04580.10750.05610.03130.01480.04630.0189
LRURecRAG0.05130.02290.11250.03370.06130.02710.09630.03020.02630.01240.03630.0153
FEARecRAG0.05630.02150.11130.03400.06630.02910.10130.03310.03000.01360.04130.0185
WGCLRAG0.06130.02780.11130.03600.08130.04310.11250.05320.03130.01430.04630.0205
NLGCLRAG0.06630.02940.11250.03800.08380.04610.11630.05630.03380.01520.04880.0227
PRTASimple [43]0.08750.04160.11250.05420.10250.05540.12000.05830.15880.07510.16750.0809
PRTAUltra [43]0.11130.05280.13000.06120.10750.05830.12630.06140.16750.07910.17500.0834
BRAVESimple (ours)0.09500.04830.11750.05820.10630.06020.12500.06370.16630.08200.18250.0854
BRAVEUltra (ours)0.1138*0.0545*0.1338*0.0631*0.1113*0.0615*0.1300*0.0649*0.1713*0.0832*0.1875*0.0887*

Table 2. Recommendation performance comparison of different models. The best and the second-best performances are marked with bold and underline, respectively. Results are averaged over five matched 160-user agent-evaluation windows. * indicates paired significance under the sampled-window evaluation with pp-value < 0.05.

First, BRAVE achieves the best overall performance across datasets and metrics. Compared with conventional recommenders, language-only rankers, and RAG-style rerankers, BRAVE obtains stronger Recall@10, NDCG@10, Recall@20, and NDCG@20. The pattern indicates that recommendation agents benefit from learning personalized recommender-mixture preferences rather than relying on a single recommender, a text-only ranker, or a one-shot LLM reranking procedure. Among all methods, BRAVEUltra^{\mathrm{Ultra}} achieves the best results on all three datasets, with significant improvements over the strongest baseline where marked.

Second, the improvement is not simply caused by a stronger recommender pool. Under the same recommender combination as PRTA, BRAVESimple consistently outperforms PRTASimple across datasets and metrics. The gains are especially clear on NDCG: for example, Amazon NDCG@20 increases from 0.0542 to 0.0582, while Goodreads NDCG@20 increases from 0.0809 to 0.0854. The gain therefore comes from value-aware reflection allocation, not merely from replacing the underlying recommender tools. With the stronger recommender pool, BRAVEUltra further improves over PRTAUltra on all datasets, showing that budgeted allocation remains useful even when the base recommenders are already competitive.

Third, BRAVE improves ranking quality in addition to hit coverage. The gains are especially clear on NDCG, which is sensitive to the ordering of relevant items near the top of the list. BRAVEUltra improves over PRTAUltra on NDCG@20 by 0.0019, 0.0035, and 0.0053 on Amazon, Yelp, and Goodreads, respectively. Selective reflection therefore appears to refine the learned recommender mixture and improve top-ranked ordering, rather than only expanding the hit set. Overall, full reflection is not necessary for every user: allocating reflection according to estimated user-level

Figure 2. Allocator ablation under the Ultra pool and 60% reflection budget. Bars report NDCG@20.

value can preserve or improve recommendation quality under a constrained reasoning budget.

Ablation Study

We compare BRAVE with cheap-selection baselines under the same budget. Random selects uniformly; HistLen prioritizes shorter histories; Agg-margin uses the top-1/top-2 margin of SuS_u; **Top-K boundary uses the KK-th/(K+1)(K+1)-th aggregate margin; Disagreement-only, Ungated additive, and Fragility-only isolate the components of Equation (7).

Figure 3. Reflection-value calibration. Left: realized NDCG@20 gain by BRAVE-score quintile; dashed/dotted lines mark selected/unselected means. Right: Spearman correlation between each allocation score and realized gain.

Figure 2 shows that the gain is not explained by selecting arbitrary uncertain or cold-start users. Aggregate-boundary baselines are competitive, supporting the rank-flip analysis in Section 5.1, but still trail the gated allocator. Disagreement-only selection is unreliable, while fragility-only is stronger but below BRAVE. The key contrast is ungated additive versus BRAVE: both use the same signals, but ungated addition admits disagreement for every user, whereas BRAVE uses it only after the fragility gate.

