Huangyu Dai

Zihan Liang, Yufei Ma, Ben Chen et al.

Post-training has become the dominant recipe for turning a language model into a competent search-augmented reasoning agent. A line of recent work pushes its performance further by adding elaborate machinery on top of this standard pipeline. These augmentations import external supervision from stronger external systems, attach auxiliary modules such as process reward models or retrospective critics, restructure the rollout itself with tree search or multi-stage curricula, or shape the reward with hand-crafted bonuses and penalties. Each addition delivers a measurable gain, but each also inflates the training pipeline and ties the recipe to resources or designs that may not always be available. We take a step back and ask whether any of this machinery is actually necessary, and propose Search-E1, a self-evolution method that lets a search-augmented agent improve through only vanilla GRPO interleaved with offline self-distillation (OFSD). After each GRPO round, the policy rolls out on its own training questions. A token-level forward KL objective then aligns the policy's inference-time distribution to its own distribution under a privileged context that exposes a more efficient sibling trajectory. Despite this simplicity, the procedure naturally provides dense per-step supervision. On seven QA benchmarks, Search-E1 reaches $0.440$ average EM with Qwen2.5-3B, surpassing all open-source baselines at both scales. Code and complete version will be made public soon.

Lingtao Mao, Huangyu Dai, Xinyu Sun et al.

Multimodal large language models are increasingly used as agent backbones that understand multimodal inputs, plan retrieval actions, invoke external tools, and reason over retrieved information. Yet existing benchmarks rarely evaluate this ability in short-video applications, where a paused frame is often visually ambiguous and answering requires vertical, long-tail, and fast-evolving domain knowledge. We introduce SVFSearch, the first open benchmark for short-video frame search in the Chinese gaming domain. SVFSearch contains 5,000 four-choice test examples and 4,198 auxiliary training examples, each centered on a paused game scene from a real short-video clip. To support fair and reproducible evaluation, SVFSearch provides a frozen offline retrieval environment with a game-domain text corpus, a topic-linked image gallery, and text, image, and multimodal retrieval interfaces, avoiding reliance on uncontrolled web search APIs. We evaluate representative paradigms ranging from direct QA and RAG workflow to Plan-Act-Replan agents and learned search models. Results reveal a large gap between model-only answering, practical agentic search, and oracle knowledge: the best open-source direct-QA model reaches 66.4%, the best practical agent achieves 79.1%, and oracle knowledge reaches 95.4%. Further analysis exposes bottlenecks in visual grounding, retrieval quality, evidence-grounded reasoning, and tool-use behavior, including over-search, answer-only shortcuts, and retrieval-induced misleading.

Zhipeng Qian, Zihan Liang, Yufei Ma et al.

Training large language models as retrieval-augmented reasoning agents typically combines reinforcement learning with an SFT cold start distilled from a stronger model. However, this paradigm overlooks two fundamental factors: the dependency structure among sub-skills, and the possibility that distillation is not the only route to capability acquisition. We study this through Plan, a structured agentic behavior for multi-hop retrieval that decomposes a question into ordered sub-questions before any retrieval is performed, so that each search step can be anchored to a pre-designed sub-question instead of drifting under the influence of partially relevant documents retrieved earlier. However, across three model families spanning 3B to 14B parameters, we find that an identical reward signal induces qualitatively different RL failure modes. This phenomenon indicates that successful training hinges not only on reward design but also on model-specific feasibility conditions: sufficient initial entropy, training stability, and prerequisite sub-skills. Motivated by this, we propose a self-bootstrapping paradigm in which a small-scale seed model generates filtered trajectories that activate Plan in any target model, eliminating the need for distillation from an external stronger model. Our pipeline activates Plan across every tested model and consistently outperforms competitive baselines on multi-hop QA benchmarks.

Ben Chen, Siyuan Wang, Yufei Ma et al.

