e-commerce platform teams

Abstract: Generative retrieval models have been increasingly adopted in search and recommendation systems. However, industrial generative retrieval models are commonly specialized to a fixed online scenario: they learn from scenario-specific interactions, support a narrow set of query forms and retrieval objectives. We study how to remove these couplings and build a unified end-to-end multimodal search system for open-ended image--text queries, heterogeneous shopping intents, and online scenarios with different resource constraints. We present \textbf{PailitaoMMSearch}, which directly generates product Semantic IDs from a raw query image and free-form instruction and scales along three dimensions. \textbf{Data scaling} uses a reusable toolset that can be composed into intent- and lifecycle-specific workflows for cold-start synthesis, quality control, real-traffic refinement, and continuous data production. \textbf{Intent scaling} progressively equips one model with e-commerce product knowledge, multiple retrieval relations, and adjustable reasoning budgets through continual pre-training (CPT), mixed-budget supervised fine-tuning (SFT), and unified reinforcement learning (RL). The RL stage combines intent-balanced optimization and Sample--Beam rollouts aligned with multi-candidate serving. \textbf{Scenario scaling} exposes quality--cost operating points across model capacity and reasoning budget. The system is deployed in multimodal search applications on Taobao Pailitao. A controlled online A/B test reports statistically significant improvements, including item page views $\textbf{+17.90\%}$, transactions $\textbf{+27.04\%}$, and gross merchandise value $\textbf{+8.94\%}$.

Abstract: This reproducibility paper explores the counterfactual explanation technique known as Learning to eXplain Recommendations (LXR). Prior work has shown that LXR can reliably identify perturbations that, if applied to a user's profile, cause the user's top ranked item to no longer be recommended, thereby explaining the item's recommendation as a function of the user's rating choices. In this paper, we extend this work in multiple ways. We expand the scope of the original LXR study to look at explanations for items other than the most highly recommended item, finding that LXR is also effective at this version of the task. We consider alternate baselines for LXR, including item popularity and latent factor similarity, and we see that for some tasks these alternatives work surprisingly well. We also examine the impact of item popularity, finding that LXR offers better explanation quality (shorter perturbations) when less popular items are being explained. Finally, we conduct a brute-force exploration of a subset of the perturbation space in order to estimate a lower bound for optimal perturbation size against which LXR and other such methods can be compared.

Abstract: Existing personalized information retrieval methods mainly model differences in user needs through user preferences and query intent, while paying little attention to the personalized expression differences that arise when different users express the same retrieval need. Existing real query data typically lack stable user identities, long-term expression histories, and intent-matched counterparts, making it difficult to systematically evaluate the impact of personalized expression differences on retrieval outcomes. To this end, we propose PersonalQuery-Benchmark (PQB). PQB models personalized syntactic expression and personalized writing errors from users' historical texts, constructs personalized syntactic expression queries and personalized writing-error queries, and validates benchmark quality through LLM assessment and independent human evaluation. PQB covers three Amazon product domains with 6,516 query instances in total, and is evaluated on five retrieval models representing lexical, learned sparse, dense bi-encoder, late-interaction, and hybrid architectures, together with two LLM-based ranking models. Results show that even when the information need is held fixed, personalized expression differences still induce measurable ranking variation; different retrieval models exhibit clearly different sensitivities to syntactic variation and writing-error variation, while higher overall retrieval effectiveness does not necessarily correspond to stronger robustness against expression variation. PQB provides a new benchmark and evaluation framework for systematically studying personalized expression robustness. The associated code and dataset are available at the following address: https://github.com/cpxjazz123/PersonalQuery-Benchmark.

Abstract: Web agents increasingly serve the same user across repeated online tasks such as booking travel, shopping, and scheduling, so their long-term usefulness depends on how well they know the user's preferences. Passive memory and personalization methods cannot obtain preferences the user never reveals, and clarification methods that value a question by its benefit to the current task overlook the benefit of reusing the answer in later tasks. We study Active Preference Acquisition (APA): which missing preferences are worth the cost of asking? We define the Prospective Personalization Value (PPV) of a missing preference as its expected cumulative utility gain over future tasks minus the effort and disclosure-risk cost of asking, and select questions under a per-interaction budget. We characterize when cumulative reuse justifies acquisition despite insufficient single-task benefit, and evaluate this regime with an initial-profile diagnostic. APA computes the counterfactual gain of each candidate preference on public task prototypes with the agent's own task solver and aggregates it by the expected reuse of each task family. We evaluate it on ActiveWebPref, a controlled longitudinal benchmark with hidden but queryable user profiles and exact task utilities. At a horizon of 20 tasks, 34.4% of budget-feasible valuable missing preferences are amortization-dependent; APA acquires 80.3% of them, whereas current-task value of information acquires 5.1%. Under per-user asking-cost caps matched to APA's realized expenditure, APA achieves higher net utility than four heuristic selection baselines. In our cost setting, asking about every unknown preference yields lower net utility than not asking, while APA recovers 84.2% of its utility gain with 27.6% of its questions. The value of acquiring a user preference should account for its reuse across future tasks, not only its benefit to the task at hand.