Reflection-Value Calibration

Figure 3 tests whether the allocation score estimates the latent value in Equation (4). For each diagnostic-window user, we run matched branches from the same initial mixture memory and recommender outputs: one receives LLM local evaluation and global comparison, while the other receives only the rank-grounded update. Their NDCG@20 difference is the realized reflection gain. The left panel shows monotone gains across BRAVE-score quintiles: low-score users have near-zero or slightly negative gains, while the highest quintile contributes most improvement.

The right panel compares allocation scores against realized gain. History length and raw disagreement correlate weakly; aggregate-margin and top-KK boundary scores are stronger but remain below BRAVE. The oracle gap is expected because BRAVE uses only pre-reflection recommender evidence, yet selected users show substantially larger realized gain than unselected users.

Cost-Quality Analysis

Table 3 contains three Amazon diagnostic blocks: the budget block varies how many users receive LLM reflection, the allocator block compares selection rules under the same 60% budget, and the controller block changes only the LLM controller. Together with Figure 4, the budget block shows a saturated cost-quality frontier. A 20% budget loses useful updates, 40% recovers most of the full-reflection benefit, and 60% gives the best top-ranked quality. Moving to 80% offers little gain, showing that more reflection is not automatically better when extra calls go to stable or unhelpfully conflicting users. The diagnostic rows in Table 3 support the same pattern: random and disagreement-only selection are weaker, fragility-only is competitive, and stronger LLMs improve absolute scores without removing the need for allocation. BRAVE thus keeps the reflection operation unchanged but moves it toward users with higher expected value.

BlockSettingBBCallsR@10N@10R@20N@20
Full reflection referenceFull PRTAUltraPRTA^{\mathrm{Ultra}}100%1.00×0.11130.05280.13000.0612
Budget: vary the reflection budget with BRAVE allocation
BudgetBRAVE20%0.31×0.10630.05060.12630.0588
BudgetBRAVE40%0.49×0.11250.05370.13250.0625
BudgetBRAVE60%0.66×0.11380.05450.13380.0631
BudgetBRAVE80%0.83×0.11380.05410.13250.0630
Allocator: vary the selection rule at 60% budget
AllocatorRandom60%0.66×0.10250.04890.12130.0572
AllocatorDisagreement-only60%0.66×0.10500.04980.12380.0584
AllocatorUngated additive60%0.66×0.10880.05140.12750.0600
AllocatorFragility-only60%0.66×0.11130.05320.13130.0623
AllocatorBRAVE gated60%0.66×0.11380.05450.13380.0631
Controller: vary the LLM controller with BRAVE allocation
Closed ctrl.GPT-4o60%0.66×0.11500.05500.13500.0637
Closed ctrl.GPT-5.6 Luna60%0.66×0.11380.05450.13380.0631
Closed ctrl.GPT-5.6 Terra60%0.66×0.11630.05560.13630.0643
Open ctrl.Phi-4-mini60%0.66×0.11250.05350.13250.0620
Open ctrl.Qwen3-8B60%0.66×0.11250.05380.13250.0624
Open ctrl.GPT-OSS-20B60%0.66×0.11380.05410.13250.0627

Table 3. Amazon cost and diagnostic blocks. The budget block changes BB, the allocator block changes only the selection rule, and the controller block changes only the LLM controller. All rows keep full-catalog ranking and inference reranking.

Figure 4. Budget-quality trade-off. The second x-axis line reports the three-tool call ratio from Equation (11).

Boundary-Swap and Disagreement Calibration

We test the distinction between ranking fragility and cross-tool disagreement. Figure 5 varies η\eta and compares ungated addition with gated estimator in Equation (7). Ungated NDCG@20 decreases as η\eta grows: high disagreement can pull the allocator toward users whose recommender tools disagree even when reflection is unlikely to improve the final ranking. BRAVE is more stable because disagreement affects priority only after the fragility gate. The selected-set overlap curve explains the mechanism. As η\eta increases, disagreement-aware selection diverges from the fragility boundary set; replacing fragile high-value users with high-disagreement low-value users can reduce ranking quality. This boundary-swap effect supports using disagreement to refine, not define, reflection opportunity.