Generative Retrieval (GR) has emerged as a promising paradigm for modern search systems. Compared to multi-stage cascaded architecture, it offers advantages such as end-to-end joint optimization and high computational efficiency. OneSearch, as a representative industrial-scale deployed generative search framework, has brought significant commercial and operational benefits. However, its inadequate understanding of complex queries, inefficient exploitation of latent user intents, and overfitting to narrow historical preferences have limited its further performance improvement. To address these challenges, we propose \textbf{OneSearch-V2}, a latent reasoning enhanced self-distillation generative search framework. It contains three key innovations: (1) a thought-augmented complex query understanding module, which enables deep query understanding and overcomes the shallow semantic matching limitations of direct inference; (2) a reasoning-internalized self-distillation training pipeline, which uncovers users' potential yet precise e-commerce intentions beyond log-fitting through implicit in-context learning; (3) a behavior preference alignment optimization system, which mitigates reward hacking arising from the single conversion metric, and addresses personal preference via direct user feedback. Extensive offline evaluations demonstrate OneSearch-V2's strong query recognition and user profiling capabilities. Online A/B tests further validate its business effectiveness, yielding +3.98\% item CTR, +3.05\% buyer conversion rate, and +2.11\% order volume. Manual evaluation further confirms gains in search experience quality, with +1.65\% in page good rate and +1.37\% in query-item relevance. More importantly, OneSearch-V2 effectively mitigates common search system issues such as information bubbles and long-tail sparsity, without incurring additional inference costs or serving latency.

Zhipeng Qian, Zihan Liang, Yufei Ma et al.

Composed Image Retrieval (CIR) enables users to search for target images using both a reference image and manipulation text, offering substantial advantages over single-modality retrieval systems. However, existing CIR methods suffer from representation space fragmentation: queries and targets comprise heterogeneous modalities and are processed by distinct encoders, forcing models to bridge misaligned representation spaces only through post-hoc alignment, which fundamentally limits retrieval performance. This architectural asymmetry manifests as three distinct, well-separated clusters in the feature space, directly demonstrating how heterogeneous modalities create fundamentally misaligned representation spaces from initialization. In this work, we propose CSMCIR, a unified representation framework that achieves efficient query-target alignment through three synergistic components. First, we introduce a Multi-level Chain-of-Thought (MCoT) prompting strategy that guides Multimodal Large Language Models to generate discriminative, semantically compatible captions for target images, establishing modal symmetry. Building upon this, we design a symmetric dual-tower architecture where both query and target sides utilize the identical shared-parameter Q-Former for cross-modal encoding, ensuring consistent feature representations and further reducing the alignment gap. Finally, this architectural symmetry enables an entropy-based, temporally dynamic Memory Bank strategy that provides high-quality negative samples while maintaining consistency with the evolving model state. Extensive experiments on four benchmark datasets demonstrate that our CSMCIR achieves state-of-the-art performance with superior training efficiency. Comprehensive ablation studies further validate the effectiveness of each proposed component.

Xinyu Sun, Huangyu Dai, Lingtao Mao et al.

E-commerce image search often takes a cropped image as the query, while each candidate is represented by full item images and structured text. This image-to-multimodal retrieval setting presents two asymmetries: a modality disparity -- a visual query must match image--text items, and a granularity disparity -- a cropped query must be compared with full images containing background context and possible distractors. Detection-based pipelines handle the granularity disparity through explicit localization but incur extra cost and error propagation, whereas CLIP-style encoders avoid detection, but are vulnerable to backgrounds or irrelevant items. To address these limitations, we propose TIGER-FG, a text-guided implicit fine-grained grounding framework for image-to-multimodal e-commerce retrieval. TIGER-FG uses item text as semantic guidance to produce target-focused item representations without object detection for retrieval. We further introduce dual distillation objectives that preserve target-region spatial consistency and query--item similarity structure, yielding more stable and discriminative multimodal representations. In addition, we construct ECom-RF-IMMR, a realistic benchmark suite with a 10M-pair training set and two evaluation benchmarks covering standard and cluttered item layouts. TIGER-FG improves Recall@1 over the strongest baseline by 6.1 and 34.4 percentage points on the two evaluation benchmarks, respectively, with only 85.7M query-side parameters and 256-dim embeddings. Results on public e-commerce benchmarks further demonstrate its generalization to noisy and one-to-many retrieval scenarios. Code and data will be released.

Yufei Ma, Zihan Liang, Ben Chen et al.

Search-augmented reasoning agents interleave internal reasoning with calls to an external retriever, and their performance relies on the quality of each issued query. However, under outcome-reward reinforcement learning, every search decision in a rollout shares the same trajectory-level reward, leaving individual queries without step-specific credit. Recent process-supervision approaches address this gap by drawing step-level signals from outside the policy, relying either on a much larger teacher model, or on sub-question annotations produced by a stronger external system. In contrast, we propose SD-Search, which derives step-level supervision from the policy itself through on-policy hindsight self-distillation, requiring neither an external teacher nor additional annotations. In SD-Search, a single model plays two roles that differ only in conditioning: a student that sees only the context available at inference time, and a teacher that additionally conditions on a compact hindsight block summarizing the search queries and final outcomes of a group of rollouts sampled from the same question. Since the teacher knows how each rollout unfolded and which ones succeeded, its query distribution implicitly marks which decisions were worth making, and the student is trained to recover this behavior by minimizing the token-level Jensen--Shannon divergence to the teacher at search-query positions. This layers a dense, step-level signal on top of GRPO's coarse trajectory reward. Crucially, this signal is produced by the policy itself within the standard RL training loop, without external model inference, auxiliary annotation pipeline, or additional training stage.