Abstract: Large-scale conversational recommendation datasets typically require manual dialogue construction or annotation. We propose SearCl, which automatically generates conversational recommendation dialogues grounded in user search and click behaviors. SearCl segments user interactions into query sessions and derives each recommendation target from content selected after a search query. A configurable probabilistic state machine controls the dialogue flow, while two LLM agents acting as the seeker and recommender generate utterances according to turn-level constraints. To support recommendation, the recommender is provided with knowledge graphs constructed from metadata about the content items used in each dialogue, capturing their attributes and relationships. Using SearCl, we generate 3,538 dialogues from content search and click interaction data, covering 1,731 unique target items. Under an LLM-as-a-judge protocol, the resulting dialogues achieve higher quality scores than dialogues from existing benchmarks. These results demonstrate that SearCl enables controllable and knowledge-grounded dataset construction without human annotations.

Abstract: We present CommerceGen-Studio, an end-to-end framework for generating commercial-grade main images for e-commerce Stock Keeping Units (SKUs). Our framework is organized along a three-stage pipeline of information extraction, semantic rewriting, and multi-condition generation: Prompt-Eval-Refine-Loop Agent (PERL-Agent) distills heterogeneous SKU metadata into concise selling-point text lists through an iterative closed loop of VLM and LLM roles scored by a reference-free rubric. Layout Thinker Optimization (LTO) aligns the rewriter with both caption and rendered-image quality via alternating domain-specific policy optimization. SKU-DiT, trained under Reinforce-Distill-Reinforce (RDR) progressive training, conditions jointly on the white-background product image, the trademark crop, and the structured layout to hold subject fidelity, trademark integrity, and character-level text accuracy together. The framework runs in production at a scale of millions of main images per day with a latency of a few seconds per image, which we believe makes it among the first systems of this kind to operate at commercial scale. We also release CommerceGen-Bench, a benchmark of hard cases collected from live traffic across all product categories. On it, our framework achieves the highest overall scores among the evaluated open-source baselines on both editing tracks, with and without prompt enhancement. It also matches much larger closed-source models, and beats several of them at Instruction Following and Product Integration - the two properties that determine whether an image can be published.

Abstract: AI agents now browse and buy on people's behalf, and benchmarks measure how well the agents perform. Far less is known about the other side of the transaction: whether a merchant's website lets an agent in, publishes terms precise enough to decide a purchase, and leaves evidence of what was agreed. This paper presents a merchant-side benchmark built on those three questions and reports a pilot of it. The central measurement, decision sufficiency, rates each criterion of a concrete buyer mandate (total cost, cancellation or return terms, service date, variant identity) as determinate, indeterminate, or unknown from the merchant's own published evidence. Readiness scores carry their evidence coverage and missingness bounds; access is reported separately for a declared bot, a declared headless browser, and a user-present browser; and each observed term is mapped to the RAILS admissibility class of the evidence that could support it in a dispute. Four large merchants received the full assessment: three refused the declared profiles at their offer pages, and no mandate could be confirmed as met, because each merchant left a total or a refund term undetermined. An access census of 555 sites of merchants that two payment processors publicly name as clients found the home page unreadable to both declared profiles at 105 sites (18.9%). A declared assessment of 330 merchants, scored by deterministic rules without a rater, found that evidence coverage rather than the score decides who can be ranked: 115 merchants (35%) reached the ranking threshold, rankable profiles scored a median of 89 to 90, and material terms was the one pillar well below the others, because the all-in total was decidable at 18 merchants. The pilot shows that the measurements can be taken; it does not validate them or rank merchants, and the weights and the projected clearing band await validation against outcomes.