Hyperparameter Sensitivity

Figure 6 shows stable performance near the default. Unlike the Amazon-only diagnostic in Figure 5, this figure averages over Amazon, Yelp, and Goodreads. Very small η\eta approaches fragility-only selection, while very large η\eta overweights disagreement and weakens ranking quality. The gate quantile τ\tau behaves similarly: loose gates admit disagreement too broadly, whereas strict gates suppress useful disagreement. The default η=0.25\eta= 0.25 and τ=0.5\tau= 0.5 are robust operating points rather than narrow optima.

Figure 5. Amazon-only disagreement calibration. Gating keeps cross-tool disagreement conditional on ranking fragility; the gray curve shows selected-set overlap with the fragility-only boundary set.

Figure 6. Sensitivity to η\eta and τ\tau. Values are NDCG@20 averaged over the three datasets under the Ultra pool; dashed lines mark defaults.

Tool-Pool Composition

Figure 7 studies how the Ultra pool is selected. We compare the PRTA Simple pool, the strongest compact two-recommender pool (NLGCL + FEARec), three substitution pools, the selected Ultra pool (NLGCL + WGCL + FEARec), and a four-recommender expansion

Figure 7. Tool-pool composition study. Each cell reports NDCG@20 normalized by the selected Ultra pool on the same dataset. L/S/X/N/W/F/D denote LightGCN/SASRec/SimpleX/NLGCL/WGCL/FEARec/DiffRec; * marks the Ultra pool.

with DiffRec. All variants keep the BRAVE allocator, budget, controller, and inference reranking fixed. Three patterns emerge. Simple is insufficient for the stronger setting, so budgeted reflection still depends on high-quality recommender evidence. NLGCL + FEARec is strong but drops about 3–6% from Ultra, indicating that WGCL adds complementary graph-neighborhood evidence. Substitution is not interchangeable: replacing WGCL with DiffRec, removing FEARec, or removing NLGCL weakens at least one dataset. A fourth recommender tool helps Amazon by about 3% but reduces Yelp and Goodreads by about 1%, suggesting that extra evidence can also introduce noisy disagreement. We therefore use NLGCL + WGCL + FEARec as the default Ultra pool.

Controller Robustness

Figure 8 tests whether the policy is tied to one controller. We compare closed-source controllers (GPT-4o, GPT-5.6 Luna, GPT-5.6 Terra) with open-source controllers (Phi-4-mini, Qwen3-8B, GPT-OSS-20B), keeping prompts, recommender outputs, and allocation fixed. Larger closed-source controllers give modest gains over GPT-5.6 Luna, while open-source controllers incur only slight drops. Thus the allocator benefits from stronger controllers but does not require the strongest available model.

Figure 8. Controller Analysis. Only LLM controller changes.

Conclusion

This paper studies reflection allocation for recommendation agents. Instead of assuming that every user should receive the same amount of LLM reflection, we argue that reflection has user-dependent value and should be spent where it is most likely to change the learned recommender mixture. We propose BRAVE, a budgeted reflection-allocation framework that estimates reflection value from recommender-side self-doubt before making training-time LLM calls. By gating cross-tool disagreement with ranking fragility, BRAVE preserves the base agent’s recommender tools, prompts, and inference reranking while selectively reducing repeated reflection.

Our analysis explains why low-margin users provide greater rank-flip opportunity and why disagreement should be treated as a conditional signal rather than an unconditional objective. Experiments on full-ranking recommendation benchmarks show that BRAVE improves the cost-quality trade-off over uniform reflection and remains effective across tool pools and LLM controllers. These results suggest that scalable recommendation agents require not only stronger reflection modules, but also principled policies for deciding when reflection is worth using.

Ethical Considerations

This work aims to make recommendation agents more efficient by reducing unnecessary LLM reflection while preserving full-catalog ranking and inference reranking. Since BRAVE changes only how training-time reflection is allocated, it does not introduce new user data requirements or additional LLM exposure beyond the base agent protocol. The main potential risk is that a poorly calibrated allocator may assign less reflection to users whose preferences are sparse, niche, or under-represented, which could weaken personalization for those users or indirectly affect item exposure.