Zihan Liang, Yufei Ma, Ben Chen et al.

Reinforcement learning has emerged as an effective paradigm for training large language models to perform search-augmented reasoning. However, existing approaches rely on trajectory-level rewards that cannot distinguish precise search queries from vague or redundant ones within a rollout group, and collapse to a near-zero gradient signal whenever every sampled trajectory fails. In this paper, we propose IG-Search, a reinforcement learning framework that introduces a step-level reward based on Information Gain (IG). For each search step, IG measures how much the retrieved documents improve the model's confidence in the gold answer relative to a counterfactual baseline of random documents, thereby reflecting the effectiveness of the underlying search query. This signal is fed back to the corresponding search-query tokens via per-token advantage modulation in GRPO, enabling fine-grained, step-level credit assignment within a rollout. Unlike prior step-level methods that require either externally annotated intermediate supervision or shared environment states across trajectories, IG-Search derives its signals from the policy's own generation probabilities, requiring no intermediate annotations beyond standard question-answer pairs. Experiments on seven single-hop and multi-hop QA benchmarks demonstrate that IG-Search achieves an average EM of 0.430 with Qwen2.5-3B, outperforming the strongest trajectory-level baseline (MR-Search) by 1.6 points and the step-level method GiGPO by 0.9 points on average across benchmarks, with particularly pronounced gains on multi-hop reasoning tasks. Despite introducing a dense step-level signal, IG-Search adds only ~6.4% to per-step training wall-clock time over the trajectory-level baseline and leaves inference latency unchanged, while still providing a meaningful gradient signal even when every sampled trajectory answers incorrectly.

Zihan Liang, Yufei Ma, ZhiPeng Qian et al.

Current e-commerce multimodal retrieval systems face two key limitations: they optimize for specific tasks with fixed modality pairings, and lack comprehensive benchmarks for evaluating unified retrieval approaches. To address these challenges, we introduce UniECS, a unified multimodal e-commerce search framework that handles all retrieval scenarios across image, text, and their combinations. Our work makes three key contributions. First, we propose a flexible architecture with a novel gated multimodal encoder that uses adaptive fusion mechanisms. This encoder integrates different modality representations while handling missing modalities. Second, we develop a comprehensive training strategy to optimize learning. It combines cross-modal alignment loss (CMAL), cohesive local alignment loss (CLAL), intra-modal contrastive loss (IMCL), and adaptive loss weighting. Third, we create M-BEER, a carefully curated multimodal benchmark containing 50K product pairs for e-commerce search evaluation. Extensive experiments demonstrate that UniECS consistently outperforms existing methods across four e-commerce benchmarks with fine-tuning or zero-shot evaluation. On our M-BEER bench, UniECS achieves substantial improvements in cross-modal tasks (up to 28\% gain in R@10 for text-to-image retrieval) while maintaining parameter efficiency (0.2B parameters) compared to larger models like GME-Qwen2VL (2B) and MM-Embed (8B). Furthermore, we deploy UniECS in the e-commerce search platform of Kuaishou Inc. across two search scenarios, achieving notable improvements in Click-Through Rate (+2.74\%) and Revenue (+8.33\%). The comprehensive evaluation demonstrates the effectiveness of our approach in both experimental and real-world settings. Corresponding codes, models and datasets will be made publicly available at https://github.com/qzp2018/UniECS.

Xinyu Nan, Lingtao Mao, Huangyu Dai et al.

Achieving visual semantic understanding requires a unified framework that simultaneously handles object detection, category prediction, and attribute recognition. However, current advanced approaches rely on global similarity and struggle to capture fine-grained category distinctions and category-specific attribute diversity, especially in large-scale e-commerce scenarios. To overcome these challenges, we introduce a detection-guided generative framework that predicts hierarchical category and attribute tokens. For each detected object, we extract refined ROI-level features and employ a BART-based generator to produce semantic tokens in a coarse-to-fine sequence covering category hierarchies and property-value pairs, with support for property-conditioned attribute recognition. Experiments on both large-scale proprietary e-commerce datasets and open-source datasets demonstrate that our approach significantly outperforms existing similarity-based pipelines and multi-stage classification systems, achieving stronger fine-grained recognition and more coherent unified inference.

