betaFeedback

WorksComputer Science

Learning to Adjudicate: Verifier-Guided Multi-Agent Reasoning for E-Commerce Relevance

Abstract

Relevance modeling is fundamental to e-commerce search, determining how well retrieved products satisfy the needs expressed in a user's query. Accurate judgments require interpreting user intent, comparing product attributes, and applying business-specific matching criteria. Current methods based on large language models (LLMs) leverage contextual understanding and reasoning, focusing on incorporating domain-specific context and enhancing reasoning performance. Despite these advances, difficult query-product pairs remain challenging when evidence from different perspectives supports conflicting conclusions. A single reasoning pass may fail to reconcile these conflicts and ensure that decisive evidence guides the final judgment. To address this challenge, we propose Collaborative Adjudication with Debate and Replay Experience (CADRE), a framework for evidence-based adjudication in e-commerce relevance assessment. CADRE organizes three expert agents sharing a frozen backbone to analyze product attributes, user intent, and business rules in separate contexts, while a trainable Verifier assesses their supporting evidence against the applicable matching criteria to determine relevance. Under expert disagreement, a Verifier-guided debate elicits targeted clarification and revision of disputed interpretations before final adjudication. We optimize the Verifier through multi-turn group relative policy optimization and adapt successful-experience replay so that trajectories yielding correct terminal labels can be reused alongside fresh rollouts in subsequent updates. For efficient online serving, we use CADRE as an offline teacher to distill its relevance judgments into a lightweight student model. Offline evaluations and online A/B tests on an e-commerce search advertising platform demonstrate that CADRE improves relevance classification and key business and relevance satisfaction metrics.

CCS Concepts

• Information systems → Relevance assessment; Similarity measures; Language models; Sponsored search advertising; • Computing methodologies → Natural language processing.

Keywords

Relevance Modeling, Large Language Model, Multi-Agent Debate, Reinforcement Learning, Experience Replay, E-Commerce Search

ACM Reference Format:

Shi Shu, Hao Chen, Baopu Qiu, Chao Wei, Weiru Zhang, Xiu li, and Xiaoyi Zeng. 2018. Learning to Adjudicate: Verifier-Guided Multi-Agent Reasoning for E-Commerce Relevance. In Woodstock ’18: ACM Symposium on Neural Gaze Detection, June 03–05, 2018, Woodstock, NY. ACM, New York, NY, USA, 13 pages. https://doi.org/XXXXXXXX.XXXXXXX

Introduction

Relevance modeling is a cornerstone of e-commerce search, as accurate query–product matching is essential to effective product discovery and a satisfying user experience. Earlier approaches predominantly rely on BERT-style encoders to learn semantic interactions between queries and products from annotated data [6, 24]. More recently, large language models (LLMs) have been adopted to bring broader knowledge and stronger contextual reasoning to relevance judgments through prompting and fine-tuning [14]. Chain-of-Thought (CoT) approaches further express these judgments as explicit reasoning steps that connect query understanding, product analysis, and the application of matching criteria [7, 20, 26].

However, difficult query–product pairs often involve conflicting evidence from user intent, product attributes, and business rules, requiring the model to determine which evidence is decisive under the applicable matching criteria. Preference optimization methods such as DPO [16] and KTO [8], together with RL methods such as GRPO [7], refine relevance reasoning through feedback on generated judgments, but resolving these conflicts remains challenging. As illustrated in Figure 1, an intent-based interpretation may override a correctly identified attribute mismatch within the same reasoning context. Even when different views are generated separately, their initial reports may leave unresolved why a particular attribute should be decisive or whether a matching rule applies, requiring further examination before a final judgment can be justified. Learning from difficult cases also poses a training challenge, since trajectories yielding correct relevance labels may be scarce within a limited rollout budget and, once discovered, may not be effectively reused in subsequent updates.

Figure 1. An illustrative case of conflict within a single CoT, where an intent-based interpretation overrides an identified attribute mismatch and yields an incorrect relevance judgment.

To address these challenges, we propose Collaborative Adjudication with Debate and Replay Experience (CADRE), a framework for evidence-based adjudication in e-commerce relevance assessment. The framework comprises three complementary components.

  1. Multi-agent analysis with a dedicated Verifier assigns product analysis, user intent understanding, and business-rule interpretation to agents with separate contexts over a shared frozen backbone. A trainable Verifier compares their analyses and relevance judgments to assess the supporting evidence and determine which matching criteria apply to the query–product pair.

  1. Verifier-guided debate selectively extends adjudication based on disagreement among the experts’ initial labels. For selected inputs, the Verifier asks focused questions to clarify disputed evidence and matching criteria, then integrates the revised expert reports into a final judgment.

  1. Successful-experience replay builds on Reinforcement Learning with Experience Replay (RLEP) [25]. We adapt this approach to multi-turn Verifier training by retaining trajectories with correct terminal labels and their original interaction contexts, and replaying them alongside fresh rollouts in subsequent updates.

For efficient online deployment, we distill CADRE’s reasoning into a lightweight student that directly scores query–product pairs. On an in-house multilingual relevance dataset, CADRE improves accuracy and Macro-F1 by 3.73 and 13.64 percentage points, respectively, over LLM-Replay-GRPO with additional context. In a 7-day online A/B test, the distilled student achieves relative gains of 1.69% in revenue per mille (RPM) and 0.76% in click-through rate (CTR), alongside an absolute increase of 0.3 percentage points in relevance satisfaction score (RS), over the existing production baseline.

Related Works

Early relevance models rely on lexical matching, with TF-IDF [17] and BM25 [18] scoring query–document pairs through term statistics. Although computationally efficient, these methods have limited ability to capture semantic associations beyond lexical overlap. Neural approaches such as DSSM [9] learn dense representations that bring semantically related queries and documents closer in a shared space. Pre-trained encoders such as BERT [6] further introduce contextual representations for semantic matching, while ColBERT [12] and Poly-encoders [10] explore different trade-offs between interaction quality and computational cost. These advances improve semantic matching, but complex product requirements and implicit shopping intent motivate methods that can incorporate broader knowledge and reason about the basis of a relevance judgment.

Recent studies therefore explore LLMs for relevance assessment, drawing on their broad knowledge and contextual reasoning capabilities [14, 21]. Subsequent work further grounds these capabilities in domain-specific knowledge and matching criteria to address the requirements of e-commerce relevance. An early effort in this direction, behavior-augmented prompting [5], supplies behavioral context for relevance prediction. More recent work further explores domain knowledge and matching criteria, with MPCoT [15] incorporating business rules alongside intent and attribute analysis, while K-CARE [4] addresses missing domain context through behavioral anchors and expert-curated prototypes. Together, these works show the value of supplying domain context that may be absent from, or insufficiently elicited by, the model’s parametric knowledge.

Another line of research seeks to enhance reasoning quality and alignment with matching rubrics through more informative supervision. LREF [20] and ADORE [8] use preference optimization to calibrate relevance judgments, while SHE [11], TaoSR-AGRL [23], and LORE [13] develop stepwise rewards, rule-aware feedback, and progressive reasoning supervision, respectively. SHE organizes relevance assessment into five stages, and TaoSR-AGRL follows a predefined category–attribute–relevance pattern. Prior work also addresses the reliability of relevance reasoning and its alignment with final predictions. TaoSR1 [7] combines preference optimization with RL to mitigate reasoning errors and inconsistencies between reasoning and final answers. More recently, Graph-GRPO [2] models relevance reasoning as a dependency graph and propagates reward signals across dependent reasoning components, enabling fine-grained credit assignment beyond uniform sequence-level feedback. These methods strengthen the supervision of relevance reasoning, while leaving room to investigate how disagreements among complementary analyses can guide targeted re-examination of evidence and matching criteria before final adjudication.