Abstract: Collaborative filtering predicts user-item interactions from implicit feedback. We focus on matrix factorization (MF) trained with the Bayesian Personalized Ranking loss. This loss updates each user embedding only through that user's observed items and sampled negatives, so sparse feedback gives MF few constraints. To constrain each user through other users, we propose Pathwise Alignment for Collaborative Filtering (PACF), which trains MF with one network shared by all training pairs. At points between user and item embeddings, a flow matching term trains the network to predict the user-to-item step, reversed and scaled for sampled negatives. A trajectory consistency term pulls each user's predicted path to end at the observed item. Both terms update the embeddings, so the network is discarded after training, thereby maintaining the efficiency of MF at inference. We prove that a network with bounded sensitivity to its input point accurately fits users who share an item only if their embeddings are close, so these users constrain one another. Against eleven discriminative and generative baselines, PACF achieves the highest performance on all seven datasets. Its relative gains over MF are largest where feedback is sparsest. The shared network thus supplies the missing constraints through other users.

Abstract: Federated Recommendation Systems (FedRSs) learn collaborative scores from interaction histories on device and rank items by them, so a higher score means a higher rank. Item text and images add semantic information that can reorder the list. Existing Federated Multimodal Recommenders (FedMRs) fuse this content into their models, which can obscure how content reorders the list. We therefore propose the SEmantic Residual Adapter (SERA), which keeps content out of an existing FedRS and adds semantic residual scores to its collaborative scores. Content can then lift an item above another only when its semantic residual score exceeds the other's by more than their collaborative score gap. This score grows with the item's content similarity to the user's interaction history and is capped at a residual bound. We prove that for any content, an item can enter the list only if its collaborative score is within twice this bound of the lowest one in the list. A larger bound lets content lift more items, including irrelevant ones, so the bound trades off how far content reaches against accuracy. With a validated bound, SERA improves all three FedRSs on all six datasets without adding personalized parameters. It outperforms FedMRs in most comparisons and lifts some FedRSs above centralized baselines.

Abstract: Large language model (LLM)-enhanced recommender systems leverage rich item semantics to support personalized recommendation. However, semantic representations alone do not determine which historical behaviors reflect persistent preferences and which mainly indicate recent interests, leaving an important aspect of user understanding unresolved. Recent advances in LLM inference show that newly available information can be used to refine the internal state during inference, thereby improving subsequent predictions. Inspired by this principle, we propose MATE (Memory Adaptation with Temporal Evidence), an adaptive user modeling framework for LLM-enhanced sequential recommendation. MATE first evaluates each newly observed interaction from two temporal perspectives: whether it is repeatedly supported by historical behaviors and whether it is consistent with recent interactions. The resulting temporal evidence controls the updates of two user-specific memories, where the long-term memory conservatively preserves persistent preferences while the short-term memory rapidly adapts to recent interests. For each recommendation, a recent-context representation dynamically determines how strongly the two memories contribute to the current user representation. During offline training, next-item prediction is jointly optimized with temporal supervision, while during online adaptation, the shared model remains fixed and only the two user memories are updated from newly observed interactions. Experiments on MovieLens-10M, Amazon Luxury Beauty, and KuaiRec show that MATE improves mean NDCG@10 over the strongest external baseline by 7.0--13.2%. Further analyses support its ability to adapt to recent interests while retaining useful information about recurring earlier preferences.

Abstract: LLM-based personal agents are emerging as persistent carriers of user semantics and intermediaries between users and recommendation platforms, maintaining richer user knowledge locally. As agents interact with one another, the conventional \textit{User--Platform} relation evolves into a \textit{User--Agent Web--Platform} information pathway, enabling distributed user-side information to complement item-side information. This new pathway, however, defies conventional recommendation: evidence is scattered across mutually opaque agents and reachable only through bounded queries, only a small portion of it is relevant to the current recommendation decision, and the responses returned by different agents are semantically heterogeneous. We therefore recast recommendation over the agent web as a \emph{task-time evidence acquisition and fusion} problem under a finite evidence budget by deciding what to ask and what to keep, rather than learning from aggregated data. We propose AgentWebRec, a user-agent-oriented framework that progressively acquires and fuses distributed evidence for each user-item decision while keeping underlying agent memories local. It grounds each decision in platform-provided item semantics and task-relevant evidence from the target user agent's private memory, and conditionally queries neighboring user agents for complementary preference patterns when local evidence is insufficient. Experiments on four InstructRec datasets show that AgentWebRec consistently outperforms baseline recommenders, and ablations verify that the evidence layers contribute complementary gains.