This risk can be mitigated through deployment-time auditing and conservative allocation design. In practice, reflection budgets should be evaluated not only by aggregate Recall or NDCG, but also across user-history lengths, item popularity groups, and application-relevant cohorts when such analysis is permitted. The allocator should rely on task-side ranking evidence rather than sensitive attributes, and fallback policies can reserve a small portion of the budget for random or fairness-aware coverage when subgroup robustness is important. Privacy risks can be controlled with the same protections required by the base recommendation agent: minimizing user text sent to the LLM controller, anonymizing user identifiers, retaining only necessary interaction evidence, and applying standard access-control and logging policies. Overall, BRAVE is designed as a cost-aware allocation layer, and its ethical deployment depends on validating that the saved computation does not come at the expense of vulnerable users or responsible exposure.

References

References

  1. [1]Keqin Bao, Jizhi Zhang, Yang Zhang, Wenjie Wang, Fuli Feng, and Xiangnan He. 2023. TALLRec: An Effective and Efficient Tuning Framework to Align Large Language Model with Recommendation. In Proceedings of the 17th ACM Conference on Recommender Systems. 1007–1014. doi:10.1145/3604915.3608857
  2. [2]Chong Chen, Min Zhang, Yongfeng Zhang, Yiqun Liu, and Shaoping Ma. 2020. Efficient Neural Matrix Factorization without Sampling for Recommendation. ACM Transactions on Information Systems 38, 2 (2020), 14:1–14:28. doi:10.1145/3373807DOI
  3. [3]Lingjiao Chen, Matei Zaharia, and James Zou. 2023. FrugalGPT: How to Use Large Language Models While Reducing Cost and Improving Performance. CoRR abs/2305.05176 (2023). doi:10.48550/arXiv.2305.05176DOI
  4. [4]Zheyu Chen, Jinfeng Xu, Hewei Wang, Shuo Yang, Zitong Wan, and Haibo Hu. 2025. Hypercomplex Prompt-aware Multimodal Recommendation. In Proceedings of the 34th ACM International Conference on Information and Knowledge Management. 403–414.arxiv.org/abs/2508.10753
  5. [5]Zheyu Chen, Jinfeng Xu, Yutong Wei, and Ziyue Peng. 2025. Squeeze and Excitation: A Weighted Graph Contrastive Learning for Collaborative Filtering. In Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval. 2769–2773.arxiv.org/abs/2504.04443
  6. [6]Xinyu Du, Huanhuan Yuan, Pengpeng Zhao, Jianfeng Qu, Fuzhen Zhuang, Guan-feng Liu, Yanchi Liu, and Victor S. Sheng. 2023. Frequency Enhanced Hybrid Attention Network for Sequential Recommendation. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval. 78–88. doi:10.1145/3539618.3591689DOI
  7. [7]Yonatan Geifman and Ran El-Yaniv. 2017. Selective Classification for Deep Neural Networks. In Advances in Neural Information Processing Systems, Vol. 30.
  8. [8]Yonatan Geifman and Ran El-Yaniv. 2019. SelectiveNet: A Deep Neural Network with an Integrated Reject Option. In Proceedings of the 36th International Conference on Machine Learning. 2151–2159. https://proceedings.mlr.press/v97/geifman19a.htmlarxiv.org/abs/1901.09192
  9. [9]Shijie Geng, Shuchang Liu, Zuohui Fu, Yingqiang Ge, and Yongfeng Zhang. 2022. Recommendation as Language Processing (RLP): A Unified Pretrain, Personalized Prompt and Predict Paradigm (P5). In Proceedings of the 16th ACM Conference on Recommender Systems. 299–315. doi:10.1145/3523227.3546767
  10. [10]Alex Graves. 2016. Adaptive Computation Time for Recurrent Neural Networks. CoRR abs/1603.08983 (2016). doi:10.48550/arXiv.1603.08983DOI