Huangyu Dai, Ben Chen, Kaidi Chen et al.

For crosslingual conversation and trade, Neural Machine Translation (NMT) is pivotal yet faces persistent challenges with monotony and repetition in generated content. Traditional solutions that rely on penalizing text redundancy or token reoccurrence have shown limited efficacy, particularly for lengthy article and e-commerce descriptions with inherent redundancy, even with the advent of Large Language Models (LLMs). This paper investigates the underlying causes of textual repetition through the lens of information entropy, attributing the phenomenon to the elevated uncertainty within the input text. To address this, a novel algorithm named Contrastive Token Learning with Similarity Decay (CTSD) is introduced, which modulates the suppression of tokens dynamically, informed by varying attention weights and inter-token distances. Furthermore, an e-commerce dataset comprised of title texts of online real items is compiled and released susceptible to hallucination translations to benchmark the algorithm. Extensive evaluations demonstrate that CTSD significantly outperforms existing approaches in precision and generalizability. Additional online A/B testing underscores its practical value, showing marked improvements in user engagement and conversion. Notably, this method has been implemented with full traffic on eight multilingual sites of alibaba.com, the largest B2B e-commerce platform in the world.

Kaidi Chen, Ben Chen, Dehong Gao et al.

Existing Neural Machine Translation (NMT) models mainly handle translation in the general domain, while overlooking domains with special writing formulas, such as e-commerce and legal documents. Taking e-commerce as an example, the texts usually include amounts of domain-related words and have more grammar problems, which leads to inferior performances of current NMT methods. To address these problems, we collect two domain-related resources, including a set of term pairs (aligned Chinese-English bilingual terms) and a parallel corpus annotated for the e-commerce domain. Furthermore, we propose a two-step fine-tuning paradigm (named G2ST) with self-contrastive semantic enhancement to transfer one general NMT model to the specialized NMT model for e-commerce. The paradigm can be used for the NMT models based on Large language models (LLMs). Extensive evaluations on real e-commerce titles demonstrate the superior translation quality and robustness of our G2ST approach, as compared with state-of-the-art NMT models such as LLaMA, Qwen, GPT-3.5, and even GPT-4.

Yufei Ma, Zihan Liang, Huangyu Dai et al.

The growing demand for larger-scale models in the development of \textbf{L}arge \textbf{L}anguage \textbf{M}odels (LLMs) poses challenges for efficient training within limited computational resources. Traditional fine-tuning methods often exhibit instability in multi-task learning and rely heavily on extensive training resources. Here, we propose MoDULA (\textbf{M}ixture \textbf{o}f \textbf{D}omain-Specific and \textbf{U}niversal \textbf{L}oR\textbf{A}), a novel \textbf{P}arameter \textbf{E}fficient \textbf{F}ine-\textbf{T}uning (PEFT) \textbf{M}ixture-\textbf{o}f-\textbf{E}xpert (MoE) paradigm for improved fine-tuning and parameter efficiency in multi-task learning. The paradigm effectively improves the multi-task capability of the model by training universal experts, domain-specific experts, and routers separately. MoDULA-Res is a new method within the MoDULA paradigm, which maintains the model's general capability by connecting universal and task-specific experts through residual connections. The experimental results demonstrate that the overall performance of the MoDULA-Flan and MoDULA-Res methods surpasses that of existing fine-tuning methods on various LLMs. Notably, MoDULA-Res achieves more significant performance improvements in multiple tasks while reducing training costs by over 80\% without losing general capability. Moreover, MoDULA displays flexible pluggability, allowing for the efficient addition of new tasks without retraining existing experts from scratch. This progressive training paradigm circumvents data balancing issues, enhancing training efficiency and model stability. Overall, MoDULA provides a scalable, cost-effective solution for fine-tuning LLMs with enhanced parameter efficiency and generalization capability.

Zexin Zheng, Huangyu Dai, Lingtao Mao et al.