Abstract: Generative recommendation, has been attracted a surge of attentions in industrial and academic research community, towards to build more smart system to build next-generation recommender. Under the significant developing wave of large language model, our team have been developed Semantic ID based OneRec/OneRec-V2. These models have been widely deployed in production and demonstrate the scaling potential of the autoregressive next-item prediction paradigm for industrial recommender systems. Building on the success of OneRec, we further explored a series of models, including OneRec-Think, OpenOneRec, and OneReason, that connect item Semantic IDs with natural language in a unified representation space and seek to unlock the potential of natural-language chain-of-thought (CoT) reasoning for recommendation. However, our preliminary works found that introducing reasoning CoT does not always improve the recommendation performance. To address this issue, OneReason strengthens the semantic alignment between items and language, introduces structured template-based supervision for interest reasoning, and applies advanced reinforcement learning techniques to make reasoning more beneficial to recommendation. As a frontier topic to building recommendation foundation models, we believe this topic has significant research value and hope to encourage more researchers to explore it together. To this end, together with the SIGIR 2026 community, we organized the KUAISHOU Explorer LLM-Rec Challenge 2026: Reasoning Generative Recommendation.

Abstract: We study how well simple mechanisms approximate gains from trade (GFT) in two-sided matching markets with independent buyer values and seller costs, where feasible outcomes form an arbitrary downward-closed family of matchings. This model includes bilateral trade and double auctions as special cases. We focus on the Generalized Random-Offerer (GRO) mechanism, which is an equal mixture of the Generalized Sellers-Offering Mechanism (GSOM) and the Generalized Buyers-Offering Mechanism (GBOM). We determine GRO's exact worst-case approximation ratio with respect to first-best GFT, showing that it is $e$. This improves the previous $3.15$ approximation guarantee (Babaioff et al. STOC 2026) for the same mechanism. In bilateral trade, the result implies a $1/e$ guarantee for the random-offerer mechanism, improving the previous $1/π$ bound (Jo 2026). Our analysis uses a two-dimensional inequality in quantile space to compare first-best GFT with optimal one-sided auction profits. For each potential trade, we identify the region of buyer values and seller costs for which the first-best allocation selects that trade, and then randomly shrink this region in quantile space, yielding posted-price rules whose expected profits can be evaluated exactly. We establish tightness of the $e$ approximation by constructing markets with regular type distributions, pairwise disjoint trading edges, and a single knapsack constraint. On these instances, GRO's expected GFT approaches a $1/e$ fraction of first-best GFT, therefore ruling out any better guarantee even under these restrictions.

Abstract: ZS-CIR aims to retrieve a target image from a reference image and a modification text without paired supervision, typically by encoding composed queries as text-dominant representations within the image-text matching space of VLPs. However, queries reconstructed by visual pseudo-word learning or MLLM-based target reasoning often deviate from the native VLP representation space due to reference noise and coarse text fusion in the former, and verbose, weakly visually grounded descriptions in the latter. In this paper, we propose a unified ZS-CIR framework (named VMIR-CVI) to reconstruct multimodal composite queries from two complementary perspectives for optimizing VLP-compatible multimodal intent representation. First, it reasons and converts the multimodal intent into a unified textual description, aligning with the native text space of the VLP backbones to produce more retrieval-compatible textual queries. Second, it reconstructs the query representation with correctly decoupled visual instance cues, reducing reference noise while preserving target-relevant content. Specifically, a VLP-aligned Multimodal Intent Reasoning (VMIR) module injects few-shot VLP-style exemplars into chain-of-thought prompts, guiding the MLLM to generate target-consistent intent queries. A Training-free Visual Instance Disentanglement (TVID) module decouples fine-grained visual instances from global reference features without additional optimization. Finally, a lightweight Hybrid-modal Intent Alignment and Fusion (HIAF) module integrates the reasoned textual intent and disentangled visual cues into a unified hybrid-modal representation for robust ZS-CIR. Extensive experiments on three CIR benchmarks, namely CIRR, CIRCO and FashionIQ, show that VMIR-CVI significantly outperforms existing baselines and achieves new state-of-the-art performance. Code and trained models will be publicly released.

Abstract: Many real systems can be represented as growing networks where new nodes and links gradually emerge. The Barabási-Albert model for growing networks, and many models inspired by it, are based on the idea that nodes compete for links. However, the strength and the very presence of this competition have not been tested. We propose a robust statistical approach to quantify how strongly nodes compete for links, and apply it to data from various real systems---commenting on online news and cinema attendance data. We find a range of possible behaviors, from the perfectly elastic case, where new entrants shape network growth in a way that leaves the rest of the system unaffected, to an intermediate case where new entrants measurably affect the rest. Perfect competition is never observed. These findings have direct implications for complex systems modeling and e-commerce applications.