Effective RL training also depends on obtaining and reusing informative trajectories within a limited rollout budget. In relevance modeling, TaoSR1 and SHE use difficulty-based filtering to improve the informativeness of training groups [7, 11], while TaoSR-AGRL introduces targeted ground-truth guidance to regenerate trajectories for low-reward cases [23]. These strategies improve the utility of newly sampled trajectories, while reusing previously discovered successes offers a complementary direction. In the broader context of LLM reasoning, RLEP [25] collects successful trajectories from an RL-trained policy and combines them with fresh rollouts during a subsequent training run. These studies highlight the complementary roles of discovering useful trajectories and retaining successful experiences, providing a basis for investigating experience reuse in multi-turn relevance assessment.

Knowledge distillation provides a practical route for deploying LLM-based relevance models under online latency constraints.

Recent work transfers supervision from reasoning LLMs to lightweight BERT models [1, 22, 26], while MPCoT [15] introduces latent reasoning distillation and Graph-GRPO [2] uses graph-node-based multi-head distillation to supervise compact students with structured reasoning signals. These studies establish knowledge distillation as a practical connection between reasoning-intensive offline teachers and efficient online relevance models.

Methods

Relevance Task Definition

As shown in Figure 2, we propose Collaborative Adjudication with Debate and Replay Experience (CADRE), a framework that combines evidence adjudication with the reuse of successful reasoning attempts. We first define the e-commerce relevance task with a multi-agent setting, then introduce the multi-agent analysis generation protocol and detail how we optimize the Verifier via multi-turn agentic GRPO and experience replay. Finally, we distill CADRE’s reasoning into a lightweight student for efficient online serving.

Figure 2. Overview of CADRE. (A) Frozen expert agents analyze query–product pairs with complementary context, while the Verifier integrates their reports and selectively initiates debate when their initial labels disagree. (B) Verifier training combines SFT initialization, multi-turn GRPO, successful-experience collection, and Replay-GRPO to reuse scarce successful trajectories and reduce repeated exploration.

In e-commerce search, relevance modeling assesses how well a product satisfies the need expressed in a user’s query under the applicable matching criteria. Given a query qq and a product title pp, we denote the input pair by x=(q,p)x=(q,p) and formulate relevance prediction as multi-class classification over a fine-grained label space Y\mathcal{Y} containing seven relevance categories, detailed in § 4.1. We write y∗∈Yy^* \in\mathcal{Y} for the gold label and y^\hat{y} for a prediction.

We extend the standard single-model formulation to three expert agents indexed by

K={Ana., Int., Rule}.(1)\mathcal{K} = \{\mathrm{Ana.},\ \mathrm{Int.},\ \mathrm{Rule}\}. \tag*{(1)}

Each expert receives the query–product pair and a role-specific input context ckc_k, and produces an analysis rkr_k and a predicted label y^k∈Y\hat{y}_k \in\mathcal{Y}:

ak:(x,ck)⟼(rk,y^k).(2)a_k : (x,c_k) \longmapsto(r_k,\hat{y}_k). \tag*{(2)}

We refer to the analysis and predicted label (rk,y^k)(r_k,\hat{y}_k) collectively as an expert report. Here, ckc_k denotes the context supplied to an expert before analysis, while rkr_k contains its reasoning and intermediate findings derived from (x,ck)(x,c_k). The contexts provide role-specific criteria or historical interaction evidence, allowing the experts to examine complementary aspects of the same query–product pair within a shared label space. The context construction and analysis procedure for each role are detailed in § 3.2.1.

All experts share a frozen backbone model MM, with specialization induced by their role prompts and input contexts. The Verifier agent VθV_\theta, with trainable parameters θ\theta, receives the query–product pair and the expert reports to produce an adjudication rationale RR and a relevance label y^\hat{y}:

Vθ:(x,{(rk,y^k)}k∈K)⟼(R,y^).(3)V_\theta: \left(x,\{(r_k,\hat{y}_k)\}_{k\in\mathcal{K}}\right) \longmapsto(R,\hat{y}). \tag*{(3)}

Verifier-Guided Multi-Agent Reasoning

Our framework organizes relevance assessment around independent expert analyses and a dedicated Verifier. The experts examine complementary evidence in separate contexts, while the Verifier evaluates their reports to determine the relevance judgment. For inputs selected by a predefined disagreement criterion, the Verifier initiates a focused debate to obtain clarification and revised analyses before making the final decision.

Multi-Agent Analysis and Adjudication

Prior work has leveraged diverse sources of contextual information to support e-commerce relevance judgments [4, 15]. We build on this idea by assigning complementary analytical roles to three expert agents, each analyzing the query-product pair x=(q,p)x=(q,p) in a separate context over a shared frozen backbone. Role-specific prompts and inputs induce specialization while holding model capacity fixed. Each expert develops its analysis without observing the other experts’ reports, leaving cross-perspective comparison to the Verifier. We freeze the expert backbone to reduce optimization overhead and retain its general-purpose response policy, while concentrating task-specific learning on the Verifier’s evidence adjudication and targeted questioning.

The Analyst Agent receives cAnac_{\mathrm{Ana}}, comprising structured comparison criteria and an attribute schema adapted from Zhao et al. [26], covering fourteen dimensions including category, brand, and material. It first generates attribute-level matching results eAnae_{\mathrm{Ana}} and then reasons over them to determine relevance. Each attribute is assessed as a match, mismatch, or unspecified; a mismatch requires explicitly conflicting values on both sides, while an attribute specified on only one side is marked as unspecified and supports neither agreement nor conflict. These comparison states describe attribute-level evidence, while the final relevance judgment also depends on the applicable matching criteria. The Ruler Agent receives cRulec_{\mathrm{Rule}}, containing the original text of the platform’s relevance guidelines concerning categories, brands, model designations, attributes, main-product versus accessory relations, and multi-SKU listings, including precedence rules for overlapping mismatch conditions. It first identifies the applicable clauses eRulee_{\mathrm{Rule}} and then derives its judgment by reasoning about the pair under those clauses. As the only expert receiving historical behavioral context, the Intent Agent receives cInt=eIntc_{\mathrm{Int}}=e_{\mathrm{Int}}, comprising the top-5 historical search terms associated with the candidate product and the top-5 product titles clicked in historical searches for the current query, to infer the shopping need expressed by the query and assess the candidate product’s functional substitutability.

Each expert k∈Kk \in\mathcal{K} produces a first-round report containing its analysis rk(1)r_k^{(1)} and a predicted label y^k(1)\hat{y}_k^{(1)}. The trainable Verifier Agent receives the query–product pair and all three reports, examines the evidence supporting each interpretation, and assesses the applicability of the matching criteria invoked by the experts. It then produces an initial adjudication rationale R(1)R^{(1)} and a label y^V(1)\hat{y}_V^{(1)}. The Verifier retains independent decision authority, so its judgment may differ from the experts’ shared label even when all three agree. The Verifier can thus retain useful observations from expert reports even when all three initial predictions are incorrect.

Selective Verifier-Guided Debate

To control the additional cost of expert interaction, we initiate debate if and only if the three experts’ first-round relevance labels are not unanimous, covering both a two-to-one label split and three distinct labels. When all three experts agree, the Verifier’s initial adjudication serves as the final prediction, even if it differs from their shared label. For inputs selected for debate, the Verifier examines the conflicting reports and generates one focused question for each expert to clarify the evidence or matching criteria underlying its judgment.