Traditional vision search, similar to search and recommendation systems, follows the multi-stage cascading architecture (MCA) paradigm to balance efficiency and conversion. Specifically, the query image undergoes feature extraction, recall, pre-ranking, and ranking stages, ultimately presenting the user with semantically similar products that meet their preferences. This multi-view representation discrepancy of the same object in the query and the optimization objective collide across these stages, making it difficult to achieve Pareto optimality in both user experience and conversion. In this paper, an end-to-end generative framework, OneVision, is proposed to address these problems. OneVision builds on VRQ, a vision-aligned residual quantization encoding, which can align the vastly different representations of an object across multiple viewpoints while preserving the distinctive features of each product as much as possible. Then a multi-stage semantic alignment scheme is adopted to maintain strong visual similarity priors while effectively incorporating user-specific information for personalized preference generation. In offline evaluations, OneVision performs on par with online MCA, while improving inference efficiency by 21% through dynamic pruning. In A/B tests, it achieves significant online improvements: +2.15% item CTR, +2.27% CVR, and +3.12% order volume. These results demonstrate that a semantic ID centric, generative architecture can unify retrieval and personalization while simplifying the serving pathway.

Zihan Wang, Zihan Liang, Zhou Shao et al.

Retrieval-Augmented Generation (RAG) has emerged as a promising approach to address key limitations of Large Language Models (LLMs), such as hallucination, outdated knowledge, and lacking reference. However, current RAG frameworks often struggle with identifying whether retrieved documents meaningfully contribute to answer generation. This shortcoming makes it difficult to filter out irrelevant or even misleading content, which notably impacts the final performance. In this paper, we propose Document Information Gain (DIG), a novel metric designed to quantify the contribution of retrieved documents to correct answer generation. DIG measures a document's value by computing the difference of LLM's generation confidence with and without the document augmented. Further, we introduce InfoGain-RAG, a framework that leverages DIG scores to train a specialized reranker, which prioritizes each retrieved document from exact distinguishing and accurate sorting perspectives. This approach can effectively filter out irrelevant documents and select the most valuable ones for better answer generation. Extensive experiments across various models and benchmarks demonstrate that InfoGain-RAG can significantly outperform existing approaches, on both single and multiple retrievers paradigm. Specifically on NaturalQA, it achieves the improvements of 17.9%, 4.5%, 12.5% in exact match accuracy against naive RAG, self-reflective RAG and modern ranking-based RAG respectively, and even an average of 15.3% increment on advanced proprietary model GPT-4o across all datasets. These results demonstrate the feasibility of InfoGain-RAG as it can offer a reliable solution for RAG in multiple applications.

Huangyu Dai, Lingtao Mao, Ben Chen et al.

Hotword customization is crucial in ASR to enhance the accuracy of domain-specific terms. It has been primarily driven by the advancements in traditional models and Audio large language models (LLMs). However, existing models often struggle with large-scale hotwords, as the recognition rate drops dramatically with the number of hotwords increasing. In this paper, we introduce a novel hotword customization system that utilizes a hotword pre-retrieval module (H-PRM) to identify the most relevant hotword candidate by measuring the acoustic similarity between the hotwords and the speech segment. This plug-and-play solution can be easily integrated into traditional models such as SeACo-Paraformer, significantly enhancing hotwords post-recall rate (PRR). Additionally, we incorporate H-PRM into Audio LLMs through a prompt-based approach, enabling seamless customization of hotwords. Extensive testing validates that H-PRM can outperform existing methods, showing a new direction for hotword customization in ASR.

Ben Chen, Huangyu Dai, Xiang Ma et al.

Semantic relevance calculation is crucial for e-commerce search engines, as it ensures that the items selected closely align with customer intent. Inadequate attention to this aspect can detrimentally affect user experience and engagement. Traditional text-matching techniques are prevalent but often fail to capture the nuances of search intent accurately, so neural networks now have become a preferred solution to processing such complex text matching. Existing methods predominantly employ representation-based architectures, which strike a balance between high traffic capacity and low latency. However, they exhibit significant shortcomings in generalization and robustness when compared to interaction-based architectures. In this work, we introduce a robust interaction-based modeling paradigm to address these shortcomings. It encompasses 1) a dynamic length representation scheme for expedited inference, 2) a professional terms recognition method to identify subjects and core attributes from complex sentence structures, and 3) a contrastive adversarial training protocol to bolster the model's robustness and matching capabilities. Extensive offline evaluations demonstrate the superior robustness and effectiveness of our approach, and online A/B testing confirms its ability to improve relevance in the same exposure position, resulting in more clicks and conversions. To the best of our knowledge, this method is the first interaction-based approach for large e-commerce search relevance calculation. Notably, we have deployed it for the entire search traffic on alibaba.com, the largest B2B e-commerce platform in the world.