Abstract: Search is one of the most important features in e-commerce, directly driving customer engagement and business growth. A good product search system must show both relevant and desirable results. However, retail search presents unique challenges. User intent can range from exact matches to open-ended discovery. Search systems must also balance multiple goals, such as relevance, revenue, and profit, while keeping response times low. Traditional keyword-based methods often fall short in handling natural language or semantic queries. Vector search helps alleviate these issues, but it can miss key intent signals or return low-precision results. In this paper, we present the design of a hybrid search system at Target that combines lexical and vector search. We describe our approach to data processing, embedding training, precision control for the final result set, multi-channel result fusion (where we compared fusion strategies and adopted weighted interleaving), and the performance optimizations used to maintain low latency for production deployment. Our method improves offline evaluation metrics, and in online A/B testing it raised click-through rate by 0.97%, order conversion by 0.98%, and demand per visitor by 1.10% over lexical-only search, while roughly halving zero-result searches. The resulting system is deployed at scale and serves millions of guests daily.

Abstract: Sequential recommendation aims to predict users' future interests from their historical interactions. Although Large Language Models (LLMs) capture rich item semantics, existing methods often struggle to align collaborative signals with textual semantic knowledge. As a result, the learned item representations fail to capture the complementary strengths of both signals, leading to suboptimal recommendation quality. To address this limitation, we propose Latent Semantic Fusion for Sequential Recommendation via Flow-based Conditional Variational Autoencoders (LSF-SR), a novel framework that uses a Conditional Variational Autoencoder (CVAE) with Normalizing Flows to fuse item ID embeddings and LLM-generated semantic signals. At the core of LSF-SR is a conditional fusion module augmented with planar or radial flows. This module learns a flexible latent space that encourages items with similar semantic profiles to cluster together within the latent manifold. Through extensive experiments on five public benchmark datasets, we demonstrate that LSF-SR consistently outperforms state-of-the-art baselines, achieving gains of up to 12.98% and 14.13% in Recall@20 and NDCG@20, respectively.

Abstract: Cross-country recommendation on modern e-commerce platforms is typically deployed with disjoint user and item ID spaces across markets, removing the shared anchors that conventional cross-domain methods rely on. Generative recommendation (GR) mitigates this by mapping items into a shared token space and training a unified model, but existing approaches keep behavior sequences strictly country-specific, so knowledge transfer occurs only at the parameter level and remains absent at the data level. Inspired by code-switching corpora in multilingual natural language processing, we propose CMRec, a cross-country GR framework that injects cross-country supervision at the data level via dual-constrained, context-aware code-mixing. CMRec first learns a shared semantic codebook from multi-modal content and behavioral co-occurrence across countries. It then uses this codebook to synthesize mixed-country sequences via token-level substitutions that satisfy both static (content) and dynamic (e.g., price, audience, popularity) constraints. Finally, it introduces a context-aware loss that reweights mixed samples according to their plausibility in the current sequence. Experiments on two real-world multi-country datasets and an online A/B test show that CMRec substantially improves recommendation quality in data-sparse countries while preserving performance in data-rich countries, achieving +1.77% advertising revenue and +2.64% orders on a large-scale e-commerce platform.

Abstract: Commercial image editing requires product identity preservation, accurate text rendering, and user appeal alongside general editing quality. We present KwaiMind, an image editing system combining general capabilities with e-commerce specialization. An agent-based data engine maintains approximately 1.8 million high-quality editing pairs. Built on a multimodal diffusion transformer, KwaiMind undergoes continued pre-training and supervised fine-tuning, followed by preference optimization and online reinforcement learning. A general-purpose vision-language judge and specialized rewards for click-through rate (CTR), text rendering, and product consistency guide specialized policies, which are consolidated through on-policy distillation. We introduce Ecom-Bench, covering 11 commercial editing tasks with task-specific visual evaluation and CTR-based ranking. KwaiMind achieves the strongest overall scores among evaluated open-source editors on ImgEdit, GEdit, both language splits of REDEdit, and Ecom-Bench visual quality, and the highest aggregate CTR ranking score among compared systems. Offline, CTR-guided optimization increases the proportion of generated images whose predicted CTR exceeds that of the original product image from 12.16% to 37.41%. In an online A/B experiment, CTR-based selection of product main images yields an approximately 2.44% relative increase in actual CTR. These results demonstrate the value of domain-specific data and reward-driven alignment for commercial image editing.