  11. [11]Xiangnan He, Kuan Deng, Xiang Wang, Yan Li, Yongdong Zhang, and Meng Wang. 2020. LightGCN: Simplifying and Powering Graph Convolution Network for Recommendation. In Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval. 639–648. doi:10.1145/3397271.3401063DOI
  12. [12]Yupeng Hou, Junjie Zhang, Zihan Lin, Hongyu Lu, Ruobing Xie, Julian J. McAuley, and Wayne Xin Zhao. 2024. Large Language Models are Zero-Shot Rankers for Recommender Systems. In Advances in Information Retrieval. 364–381. doi:10.1007/978-3-031-56060-6_24DOI
  13. [13]Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Ejin Bang, Andrea Madotto, and Pascale Fung. 2023. Survey of Hallucination in Natural Language Generation. Comput. Surveys 55, 12 (2023), 248:1–248:38. doi:10.1145/3571730DOI
  14. [14]Chumeng Jiang, Jiayin Wang, Weizhi Ma, Charles L. A. Clarke, Shuai Wang, Chuhan Wu, and Min Zhang. 2025. Beyond Utility: Evaluating LLM as Recommender. In Proceedings of the ACM Web Conference 2025. 3850–3862. doi:10.1145/3696410.3714759DOI
  15. [15]Wang-Cheng Kang and Julian J. McAuley. 2018. Self-Attentive Sequential Recommendation. In IEEE International Conference on Data Mining. 197–206. doi:10.1109/ICDM.2018.00035
  16. [16]Ehud Karpas, Omri Abend, Yonatan Belinkov, Barak Lenz, Opher Lieber, Nir Ratner, Yoav Shoham, Hofit Bata, Yoav Levine, Kevin Leyton-Brown, Dor Muhlgy, Noam Rozen, Erez Schwartz, Gal Shachaf, Shai Shalev-Shwartz, Amnon Shashua, and Moshe Tenenholtz. 2022. MKRL Systems: A Modular, Neuro-Symbolic Architecture that Combines Large Language Models, External Knowledge Sources and Discrete Reasoning. CoRR abs/2205.00445 (2022). doi:10.48550/arXiv.2205.00445DOI
  17. [17]David D. Lewis and William A. Gale. 1994. A Sequential Algorithm for Training Text Classifiers. In Proceedings of the 17th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval. 3–12. doi:10.1007/978-1-4471-2099-5_1
  18. [18]Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, Shashank Gupta, Bodhisattwa Prasad Majumder, Katherine Hermann, Sean Welleck, Amir Yazdanbakhsh, and Peter Clark. 2023. Self-Refine: Iterative Refinement with Self-Feedback. In Advances in Neural Information Processing Systems, Vol. 36.arxiv.org/abs/2303.17651
  19. [19]Kelong Mao, Jieming Zhu, Jinping Wang, Quanyu Dai, Zhenhua Dong, Xi Xiao, and Xiangqiang He. 2021. SimpleX: A Simple and Strong Baseline for Collaborative Filtering. In Proceedings of the 30th ACM International Conference on Information and Knowledge Management. 1243–1252. doi:10.1145/3459637.3482297
  20. [20]Jiamno Ji, Jiacheng Li, and Julian J. McAuley. 2019. Justifying Recommendations using Distantly-Labeled Reviews and Fine-Grained Aspects. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 17th International Joint Conference on Natural Language Processing. 188–197. doi:10.18653/v1/D19-1018
  21. [21]OpenAI. 2026. GPT-5: Frontier Intelligence That Scales with Your Ambition. https://openai.com/index/gpt-5-6/. Accessed: 2026-08-02.DOI
  22. [22]Shishir G. Patil, Tianjun Zhang, Xin Wang, and Joseph E. Gonzalez. 2024. Gorilla: Large Language Model Connected with Massive APIs. In Advances in Neural Information Processing Systems, Vol. 37.arxiv.org/abs/2305.15334
  23. [23]Stephen E. Robertson and Hugo Zaragoza. 2009. The Probabilistic Relevance Framework: BM25 and Beyond. Foundations and Trends in Information Retrieval 3, 4 (2009), 333–389. doi:10.1561/1500000019DOI
  24. [24]Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, Roberta Raileanu, Maria Lomeli, Eric Hambro, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. 2023. Toolformer: Language Models Can Teach Themselves to Use Tools. In Advances in Neural Information Processing Systems.arxiv.org/abs/2302.04761