Each expert receives its own question together with its first-round report and returns a revised analysis and label (rk(2),y^k(2))(r_k^{(2)},\hat{y}_k^{(2)}). The questions direct attention to disputed interpretations, such as whether an attribute mismatch is decisive or whether a business rule applies to the current pair. Experts respond independently, without observing one another’s reports or communicating directly, while the Verifier coordinates the examination of disagreements across their analyses. After this single round of questions and responses, the Verifier integrates the initial reports and subsequent feedback to produce the final adjudication rationale R(2)R^{(2)} and label y^V(2)\hat{y}_V^{(2)}. This bounded protocol allows disputed evidence to be re-examined while concentrating additional interaction on cases with expert disagreement.

Verifier Training with Experience Replay

We train the Verifier to perform the adjudication and questioning described in § 3.2, while keeping the shared backbone Qwen3-14B of the three expert agents frozen. As illustrated in Figure 2, training proceeds through supervised initialization, on-policy optimization, experience collection, and optimization with replay. Supervised fine-tuning (SFT) on multi-turn interaction data produces an initial checkpoint θSFT\theta_{\mathrm{SFT}}, which is subsequently optimized with group relative policy optimization (GRPO). The resulting checkpoint θGRPO\theta_{\mathrm{GRPO}} is used to collect successful interaction trajectories. Replay-GRPO then restarts from θSFT\theta_{\mathrm{SFT}} and combines fresh rollouts with the collected experiences.

The original dataset provides relevance labels but does not contain annotated adjudications or questions for Verifier training. Using only the training split, we construct synthetic supervision by combining expert reports generated by the frozen Qwen3-14B backbone with Verifier outputs generated through the Qwen3.8-Max API. The expert models produce their initial reports and subsequent revisions, while Qwen3.8-Max generates the Verifier’s adjudications and expert-specific questions, with the annotated relevance label supplied as additional guidance in its prompt. During SFT, this label-bearing guidance is removed from the Verifier’s input, while the expert reports remain part of the interaction context and the generated Verifier responses serve as supervision targets. The resulting checkpoint θSFT\theta_{\mathrm{SFT}} initializes subsequent multi-turn policy optimization.

Multi-Turn Policy Optimization

We formulate Verifier training as a single-agent reinforcement learning problem in which the frozen experts and message templates constitute the interaction environment. The Verifier generates adjudications and focused questions, while the environment supplies instructions and expert responses. For an input that satisfies the predefined debate-triggering criterion, we represent the complete interaction as τ=(u1,o1,u2,o2,u3,o3)\tau= (u_1, o_1, u_2, o_2, u_3, o_3), where u1u_1 contains the query–product pair and the initial expert reports, u2u_2 provides the instruction for generating focused questions, and u3u_3 contains the revised expert reports. The Verifier outputs o1o_1, o2o_2, and o3o_3 correspond to the initial adjudication, the three expert-specific questions, and the final adjudication, respectively. When the input does not satisfy the debate-triggering criterion, the trajectory consists only of (u1,o1)(u_1, o_1), and the Verifier’s initial adjudication serves as the final output. In either branch, the Verifier determines its judgment independently of whether the expert labels agree.

We assign a policy-loss mask m=1m = 1 to Verifier-generated tokens and m=0m = 0 to environment messages and expert responses, which remain part of the conditioning context. Table 4 in Appendix § 7 details the interaction steps and their masks. Starting from θSFT\theta_{\mathrm{SFT}}, we sample a group of complete trajectories for each training input and assign each trajectory a scalar reward r(τ)r(\tau), defined in § 3.5. GRPO computes group-relative advantages from these rewards and applies the clipped policy objective to Verifier-generated tokens under their corresponding interaction histories.

StepInteractionmm
1Verifier: Initial adjudication from the three expert reports.1
2Environment: Question instruction with the initial judgment and expert reports.0
3Verifier: One focused question per expert.1
4Experts: Each revises its report using its own question and initial report.0
5Environment: Append revised reports to the Verifier’s context.0
6Verifier: Final adjudication from the initial and revised reports.1

Table 4. Debate interactions and policy-loss masks.

Reward Design

We assign each complete Verifier trajectory τ\tau a scalar reward combining adjudication quality and debate quality:

r(τ)=radj+Gatedebate rshape,(4)r(\tau) = r_{\mathrm{adj}} + \mathrm{Gate}_{\mathrm{debate}}\, r_{\mathrm{shape}}, \tag*{(4)}

where Gatedebate=1\mathrm{Gate}_{\mathrm{debate}} = 1 when the three experts’ first-round labels are not unanimous and is zero otherwise.

Adjudication quality. The reward radjr_{\mathrm{adj}} combines format validity rfmtr_{\mathrm{fmt}}, hierarchical label accuracy rresr_{\mathrm{res}}, and reasoning checks rreasonr_{\mathrm{reason}}. The terminal prediction is the Verifier’s initial judgment without debate and its final judgment after debate. Outcome supervision assigns full credit to an exact label match and partial credit to a correct relevant/irrelevant superclass, while reasoning checks assess rationale–label consistency and adjudication completeness.

Outcome and reasoning rewards are enabled only when the terminal label is valid.

Debate quality. The reward rshaper_{\mathrm{shape}} evaluates the initial judgment, question quality, and expert revisions. It rewards a correct initial decision and uses question length and references to input tokens as heuristics for question specificity, while penalizing malformed questions and disclosure of the Verifier’s own judgment. Expert revisions receive greater credit for correcting an error than for preserving a correct prediction, with penalties for regressions. These revisions are evaluated against the gold label, allowing experts to disagree with the Verifier’s initial judgment. Detailed scoring rules and coefficients are provided in Appendix § 9.§ 1.

Successful-Experience Replay

Although the reward provides supervision for adjudication and questioning, a limited rollout budget may yield no successful trajectory for a difficult input, while occasional successes are not automatically retained for subsequent updates. Following RLEP [25], we adapt successful-experience replay to the Verifier’s interaction with frozen experts through a per-input experience pool B(x)\mathcal{B}(x). The pool retains trajectories with correct terminal labels together with their original interaction contexts, allowing successful adjudication and questioning attempts to be reused alongside fresh rollouts during Verifier training.

After on-policy GRPO, the checkpoint θGRPO\theta_{\mathrm{GRPO}} generates 16 complete trajectories per input using the same Verifier–expert interaction protocol as in training. We define success by a parseable terminal prediction that exactly matches the gold label, so every retained trajectory satisfies

τ∈B(x)⟹rres(τ)=1∧rfmt(τ)=1.(5)\tau\in\mathcal{B}(x) \Longrightarrow r_{\mathrm{res}}(\tau) = 1 \land r_{\mathrm{fmt}}(\tau) = 1. \tag*{(5)}

Among the trajectories satisfying these conditions, we retain at most eight per input, ranked by their total reward. The auxiliary reward terms therefore determine the ranking of successful candidates, while admission requires an exactly correct terminal label. Each stored trajectory preserves its original token IDs and policy-loss mask, maintaining the alignment between Verifier outputs and interleaved environment messages.