Abstract: Sequential recommendation must operate under long-tailed item distributions and popularity-driven concentration, often forcing practitioners to trade short-list accuracy against long-tail exposure. In this work, we study feature decorrelation as a mechanism for shaping representation geometry in dot-product sequential recommenders, and analyze how this, in turn, affects popularity-driven concentration. We propose a decorrelation-regularized training framework that augments next-item prediction with an auxiliary redundancy-reduction term, and instantiate it with BT-SR, which uses the Barlow Twins objective. To form label-consistent positive pairs without synthetic corruptions, we pair user histories that share the same next-item target. Beyond accuracy, we provide a geometric analysis showing how decorrelation suppresses shared low-rank directions in the user representation space that can give popular items a global scoring advantage, and we introduce a bucket-based alignment concentration metric to quantify this effect. Experiments on five public benchmarks show that BT-SR consistently improves next-item ranking quality, while the decorrelation strength acts as a simple control knob that reallocates accuracy across head and tail items, enabling accuracy-exposure trade-offs. Our analysis also reveals that the impact on head-vs-tail exposure differs across datasets, reflecting interactions between decorrelation and data temporal structure.

Abstract: Empirical evaluation is central to recommender-systems (RecSys) research, but turning experimental designs into executable code remains a manual and error-prone task. We present AutoRecLab, a Python-based autonomous RecSys lab that automates RecSys experiments from natural-language prompts. Given a research idea, AutoRecLab derives explicit experiment requirements, builds and validates a prototype, and iteratively expands it into the requested full experiment. The workflow combines retrieval-augmented generation (RAG) for documentation lookup, static type verification, and execution-steered tree search. In our demonstration, AutoRecLab autonomously implements an explicit-to-implicit feedback conversion study. In a baseline comparison across six algorithms and three datasets, 8 of 9 runs succeed at an average cost of approx- imately $1 per run with GPT-5.4-mini.

Abstract: Multi-interest recommendation represents each user with multiple interest vectors for fine-grained candidate matching, yet it often suffers from interest collapse, where the learned interests converge to similar representations. We highlight the prevailing single-positive paradigm as one important factor behind this issue. Since each instance provides only one positive item, intents are optimized independently, potentially causing the same best-matching interest to be repeatedly updated toward different positives while leaving the others under-supervised. Moreover, existing methods rarely model how strongly a user activates each interest, leaving scores from different interest channels incomparable at inference. To address these problems, we propose MIMA, a Multi-Interest recommendation framework built on Multi-positive exclusive Assignment. MIMA groups items co-occurring within the same request into a positive set, generates complementary interests with a causal Transformer decoder, and exclusively assigns each positive to supervise a distinct interest via Hungarian matching, so that interest differentiation emerges from the training objective itself rather than auxiliary regularization. A lightweight routing module further estimates user-interest activation probabilities to calibrate scores across interest channels. Experiments on three public datasets and an industrial dataset show that MIMA consistently outperforms state-of-the-art baselines, and an online A/B test yields significant business gains.

Abstract: Industrial recommender systems use cascaded stages with different objectives, feature spaces, and latency constraints. Optimizing pre-ranking and ranking separately can create cross-stage inconsistency: upstream models may filter out items preferred by downstream rankers, and independently tuned downstream fusion can offset upstream improvements. Existing multi-task fusion methods focus on multi-objective fusion within the ranking stage, and cross-stage methods typically only add a downstream score factor to upstream ranking. Joint optimization of fusion modules across both stages remains largely unexplored. We propose UniRec, a Unified Cross-stage Recommendation Fusion model. First, the two fusion agents partially share input embeddings and are trained in a single computation graph, so gradients from either stage propagate through the shared representation and influence the other. Second, we introduce a dual-axis preference alignment objective: a vertical cross-stage consistency term transfers downstream pairwise preferences to the upstream fusion score, and a horizontal compact aggregation term reorganizes dozens of pairwise objectives over heterogeneous prior signals into bidirectional preference evidence. Third, we find that unconstrained end-to-end fusion optimization can exploit imbalances in item attribute distributions, over-concentrating on high-reward regions at the cost of other objectives. We therefore add an attribute group-relative regularization that computes advantages within attribute groups and normalizes the policy over the same groups, so uniformly promoting an entire high-reward group yields no optimization gain. Offline, UniRec consistently outperforms single-stage fusion and cross-stage coordination baselines. Online A/B tests show a 0.616\% gain in app usage duration. UniRec is fully deployed on the Kuaishou platform.