  25. [25]Burr Settles. 2009. Active Learning Literature Survey. Technical Report 1648. University of Wisconsin-Madison.
  26. [26]H. Sebastian Seung, Manfred Opper, and Haim Sompolinsky. 1992. Query by Committee. In Proceedings of the Fifth Annual Workshop on Computational Learning Theory. 287–294. doi:10.1145/130385.130417
  27. [27]Yongliang Shen, Kaitao Song, Xu Tan, Dongsheng Li, Weiming Lu, and Yueting Zhuang. 2023. HuggingGPT: Solving AI Tasks with ChatGPT and its Friends in Hugging Face. In Advances in Neural Information Processing Systems, Vol. 36.arxiv.org/abs/2303.17580
  28. [28]Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. 2023. Reflexion: Language Agents with Verbal Reinforcement Learning. In Advances in Neural Information Processing Systems, Vol. 36.arxiv.org/abs/2303.11366
  29. [29]Simon Tong and Daphne Koller. 2001. Support Vector Machine Active Learning with Applications to Text Classification. Journal of Machine Learning Research 2 (2001), 45–66.
  30. [30]Mengting Wan and Julian McAuley. 2018. Item Recommendation on Monotonic Behavior Chains. In Proceedings of the 12th ACM Conference on Recommender Systems. 86–94. doi:10.1145/3240323.3240369DOI
  31. [31]Lei Wang, Jingsen Zhang, Hao Yang, Zhi-Yuan Chen, Jiakai Tang, Zeyu Zhang, Xu Chen, Yankai Lin, Hao Sun, Ruihua Song, Xin Zhao, Jun Xu, Zhicheng Dou, Jun Wang, and Ji-Rong Wen. 2025. User Behavior Simulation with Large Language Model-based Agents. ACM Transactions on Information Systems 43, 2 (2025), 55:1–55:37. doi:10.1145/3708985DOI
  32. [32]Peiyi Wang, Lei Li, Liang Chen, Zefan Cai, Dawei Zhu, Binghuai Lin, Yunbo Cao, Lingpeng Kong, Qi Liu, Tianyu Liu, and Zhifang Sui. 2024. Large Language Models are not Fair Evaluators. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 9440–9450. doi:10.18653/v1/2024.acl-long.511
  33. [33]Wenjie Wang, Yiyan Xu, Fuli Feng, Xinyu Lin, Xiangnan He, and Tat-Seng Chua. 2023. Diffusion Recommender Model. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval. 832–841. doi:10.1145/3539618.3591663DOI
  34. [34]Shirley Wu, Shiyu Zhao, Qian Huang, Kexin Huang, Michihiro Yasumaga, Kaidi Cao, Vassilis N. Ioannidis, Karthik Subbian, Jure Leskovec, and James Y. Zou. 2024. AvaTaR: Optimizing LLM Agents for Tool Usage via Contrastive Reasoning. In Advances in Neural Information Processing Systems.arxiv.org/abs/2406.11200
  35. [35]Ji Xin, Raphael Tang, Jaejun Lee, Yaoliang Yu, and Jimmy Lin. 2020. DeeBERT: Dynamic Early Exiting for Accelerating BERT Inference. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. 2246–2251. doi:10.18653/v1/2020.acl-main.204DOI
  36. [36]Jinfeng Xu, Zheyu Chen, Ziyue Peng, Zheng Lin, Wenhao Yuan, Jian Chen, Shujie Li, and Edith Ngai. 2026. Beyond Successor Accuracy: State Retention for Recursive Self-Improvement in Recommendation. arXiv preprint arXiv:2610.07105 (2026).
  37. [37]Jinfeng Xu, Zheyu Chen, Ziyue Peng, Shuo Yang, Jinze Li, Zewei Liu, Shujie Li, Yipeng Du, and Edith C. H. Ngai. 2026. One Graph, Multiple Gains: Single High-Quality Item-Item Graph for Multimodal Recommendation. In Proceedings of the 34th ACM International Conference on Multimedia. 12508–12516.arxiv.org/abs/2607.24607
  38. [38]Jinfeng Xu, Zheyu Chen, Ziyue Peng, Shuo Yang, Jinze Li, Wenhao Yuan, Jian Chen, Shujie Li, and Edith C. H. Ngai. 2026. Neural Tree Collaborative Filtering: Rethinking Graph Collaborative Filtering as Tree Collaborative Filtering with Curvature-Aware Propagation Depth. In Proceedings of the 35th ACM International Conference on Information and Knowledge Management. 7991–7996.arxiv.org/abs/2608.10297