Replay-GRPO restarts from θSFT\theta_{\mathrm{SFT}}, the same initialization used for on-policy GRPO. For each input, we combine GG fresh rollouts with MM trajectories sampled from B(x)\mathcal{B}(x), with empty pools contributing none. Excess replay rows are randomly subsampled, while batch deficits are filled by resampling with replacement within the original input groups; if the budget cannot be met, the update uses fresh-only GRPO. All trajectories are scored through a unified reward pipeline, and group-relative advantages are computed over the actual group:

Ai=ri−mean⁡(r1,…,rnx)std⁡(r1,…,rnx)+ϵ,i=1,…,nx.(6)A_i = \frac{r_i - \operatorname{mean}(r_1,\ldots,r_{n_x})}{\operatorname{std}(r_1,\ldots,r_{n_x}) + \epsilon}, \qquad i = 1,\ldots,n_x. \tag*{(6)}

Here, nxn_x denotes the actual number of trajectories for input xx in the update, and rir_i is the reward of the ii-th trajectory. We set ϵ=10−6\epsilon= 10^{-6} to prevent division by zero, yielding zero advantages for equal-reward groups. Fresh and replayed trajectories enter the same clipped GRPO objective, with the policy loss applied only to Verifier-generated tokens. Old log-probabilities are recomputed using the update-start actor, without importance correction for historical generating policies. Reusing scarce successful trajectories allows each discovered success to contribute to multiple updates, reducing reliance on repeated exploration to recover successful interactions.

Knowledge Distillation

CADRE adjudicates each pair through multiple expert generations and a Verifier, which is too costly for low-latency online serving. We therefore train a lightweight student gϕg_{\phi}: an mBERT encoder with a seven-way classification head that maps x=(q,p)x=(q,p) directly to y^∈Y\hat{y}\in\mathcal{Y}. Let pϕ(⋅∣x)p_{\phi}(\cdot\mid x) denote the student’s predicted distribution over Y\mathcal{Y}. Gold relevance labels supervise classification, while CADRE serves as an offline teacher providing CoTs as auxiliary supervision. Following prior work on reasoning distillation [22], we transfer this reasoning signal into the student’s representation rather than generating it at inference time.

Let cc denote the teacher’s CoT for input xx. We use a frozen BGE-M3 encoder [3] as fembf_{\mathrm{emb}} to obtain a dense target representation ecot=femb(c)e_{\mathrm{cot}}=f_{\mathrm{emb}}(c). The student encodes xx with mBERT and takes its h[CLS]h_{\mathrm{[CLS]}} representation, which a projection layer maps to er=gproj(x)e_r=g_{\mathrm{proj}}(x). In addition to the standard classification loss LCE(pϕ(⋅∣x),y∗)\mathcal{L}_{\mathrm{CE}}(p_{\phi}(\cdot\mid x),y^*), we align the student’s representation with the teacher’s reasoning through a cosine distance:

L=LCE(pϕ(⋅∣x),y∗)+λLcos(er,ecot),Lcos=1−er⋅ecot∥er∥∥ecot∥(7)\mathcal{L}=\mathcal{L}_{\mathrm{CE}}(p_{\phi}(\cdot\mid x),y^*)+\lambda\mathcal{L}_{\mathrm{cos}}(e_r,e_{\mathrm{cot}}),\qquad\mathcal{L}_{\mathrm{cos}}=1-\frac{e_r\cdot e_{\mathrm{cot}}}{\lVert e_r\rVert\lVert e_{\mathrm{cot}}\rVert} \tag*{(7)}

where λ\lambda balances reasoning distillation against gold-label supervision. The CoT branch and fembf_{\mathrm{emb}} are used only during training; at inference the student scores xx alone, with no multi-agent interaction and no reasoning generation.

Experiments

Dataset

We evaluate CADRE on an in-house e-commerce dataset with 78,997 training and 18,710 test query–product pairs. Each pair is labeled as Strongly Relevant, Substitute, or one of five irrelevant categories: Brand Mismatch, Category Mismatch, Model Mismatch, Attribute Mismatch, and Main Product vs. Accessory. When multiple mismatch conditions apply to the same query-product pair, annotators assign a single label, following the most severe applicable mismatch. To mitigate label imbalance, we use label-specific sampling quotas targeting an approximately 1:1 ratio between relevant and irrelevant pairs and roughly balanced representation of the five irrelevant categories. The training and test splits have similar label proportions. Detailed dataset statistics are provided in Appendix § 8.

Baselines and Evaluation Metrics

Baselines

All teacher-stage methods in Table 1 use Qwen3-14B as the backbone. Additional context C=(cAna,cInt,cRule)C=(c_{\mathrm{Ana}},c_{\mathrm{Int}},c_{\mathrm{Rule}}) comprises the attribute schema and comparison criteria, historical search and click context associated with the query and product, and the full text of the platform relevance guidelines, respectively. We evaluate single-model baselines with and without CC. The context-augmented variants directly process the same source information distributed across the three experts. LLM-Base denotes the off-the-shelf checkpoint without task-specific training. LLM-SFT applies supervised fine-tuning using relevance reasoning and predictions generated by Qwen3-8-Max under cold-label guidance. This follows the same supervision-generation approach as Verifier-SFT, with the prompt format adapted to single-model relevance prediction. LLM-GRPO and LLM-Replay-GRPO apply GRPO and Replay-GRPO, respectively, starting from the supervised LLM-SFT checkpoint.

MethodSeven-Class ClassificationIrrelevant Classification
AccuracyMacro-F1Weighted-F1PrecisionRecallF1
Teacher models (Qwen3-14B)
Vanilla single model
LLM-Base0.54080.47990.60920.75110.73180.7414
LLM-SFT0.72780.63050.72470.85070.79160.8201
LLM-GRPO0.77490.61730.77200.85620.89240.8739
LLM-Replay-GRPO0.79060.63690.78770.86150.90070.8806
Single model with additional context
LLM-Base0.64310.49640.64260.79860.72000.7573
LLM-SFT0.73500.63800.73200.86380.79450.8277
LLM-GRPO0.78230.62680.78050.85930.89510.8768
LLM-Replay-GRPO0.79610.64020.79150.8629<u>0.8994</u>0.8808
Multi-agent with additional context; without debate
Verifier-Base0.63690.54020.66850.74490.89740.8141
Expert Oracle0.8051–––––
Verifier-SFT0.80200.73510.80450.87050.88680.8786
Verifier-GRPO0.82390.76430.82300.89390.88610.8900
Verifier-Replay-GRPO<u>0.8311</u><u>0.7729</u><u>0.8296</u>0.90280.8815<u>0.8920</u>
Multi-agent with additional context; with debate
Verifier-SFT0.79560.71360.79610.87120.87910.8751
Verifier-GRPO0.82750.76490.82540.90500.87290.8887
CADRE (Ours)0.83340.77660.8321<u>0.9035</u>0.88400.8936
Student models (mBERT; CoT teacher: CADRE)
Student-Label-Only0.74560.67260.72920.86790.73870.7981
Student-KD (Experts CoTs)0.76940.72290.76630.8457<u>0.8238</u>0.8346
Student-KD (Verifier CoT)<u>0.7795</u><u>0.7306</u><u>0.7725</u><u>0.8609</u>0.8137<u>0.8366</u>
Student-KD (Experts CoTs + Verifier CoT)0.78090.73910.77970.84610.84710.8466

Table 1. Teacher and student performance on the in-house dataset, with the best results in each group shown in bold and the second-best results underlined.

For multi-agent methods, Verifier-Base uses the off-the-shelf checkpoint to adjudicate expert reports. Verifier-SFT applies supervised fine-tuning, while Verifier-GRPO and Verifier-Replay-GRPO apply the corresponding RL methods after supervised initialization. We compare variants with and without selective debate, with CADRE denoting Verifier-Replay-GRPO with debate. Using the same initial expert reports as CADRE, Expert Oracle measures the maximum Accuracy achievable by selecting among the three experts’ initial labels, counting an input as correctly classified whenever at least one expert predicts its gold seven-class label. Only Accuracy is reported because the oracle leaves the selected label unspecified when all experts are incorrect, preventing the other metrics from being uniquely determined.

For fair offline comparisons, all student variants use mBERT and the same 78,997-pair training split used for teacher training. Gold labels supervise classification, with CADRE providing CoTs for the distillation variants. Student-Label-Only uses gold labels alone. Student-KD (Experts CoTs) additionally uses concatenated CoTs from the three expert agents, while Student-KD (Verifier CoT) uses the Verifier’s adjudication rationale. Student-KD (Experts CoTs + Verifier CoT) combines both sources of reasoning supervision. For inputs that do not trigger debate, only first-round CoTs are available.

Evaluation Metrics

For offline evaluation, we assess both fine-grained relevance classification and irrelevant-product identification. For seven-class classification, we report Accuracy (ACC), Macro-F1, and Weighted-F1. Since irrelevant product exposure can degrade the search experience [2], we additionally report Precision, Recall, and F1 for irrelevant classification. We map both predicted and gold labels to a binary label space, treating Strongly Relevant and Substitute as negative and the remaining five categories as positive (irrelevant). These metrics assess the model’s ability to identify irrelevant products while limiting erroneous rejection of relevant candidates.

Implementation Details

We conduct experiments on a single node with 8 NVIDIA H100 GPUs (80 GB each). All agents in CADRE use Qwen3-14B as their backbone. The experts share a frozen model instance with separate contexts, while only the Verifier is optimized. We implement this interaction and masking procedure through the agent loop in verl [19].4 Within each experimental setting, GRPO and Replay-GRPO start from the same SFT checkpoint and sample G=8G=8 fresh trajectories per input, with Replay-GRPO additionally incorporating M=2M=2 stored successful trajectories. Detailed training and replay configurations are provided in Appendix § 9.

Performance

Overall Performance

As shown in Table 1, CADRE achieves the highest reported seven-class classification scores, with 0.8334 ACC, 0.7766 Macro-F1, and 0.8321 Weighted-F1. Compared with LLM-Replay-GRPO with additional context, these results represent improvements of 3.73, 13.64, and 4.06 percentage points, respectively. CADRE also achieves the highest irrelevant-class F1 of 0.8936, indicating stronger overall performance in identifying irrelevant query–product pairs. Together with its fine-grained mismatch predictions, this capability can support filtering decisions that prioritize severe mismatches, helping reduce irrelevant product exposure and improve the search experience.

Expert Oracle achieves an accuracy of 0.8051. CADRE exceeds this coverage by 2.83 percentage points, indicating that the framework can correctly adjudicate some instances where all three initial expert labels are incorrect. This result supports CADRE’s robustness to imperfect initial expert predictions, as its final adjudication draws on supporting evidence and any debate feedback without requiring every expert to predict the correct label. Appendix D provides qualitative examples of independent adjudication and correction through Verifier-guided debate.

Effectiveness of CADRE

With additional context, Verifier-SFT without debate outperforms LLM-SFT by 6.70, 9.71, and 7.25 percentage points in ACC, Macro-F1, and Weighted-F1, respectively. Without debate, Verifier-GRPO outperforms Verifier-SFT by 2.19, 2.92, and 1.85 percentage points in ACC, Macro-F1, and Weighted-F1, respectively, demonstrating the benefit of reinforcement learning for adjudicating expert reports. CADRE also outperforms Verifier-Replay-GRPO without debate across all six reported metrics, supporting the contribution of selective debate within the Replay-GRPO setting. By incorporating selective debate and successful-experience replay, CADRE further extends the gains over the same Verifier-SFT baseline to 3.14, 4.15, and 2.76 points, respectively.

Effectiveness of KD

We compare mBERT students trained with different CoT components produced by CADRE. Compared with gold-label supervision alone, the best configuration, Experts CoTs + Verifier CoT, improves ACC, Macro-F1, and Weighted-F1 by 3.53, 6.65, and 5.05 percentage points, respectively, and irrelevant-class F1 by 4.85 points. These results support the value of the teacher’s reasoning beyond label supervision. Among the CoT-based variants, Verifier CoT outperforms the concatenated Experts CoTs across all seven-class metrics (e.g., Macro-F1 0.7306 vs. 0.7229). This suggests that the trained Verifier’s adjudicated reasoning provides more effective distillation supervision than simply concatenating expert analyses.

MethodACCMacro-F1Weighted-F1
Full CADRE0.83340.77660.8321
w/o reasoning reward0.82610.76850.8241
w/o shaping reward0.82670.76950.8254

Table 2. Ablation of the composite reward.

Figure 3. Test-set accuracy of GRPO, GRPO-16, and CADRE (Replay-GRPO) in the debate-enabled setting.

Ablation Studies

We examine three components of CADRE, including the reward components, the debate protocol, and the successful-experience replay. We count the initial expert reports as the first expert-response round and each subsequent round of revisions as an additional round. Unless otherwise specified, the variants retain three expert inputs and use two expert-response rounds, consisting of the initial reports and one round of revisions following Verifier-guided questioning, without direct communication among experts.

Reward Components

Reward Components. Table 2 shows that both auxiliary reward terms contribute to performance. Removing the reasoning reward decreases ACC, Macro-F1, and Weighted-F1 by 0.73, 0.81, and 0.80 percentage points, respectively. Removing debate shaping produces corresponding decreases of 0.67, 0.71, and 0.67 points. These results support rewarding both a consistent adjudication process and productive expert revisions in addition to final-label correctness.

MethodACCM-F1W-F1
CADRE (2 rounds, independent)0.83340.77660.8321
3 rounds, independent0.83390.77680.8324
2 rounds, initial-stage sharing0.82200.76100.8219
2 rounds, debate-stage sharing0.81580.75210.8168

Table 3. Ablation of debate rounds and expert report sharing. M-F1 and W-F1 denote Macro-F1 and Weighted-F1.

Debate Protocol

Debate Protocol. Table 3 examines expert report sharing at different stages and the number of expert-response rounds. CADRE uses two rounds, keeping both initial analyses and subsequent expert revisions independent. With initial-stage sharing, experts generate their first-round reports sequentially, and each expert can observe the reports produced by preceding experts; the subsequent debate retains the original Verifier-guided revision protocol. This variant reduces ACC, Macro-F1, and Weighted-F1 by 1.14, 1.56, and 1.02 percentage points relative to CADRE, respectively. This degradation is consistent with the concern illustrated in Figure 1, where one perspective overrides decisive evidence identified from another. With debate-stage sharing, initial analyses remain independent, but each expert can directly observe the other experts’ analyses and predicted labels when revising its own report. This variant reduces the same metrics by 1.76, 2.45, and 1.53 percentage points, respectively. Exposure to peer judgments may encourage premature convergence at either stage, reducing the complementary evidence available to the Verifier. These results support preserving independent expert analyses and revisions.

With independent experts, extending the interaction from two to three response rounds yields only marginal gains of 0.05, 0.02, and 0.03 percentage points in ACC, Macro-F1, and Weighted-F1, respectively, while requiring an additional round of expert calls.

Successful-Experience Replay

Successful-Experience Replay. As shown in Table 1, successful-experience replay improves Verifier performance under both interaction settings. Without debate, Verifier-Replay-GRPO outperforms Verifier-GRPO by 0.72, 0.86, and 0.66 percentage points in ACC, Macro-F1, and Weighted-F1, respectively; with debate, the corresponding gains of CADRE over Verifier-GRPO are 0.59, 1.17, and 0.67 points. Figure 3 compares accuracy on the test set in the debate-enabled setting, with all methods initialized from the same SFT checkpoint. These test-set evaluations were conducted retrospectively and were not used for checkpoint selection, early stopping, or hyperparameter tuning. By replaying successful trajectories, CADRE consistently achieves higher accuracy than GRPO and GRPO-16, whereas increasing fresh rollouts from 8 to 16 yields only modest gains over standard GRPO in this setting.

Online Evaluation

For deployment, we train the best student on 1.64 million human-labeled query–product pairs, using gold labels for classification and CADRE-generated CoTs for reasoning distillation. We conducted a 7-day online A/B test in an e-commerce search advertising system, allocating 10% of user traffic each to this student and the existing production baseline. During serving, predefined business rules and calibrated thresholds map query–item relevance predictions to discrete levels, filtering candidates below the relevance threshold before ranking. We assess business performance using Revenue Per Mille (RPM) and Click-Through Rate (CTR), and relevance using Relevance Satisfaction Score (RS), defined as the platform’s traffic-weighted strong-relevance metric. We sample results for the same queries from the two experimental buckets under the same conditions and submit them to human experts for relevance assessment. The query set covers approximately 90% of search traffic, and RS aggregates query-level strong-relevance rates weighted by each query’s search traffic. Compared with the existing production baseline, the CADRE-distilled student achieves relative gains of 1.69% in RPM and 0.76% in CTR, alongside an absolute increase of 0.3 percentage points in RS. These results demonstrate that the CADRE-distilled student improves commercial performance while also increasing the traffic-weighted strong-relevance rate.

Conclusion

In this paper, we introduced CADRE to improve evidence integration in e-commerce relevance modeling. Its central design separates complementary expert analysis from learned adjudication, enabling a Verifier to weigh evidence across perspectives and seek clarification through selective debate. Multi-turn reinforcement learning optimizes this adjudication process, while experience replay preserves successful trajectories for reuse during training. We further transfer CADRE’s reasoning to a lightweight student for online deployment. Offline experiments demonstrate improved relevance prediction, while online A/B testing confirms that the distilled student improves business performance and relevance satisfaction over the existing production baseline. Given the current fixed-round debate protocol, future work will explore adaptive policies for expert selection, evidence clarification, and stopping to tailor the scope and duration of interaction to input difficulty while accounting for computational cost.

References

Interaction Protocol and Policy-Loss Masking

Table 4 summarizes the debate interaction sequence. Only Verifier-generated tokens contribute to the policy loss (m=1m = 1), while expert outputs and environment messages are masked (m=0m = 0). For inputs not selected for debate, the initial adjudication serves as the final prediction.

Dataset Statistics

Table 5 reports the final cleaned labels for 97,707 pairs, with 78,997 training and 18,710 test pairs whose class proportions closely match. Relevant and irrelevant pairs account for 52.27% and 47.73%, respectively. Macro-F1 complements Accuracy and Weighted-F1 by weighting classes equally despite their unequal frequencies. The five most represented languages are English, Spanish, Portuguese, Korean, and French.

LabelCountShare (%)
Strongly Relevant49,95551.13
Substitute1,1141.14
Category Mismatch9,3269.54
Brand Mismatch9,1359.35
Model Mismatch9,1459.36
Attribute Mismatch10,45410.70
Main Product vs. Accessory8,5788.78
Total97,707100.00

Table 5. Seven-class label distribution of the full dataset.

Implementation and Training Configuration

The experts share a frozen Qwen3-14B instance at a dedicated vLLM endpoint, with separate prompts, contexts, and conversation histories, each incurring a generation cost. The Verifier is a separate model initialized through full-parameter SFT on multi-turn interactions using LLaMA-Factory5. Training runs for 3 epochs with a learning rate of 1×10−51 \times10^{-5} and a batch size of 16, using BF16 precision and ZeRO-2.

Both GRPO and Replay-GRPO use ver1, starting from θSFT\theta_{\mathrm{SFT}} and training for 3 epochs with a learning rate of 1×10−61 \times10^{-6}, a batch size of 256, and G=8G = 8 fresh trajectories per input.

For replay, θGRPO\theta_{\mathrm{GRPO}} generates 16 trajectories per input at temperature 1.0 using the same Verifier–expert interaction protocol as in training. Following Section 3.3.3, we retain at most K=8K = 8 trajectories with parseable, exactly correct terminal predictions, ranked by total reward. Replay-GRPO adds M=2M = 2 stored trajectories per input to the fresh rollouts.

The mBERT-base student is trained for 3 epochs with a learning rate of 2×10−52 \times10^{-5}, a batch size of 128, and λ=0.1\lambda= 0.1. For each training pair, we concatenate the CoTs generated across all available rounds by the Analyst, Intent, and Ruler Agents and the Verifier. Each variant includes only the CoTs from its corresponding agents. The concatenated text is encoded by a frozen BGE-M3 model into a single embedding ecot\mathbf{e}_{\mathrm{cot}}. Since the concatenated CoTs fit within BGE-M3’s maximum length of 8,192 tokens, no truncation is applied. No correctness-based filtering is applied to the CoTs used for distillation.

Reward Implementation

This section specifies the scoring rules and coefficients used in Section 3.3.2.

Adjudication quality. We compute

radj=0.1rfmt+Gate⁡fmt(0.8rres+0.1rreason)(8)r_{\mathrm{adj}} = 0.1r_{\mathrm{fmt}} + \operatorname{Gate}_{\mathrm{fmt}}\left(0.8r_{\mathrm{res}} + 0.1r_{\mathrm{reason}}\right) \tag*{(8)}

where Gate⁡fmt=rfmt∈{0,1}\operatorname{Gate}_{\mathrm{fmt}} = r_{\mathrm{fmt}} \in\{0,1\} indicates whether the terminal prediction is parseable as a valid relevance label. Let y^\hat{y} denote the terminal prediction, which is y^(1)\hat{y}^{(1)} without debate and y^(2)\hat{y}^{(2)} otherwise, and let y∗y^* denote the gold label. For a valid prediction, the outcome reward is

rres={1,y^=y∗,0.65,y^≠y∗ and ϕ(y^)=ϕ(y∗),0,otherwise,(9)r_{\mathrm{res}} = \begin{cases} 1, & \hat{y} = y^*, \\ 0.65, & \hat{y} \ne y^* \text{ and } \phi(\hat{y}) = \phi(y^*), \\ 0, & \text{otherwise}, \end{cases} \tag*{(9)}

where ϕ\phi maps each fine-grained label to its relevant or irrelevant superclass. Invalid terminal labels also receive rres=0r_{\mathrm{res}} = 0. The reasoning reward is rreason=(cself+cstruct)/2r_{\mathrm{reason}} = (c_{\mathrm{self}} + c_{\mathrm{struct}})/2, where cselfc_{\mathrm{self}} is computed by a deterministic rule that checks whether the label extracted from the rationale’s concluding step matches the terminal prediction, while cstructc_{\mathrm{struct}} checks completion of the adjudication steps, coverage of all three expert verdicts, and inclusion of a concluding statement.

Figure 4. Correcting unanimous expert errors without debate.

Verifier-guided debate refining expert analyses for a bag-accessory query

Figure 5. Debate refines expert analysis while preserving a correct adjudication.

Debate quality. When the three experts’ first-round labels are not unanimous, the debate reward is

rshape=bR1+pfmt+pleak+sspec+sshift.(10)r_{\mathrm{shape}} = b_{\mathrm{R1}} + p_{\mathrm{fmt}} + p_{\mathrm{leak}} + s_{\mathrm{spec}} + s_{\mathrm{shift}}. \tag*{(10)}

The first-round bonus bR1b_{\mathrm{R1}} is 0.05 when the initial Verifier judgment matches the gold label and zero otherwise. Question quality is assessed through format and leakage penalties together with a heuristic specificity bonus. A missing question or one whose length falls outside the permitted range incurs pfmt=−0.1p_{\mathrm{fmt}}=-0.1, while a focused question that matches a predefined regular-expression pattern for explicit disclosure of the Verifier’s own judgment incurs pleak=−0.1p_{\mathrm{leak}}=-0.1. Each penalty is zero when its corresponding violation is absent. The specificity bonus is sspec=0.05ρs_{\mathrm{spec}} = 0.05\rho, where ρ\rho is the fraction of questions that exceed 60 characters and cite a query or product-title token containing at least four characters.

For expert revisions, we compare each expert’s predictions before and after debate using ck(t)=1[yk(t)=y∗]c_k^{(t)} = 1[y_k^{(t)} = y^*], which indicates whether expert kk predicts the gold label in round tt. The four cases are an incorrect prediction becoming correct, a correct prediction remaining correct, an incorrect prediction remaining incorrect, and a correct prediction becoming incorrect.

The expert-revision reward is

sshift=0.1∣K∣∑k∈KΔk,s_{\mathrm{shift}} = \frac{0.1}{|\mathcal{K}|} \sum_{k \in\mathcal{K}} \Delta_k,
Δk={1,(ck(1),ck(2))=(0,1),0.3,(ck(1),ck(2))=(1,1),0,(ck(1),ck(2))=(0,0),−1,(ck(1),ck(2))=(1,0).(11)\Delta_k = \begin{cases} 1, & (c_k^{(1)}, c_k^{(2)}) = (0,1),\\ 0.3, & (c_k^{(1)}, c_k^{(2)}) = (1,1),\\ 0, & (c_k^{(1)}, c_k^{(2)}) = (0,0),\\ -1, & (c_k^{(1)}, c_k^{(2)}) = (1,0). \end{cases} \tag*{(11)}

Case Studies

Figures 4–6 present condensed and paraphrased representations of actual model outputs on three test-set cases, illustrating independent adjudication, expert refinement, and Verifier correction. The figures summarize the key evidence, judgments, and interactions for readability, rather than reproduce complete transcripts; full role-specific input contexts are omitted for brevity.

Figure 6. Debate corrects an initial Verifier error without expert consensus.

Correcting Unanimous Expert Errors

Figure 4 pairs “realme gt7 pro case” with a case advertised for the GT7. All three experts predict Strongly Relevant, with the Analyst overlooking “Pro” and the Intent and Ruler Agents inferring compatibility from shopping intent and a shared product line. The Verifier checks these interpretations against the query and title and rejects the unsupported compatibility assumption, yielding Model Mismatch. Expert unanimity prevents debate but does not restrict the Verifier’s decision, illustrating independent adjudication when all initial expert labels are incorrect.

Refining Expert Analysis through Debate

Figure 5 pairs “ita bag accessories” with a replacement strap for a Longchamp Mini Bag. The Analyst predicts Strongly Relevant from the shared accessory category, the Intent Agent interprets Longchamp as a brand conflict and predicts Brand Mismatch, and the Ruler assigns Category Mismatch based on accessory function. The Verifier uses the Intent report’s click evidence, including charms and badge kits, to identify a decorative need that the strap’s carrying function does not satisfy, yielding Attribute Mismatch. During debate, it relays this evidence to the Analyst and asks the Intent Agent to distinguish a target-bag reference from a brand requirement. Both revise to Attribute Mismatch, while the Ruler retains Category Mismatch under a finer functional boundary. The Verifier maintains its correct judgment, illustrating how the expert refinement rewarded by sshifts_{\mathrm{shift}} in Eq. (11) can improve supporting analyses without changing the final label.

Correcting Verifier Errors through Debate

Figure 6 pairs “short human hair for black women” with 24-inch hair extensions. The Analyst predicts Substitute based on possible trimming, the Intent Agent prioritizes click-supported interest in extensions and predicts Strongly Relevant, and the Ruler predicts Attribute Mismatch from the explicit length conflict. The Verifier initially accepts trimming and predicts Strongly Relevant. During debate, its questions examine whether substitution depends on modifying the item, whether behavioral evidence justifies relaxing the length requirement, and how the matching rubric treats hypothetical trimming. The Analyst retains Substitute, while the Intent Agent revises to Attribute Mismatch and the Ruler reinforces assessment at the listed length. The Verifier then recognizes that possible trimming does not resolve the mismatch in the product as offered and corrects its prediction to Attribute Mismatch, despite persistent expert disagreement.

References

  1. [1]Sanjay Agrawal, Faizan Ahemad, and Vivek Varadarajan Sembium. 2025. Rationale-Guided Distillation for E-Commerce Relevance Classification: Bridging Large Language Models and Lightweight Cross-Encoders. In Proceedings of the 1st International Conference on Computational Linguistics: Industry Track, Owen Rambow, Leo Wanner, Marianna Apidianaki, Hend Al-Khalifa, Barbara Di Eugenio, Steven Schockaert, Kareem Darwish, and Apoorv Agarwal (Eds.). Association for Computational Linguistics, Abu Dhabi, UAE, 136–148. https://aclanthology.org/2025.coling-industry.12/
  2. [2]Jiarui Che, Yifei Chen, Zhixing Tian, Chenyang Wang, and Ziguang Cheng. 2026. Graph-GRPO: Dependency-Aware Credit Assignment for Generative E-commerce Search Relevance. arXiv:2605.31003 https://arxiv.org/abs/2605.31003
  3. [3]Jianlyu Chen, Shitao Xiao, Peitian Zhang, Kun Luo, Defu Lian, and Zheng Liu. 2024. M3-Embedding: Multi-Linguality, Multi-Functionality, Multi-Granularity Text Embeddings Through Self-Knowledge Distillation. In Findings of the Association for Computational Linguistics: ACL 2024, Lun-Wei Ku, Andre Martins, and Vivek Srikumar (Eds.). Association for Computational Linguistics, Bangkok, Thailand, 2318–2335. doi:10.18653/v1/2024.findings-acl.137DOI
  4. [4]Yifei Chen, Zhixing Tian, Chenyang Wang, and Ziguang Cheng. 2026. K-CARE: Knowledge-driven Symmetrical Contextual Anchoring and Analogical Prototype Reasoning for E-commerce Relevance. arXiv:2604.25683 [cs.IR] https://arxiv.org/abs/2604.25683
  5. [5]Zeyuan Chen, Haiyan Wu, Kaixin Wu, Wei Chen, Mingjie Zhong, Jia Xu, Zhongyi Liu, and Wei Zhang. 2025. Towards Boosting LLMs-driven Relevance Modeling with Progressive Retrieved Behavior-augmented Prompting. In Proceedings of the 31st International Conference on Computational Linguistics: Industry Track, 784–793.arxiv.org/abs/2408.09439
  6. [6]Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In Proceedings of NAACL-HLT. 4171–4186.arxiv.org/abs/1810.04805
  7. [7]Chenche Dong, Shaowei Yao, Pengkun Jiao, Jianhui Yang, Yiming Jin, Zerui Huang, Xiaojang Zhou, Dan Ou, Haihong Tang, and Bo Zheng. 2026. TaoSR1: The Thinking Model for E-commerce Relevance Search. In Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2 (Jeju Island, Republic of Korea) (KDD ’26). Association for Computing Machinery, New York, NY, USA, 7174–7184. doi:10.1145/3770855.3818489
  8. [8]Zheng Fang, Donghao Xie, Ming Pang, Chunyu Yuan, Xue Jiang, Changping Peng, Zhangang Lin, and Zheng Luo. 2025. ADORE: Autonomous Domain-Oriented Relevance Engine for E-commerce. In Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR ’25). Association for Computing Machinery, New York, NY, USA, 4259–4263. doi:10.1145/3726302.3731944
  9. [9]Po-Sen Huang, Xiaodong He, Jianfeng Gao, Li Deng, Alex Acero, and Larry Heck. 2013. Learning deep structured semantic models for web search using clickthrough data. In Proceedings of the 22nd ACM International Conference on Information & Knowledge Management (San Francisco, California, USA) (CIKM ’13). Association for Computing Machinery, New York, NY, USA, 2333–2338. doi:10.1145/2505515.2505665DOI
  10. [10]Samuel Humeau, Kurt Shuster, Marie-Anne Lachaux, and Jason Weston. 2020. Poly-encoders: Architectures and Pre-training Strategies for Fast and Accurate Multi-sentence Scoring. In International Conference on Learning Representations.
  11. [11]Pengkun Jiao, Yiming Jin, Jianhui Yang, Chenhe Dong, Zerui Huang, Shaowei Yao, Xiaojang Zhou, Dan Ou, and Haihong Tang. 2026. SHE: Stepwise Hybrid Examination Reinforcement Learning Framework for E-commerce Search Relevance. arXiv:2510.07972 [cs.AI] https://arxiv.org/abs/2510.07972
  12. [12]Omar Khatab and Matei Zaharia. 2020. Colbert: Efficient and effective passage search via contextualized late interaction over bert. In Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval, 39–48.
  13. [13]Chenji Liu, Zhuo Chen, Hui Zhao, Zhiyuan Zeng, Gang Zhao, Junjie Ren, Ruicong Xu, Haoran Li, Songyan Liu, Pengjie Wang, Jian Xu, and Bo Zheng. 2026. LORE: A Large Generative Model for Search Relevance. arXiv:2512.03025 [cs.IR] https://arxiv.org/abs/2512.03025
  14. [14]Navid Mehrdad, Hrushikesh Mohapatra, Mossaab Bagdouri, Prijith Chandran, Alessandro Magnani, Xunfan Cai, Ajit Puthenputhussy, Sachin Yadav, Tony Lee, ChengXiang Zhai, et al. 2024. Large language models for relevance judgment in product search. arXiv preprint arXiv:2406.00247 (2024).
  15. [15]Baopu Qiu, Hao Chen, Yuannong Wu, Changtong Zan, Chao Wei, Weiru Zhang, and Xiaoyi Zeng. 2026. Thinking Broad, Acting Fast: Latent Reasoning Distillation from Multi-Perspective Chain-of-Thought for E-Commerce Relevance. In Proceedings of the ACM Web Conference 2026 (United Arab Emirates) (WWW ’26). Association for Computing Machinery, New York, NY, USA, 8171–8182. doi:10.1145/3774904.3792812DOI
  16. [16]Rafael Rafailov, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D. Manning, and Chelsea Finn. 2023. Direct preference optimization: your language model is secretly a reward model. In Proceedings of the 37th International Conference on Neural Information Processing Systems (New Orleans, LA, USA) (NIPS ’23). Curran Associates Inc., Red Hook, NY, USA, Article 2338, 14 pages.arxiv.org/abs/2305.18290
  17. [17]Juan Ramos et al. 2003. Using tf-idf to determine word relevance in document queries. In Proceedings of the first instructional conference on machine learning, Vol. 242. New Jersey, USA, 29–48.
  18. [18]Stephen Robertson, Hugo Zaragoza, et al. 2009. The probabilistic relevance framework: BM25 and beyond. Foundations and Trends® in Information Retrieval 3, 4 (2009), 333–389.
  19. [19]Guangming Sheng, Chi Zhang, Zilingfeng Ye, Xibin Wu, Wang Zhang, Ru Zhang, Yanghua Peng, Haibin Lin, and Chuan Wu. 2025. HybridFlow: A Flexible and Efficient RLHF Framework. In Proceedings of the Twentieth European Conference on Computer Systems (Rotterdam, Netherlands) (EuroSys ’25). Association for Computing Machinery, New York, NY, USA, 1279–1297. doi:10.1145/3689031.3696075DOI
  20. [20]Tian Tang, Zhixing Tian, Zhenyu Zhu, Chenyang Wang, Haiqing Hu, Guoyu Tang, Lin Liu, and Sulong Xu. 2025. LREF: A Novel LLM-based Relevance Framework for E-commerce Search. In Companion Proceedings of the ACM on Web Conference 2025 (WWW ’25). ACM, 468–475. doi:10.1145/3701716.3715246DOI
  21. [21]Han Wang, Mukuntha Narayanan Sundararaman, Onur Gungor, Yu Xu, Krishna Kamath, Rakesh Chalasani, Kurchi Subhra Hazra, and Jinfeng Rao. 2024. Improving Interest Search Relevance Using Large Language Models. arXiv:2410.17152 [cs.IR] https://arxiv.org/abs/2410.17152
  22. [22]Runze Xia, Yupeng Ji, Yuxi Zhou, Haodong Liu, Teng Zhang, and Piji Li. 2026. From Reasoning LLMs to BERT: A Two-Stage Distillation Framework for Search Relevance. In Proceedings of the ACM Web Conference 2026 (United Arab Emirates) (WWW ’26). Association for Computing Machinery, New York, NY, USA, 822–8231. doi:10.1145/3774904.3792801
  23. [23]Jianhui Yang, Yiming Jin, Pengkun Jiao, Chenhe Dong, Zerui Huang, Shaowei Yao, Xiaojang Zhou, Dan Ou, and Haihong Tang. 2026. TaoSR-AGRL: Adaptive Guided Reinforcement Learning Framework for E-commerce Search Relevance. In Proceedings of the ACM Web Conference 2026 (Dubai, United Arab Emirates) (WWW ’26). Association for Computing Machinery, New York, NY, USA, 7955–7966. doi:10.1145/3774904.3792799DOI
  24. [24]Shaowei Yao, Jiwei Tan, Xi Chen, Juhao Zhang, Xiaoyi Zeng, and Keping Yang. 2022. RepBERT: Distilling BERT to an Efficient Representation-Based Relevance Model for E-Commerce. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (Washington DC, USA) (KDD ’22). Association for Computing Machinery, New York, NY, USA, 4363–4371. doi:10.1145/3534678.3539090DOI
  25. [25]Hongzhi Zhang, Jia Fu, Jingyuan Zhang, Kai Fu, Qi Wang, Fuzheng Zhang, and Guorui Zhou. 2025. RLEP: Reinforcement Learning with Experience Replay for LLM Reasoning. arXiv:2507.07451 [cs.CL]. https://arxiv.org/abs/2507.07451
  26. [26]Gang Zhao, Ximing Zhang, Chenji Lu, Hui Zhao, Tianshu Wu, Pengjie Wang, Jian Xu, and Bo Zheng. 2025. Explainable LLM-driven Multi-dimensional Distillation for E-Commerce Relevance Learning. In Companion Proceedings of the ACM on Web Conference 2025 (WWW ’25). ACM, 631–640. doi:10.1145/3701716.3715222DOI

Paper details

vlnc:W_34xusstz
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