  39. [39]Jinfeng Xu, Zheyu Chen, Shuo Yang, Jinze Li, Hewei Wang, Wei Wang, Xiping Hu, and Edith Ngai. 2025. NLGCL: Naturally Existing Neighbor Layers Graph Contrastive Learning for Recommendation. In Proceedings of the Nineteenth ACM Conference on Recommender Systems. 319–329.arxiv.org/abs/2507.07522
  40. [40]Jinfeng Xu, Zheyu Chen, Shuo Yang, Jinze Li, Wei Wang, Xiping Hu, Steven Hoi, and Edith Ngai. 2026. A survey on multimodal recommender systems: Recent advances and future directions. IEEE Transactions on Multimedia (2026).arxiv.org/abs/2502.15711
  41. [41]Wenda Xu, Guanglei Zhu, Xuandong Zhao, Liangming Pan, Lei Li, and William Yang Wang. 2024. Pride and Prejudice: LLM Amplifies Self-Bias in Self-Refinement. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 15474–15492. doi:10.18653/v1/2024.acl-long.826DOI
  42. [42]Mingdai Yang, Nurendra Choudhary, Jiangshu Du, Edward W. Huang, Philip Yu, Karthik Subbian, and Danai Koutra. 2026. AgentDR: Dynamic Recommendation with Implicit Item-Item Relations via LLM-based Agents. In Proceedings of the ACM Web Conference 2026. 6230–6241. doi:10.1145/3774904.3792304DOI
  43. [43]Mingdai Yang, Zhiwei Liu, Weizhi Zhang, Yibo Wang, Hao Peng, and Philip S. Yu. 2026. Personalized Recommendation Tool Learning via Autonomous Language Agents. In Proceedings of the 20th ACM Conference on Recommender Systems. 1–6. doi:10.1145/3773078.3813794arxiv.org/abs/2607.19739
  44. [44]Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik R. Narasimhan, and Yuan Cao. 2023. ReAct: Synergizing Reasoning and Acting in Language Models. In International Conference on Learning Representations. https://openreview.net/forum?id=WE_vluYUL-Xarxiv.org/abs/2210.03629
  45. [45]Yelp. 2024. Yelp Open Dataset. https://www.yelp.com/dataset
  46. [46]Qihang Yu, Kairui Fu, Zheyu Zhang, Xuanhiu Wu, Chen Lin, Feng Wei, Bo Zheng, and Fei Wu. 2026. ThinkRec: Thinking-based Recommendation via LLM. In Proceedings of the ACM Web Conference 2026. 5698–5709. doi:10.1145/3774904.3792070DOI
  47. [47]Murong Yue, Jie Zhao, Min Zhang, Liang Du, and Ziyu Yao. 2024. Large Language Model Cascades with Mixture of Thought Representations for Cost-Efficient Reasoning. In International Conference on Learning Representations.arxiv.org/abs/2310.03094
  48. [48]Zhenrui Yue, Yueqi Wang, Zhankui He, Huimin Zeng, Julian McAuley, and Dong Wang. 2024. Linear recurrent units for sequential recommendation. In Proceedings of the 17th ACM international conference on web search and data mining. 930–938.
  49. [49]An Zhang, Yuxin Chen, Leheng Sheng, Xiang Wang, and Tat-Seng Chua. 2024. On Generative Agents in Recommendation. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval. 1807–1817. doi:10.1145/3626772.3657844DOI
  50. [50]Junjie Zhang, Yupeng Hou, Ruobing Xie, Wenqi Sun, Julian J. McAuley, Wayne Xin Zhao, Leyu Lin, and Ji-Rong Wen. 2024. AgentCF: Collaborative Learning with Autonomous Language Agents for Recommender Systems. In Proceedings of the ACM Web Conference 2024. 3679–3689. doi:10.1145/3589334.3645537DOI
  51. [51]Yuyue Zhao, Jiancan Wu, Xiang Wang, Wei Tang, Dingxian Wang, and Maarten de Rijke. 2024. Let Me Do It For You: Towards LLM Empowered Recommendation via Tool Learning. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval. 1796–1806. doi:10.1145/3626772.3657828
  52. [52]Andy Zhou, Kai Yan, Michal Shlapentokh-Rothman, Haohan Wang, and Yu-Xiong Wang. 2024. Language Agent Tree Search Unifies Reasoning, Acting, and Planning in Language Models. In Proceedings of the 41st International Conference on Machine Learning. 62138–62160.arxiv.org/abs/2310.04406

Paper details

vlnc:W_kwdr9czw
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents