Hansi Zeng

Bowen Jin, Hansi Zeng, Guoyin Wang et al.

Semantic identifier (ID) is an important concept in information retrieval that aims to preserve the semantics of objects such as documents and items inside their IDs. Previous studies typically adopt a two-stage pipeline to learn semantic IDs by first procuring embeddings using off-the-shelf text encoders and then deriving IDs based on the embeddings. However, each step introduces potential information loss, and there is usually an inherent mismatch between the distribution of embeddings within the latent space produced by text encoders and the anticipated distribution required for semantic indexing. It is non-trivial to design a method that can learn the document's semantic representations and its hierarchical structure simultaneously, given that semantic IDs are discrete and sequentially structured, and the semantic supervision is deficient. In this paper, we introduce LMIndexer, a self-supervised framework to learn semantic IDs with a generative language model. We tackle the challenge of sequential discrete ID by introducing a semantic indexer capable of generating neural sequential discrete representations with progressive training and contrastive learning. In response to the semantic supervision deficiency, we propose to train the model with a self-supervised document reconstruction objective. We show the high quality of the learned IDs and demonstrate their effectiveness on three tasks including recommendation, product search, and document retrieval on five datasets from various domains. Code is available at https://github.com/PeterGriffinJin/LMIndexer.

Hansi Zeng, Hamed Zamani, Vishwa Vinay

Recent work has shown that more effective dense retrieval models can be obtained by distilling ranking knowledge from an existing base re-ranking model. In this paper, we propose a generic curriculum learning based optimization framework called CL-DRD that controls the difficulty level of training data produced by the re-ranking (teacher) model. CL-DRD iteratively optimizes the dense retrieval (student) model by increasing the difficulty of the knowledge distillation data made available to it. In more detail, we initially provide the student model coarse-grained preference pairs between documents in the teacher's ranking and progressively move towards finer-grained pairwise document ordering requirements. In our experiments, we apply a simple implementation of the CL-DRD framework to enhance two state-of-the-art dense retrieval models. Experiments on three public passage retrieval datasets demonstrate the effectiveness of our proposed framework.

Bowen Jin, Hansi Zeng, Zhenrui Yue et al.

Efficiently acquiring external knowledge and up-to-date information is essential for effective reasoning and text generation in large language models (LLMs). Prompting advanced LLMs with reasoning capabilities to use search engines during inference is often suboptimal, as the LLM might not fully possess the capability on how to interact optimally with the search engine. This paper introduces Search-R1, an extension of reinforcement learning (RL) for reasoning frameworks where the LLM learns to autonomously generate (multiple) search queries during step-by-step reasoning with real-time retrieval. Search-R1 optimizes LLM reasoning trajectories with multi-turn search interactions, leveraging retrieved token masking for stable RL training and a simple outcome-based reward function. Experiments on seven question-answering datasets show that Search-R1 improves performance by 41% (Qwen2.5-7B) and 20% (Qwen2.5-3B) over various RAG baselines under the same setting. This paper further provides empirical insights into RL optimization methods, LLM choices, and response length dynamics in retrieval-augmented reasoning. The code and model checkpoints are available at https://github.com/PeterGriffinJin/Search-R1.

Mingchen Li, Hansi Zeng, Zhuo Qian et al.

Retrieval is increasingly moving from one-shot matching toward interactive reasoning, where language agents iteratively inspect evidence, reformulate queries, and search again. Training such agents raises a credit-assignment challenge: executable actions such as queries or summaries can be directly evaluated by the retriever, while latent reasoning steps are not directly observable and only affect future executable actions. This asymmetry makes outcome-level reward assignment unreliable, as the same final reward may credit reasoning steps that did not actually shape retrieval success. We propose RICE-PO, a critic-free policy optimization framework that converts retrieval interactions into localized learning signals. RICE-PO selects high-uncertainty executable actions as anchors, evaluates local counterfactual branches using retrieval metrics, and propagates credit to latent reasoning steps only when reasoning-to-action influence is strong and future residual effects are stable. On BRIGHT and BEIR, RICE-PO consistently outperforms prompt-based agents and group-based RL baselines under the same retriever setting. These results show that the structure of agent-environment interaction itself can provide useful supervision for training reasoning-based retrieval agents.

Reshmi Ghosh, Harjeet Singh Kajal, Sharanya Kamath et al.

Breaking down a document or a conversation into multiple contiguous segments based on its semantic structure is an important and challenging problem in NLP, which can assist many downstream tasks. However, current works on topic segmentation often focus on segmentation of structured texts. In this paper, we comprehensively analyze the generalization capabilities of state-of-the-art topic segmentation models on unstructured texts. We find that: (a) Current strategies of pre-training on a large corpus of structured text such as Wiki-727K do not help in transferability to unstructured conversational data. (b) Training from scratch with only a relatively small-sized dataset of the target unstructured domain improves the segmentation results by a significant margin. We stress-test our proposed Topic Segmentation approach by experimenting with multiple loss functions, in order to mitigate effects of imbalance in unstructured conversational datasets. Our empirical evaluation indicates that Focal Loss function is a robust alternative to Cross-Entropy and re-weighted Cross-Entropy loss function when segmenting unstructured and semi-structured chats.

Junda Wang, Zonghai Tao, Hansi Zeng et al.

Complex clinical decision making often fails not because a model lacks facts, but because it cannot reliably select and apply the right procedural knowledge and the right prior example at the right reasoning step. We frame clinical question answering as an agent problem with two explicit, retrievable resources: skills, reusable clinical procedures such as guidelines, protocols, and pharmacologic mechanisms; and experience, verified reasoning trajectories from previously solved cases (e.g., chain-of-thought solutions and their step-level decompositions). At test time, the agent retrieves both relevant skills and experiences from curated libraries and performs lightweight test-time adaptation to align the language model's intermediate reasoning with clinically valid logic. Concretely, we build (i) a skills library from guideline-style documents organized as executable decision rules, (ii) an experience library of exemplar clinical reasoning chains indexed by step-level transitions, and (iii) a step-aware retriever that selects the most useful skill and experience items for the current case. We then adapt the model on the retrieved items to reduce instance-step misalignment and to prevent reasoning from drifting toward unsupported shortcuts. Experiments on medical question-answering benchmarks show consistent gains over strong medical RAG baselines and prompting-only reasoning methods. Our results suggest that explicitly separating and retrieving clinical skills and experience, and then aligning the model at test time, is a practical approach to more reliable medical agents.

Hansi Zeng, Liam Collins, Bhuvesh Kumar et al.

Agentic search -- the task of training agents that iteratively reason, issue queries, and synthesize retrieved information to answer complex questions -- has achieved remarkable progress through reinforcement learning (RL). However, existing approaches such as Search-R1, treat the retrieval system as a fixed tool, optimizing only the reasoning agent while the retrieval component remains unchanged. A preliminary experiment reveals that the gap between an oracle and a fixed retrieval system reaches up to +26.8% relative F1 improvement across seven QA benchmarks, suggesting that the retrieval system is a key bottleneck in scaling agentic search performance. Motivated by this finding, we propose CoSearch, a framework that jointly trains a multi-step reasoning agent and a generative document ranking model via Group Relative Policy Optimization (GRPO). To enable effective GRPO training for the ranker -- whose inputs vary across reasoning trajectories -- we introduce a semantic grouping strategy that clusters sub-queries by token-level similarity, forming valid optimization groups without additional rollouts. We further design a composite reward combining ranking quality signals with trajectory-level outcome feedback, providing the ranker with both immediate and long-term learning signals. Experiments on seven single-hop and multi-hop QA benchmarks demonstrate consistent improvements over strong baselines, with ablation studies validating each design choice. Our results show that joint training of the reasoning agent and retrieval system is both feasible and strongly performant, pointing to a key ingredient for future search agents.

Simeng Han, Frank Palma Gomez, Tu Vu et al.

Traditional text embedding benchmarks primarily evaluate embedding models' capabilities to capture semantic similarity. However, more advanced NLP tasks require a deeper understanding of text, such as safety and factuality. These tasks demand an ability to comprehend and process complex information, often involving the handling of sensitive content, or the verification of factual statements against reliable sources. We introduce a new benchmark designed to assess and highlight the limitations of embedding models trained on existing information retrieval data mixtures on advanced capabilities, which include factuality, safety, instruction following, reasoning and document-level understanding. This benchmark includes a diverse set of tasks that simulate real-world scenarios where these capabilities are critical and leads to identification of the gaps of the currently advanced embedding models. Furthermore, we propose a novel method that reformulates these various tasks as retrieval tasks. By framing tasks like safety or factuality classification as retrieval problems, we leverage the strengths of retrieval models in capturing semantic relationships while also pushing them to develop a deeper understanding of context and content. Using this approach with single-task fine-tuning, we achieved performance gains of 8\% on factuality classification and 13\% on safety classification. Our code and data will be publicly available.

Hansi Zeng, Zoey Li, Yifan Gao et al.

Research Agents enable models to gather information from the web using tools to answer user queries, requiring them to dynamically interleave internal reasoning with tool use. While such capabilities can in principle be learned via reinforcement learning with verifiable rewards (RLVR), we observe that agents often exhibit poor exploration behaviors, including premature termination and biased tool usage. As a result, RLVR alone yields limited improvements. We propose SynPlanResearch-R1, a framework that synthesizes tool-use trajectories that encourage deeper exploration to shape exploration during cold-start supervised fine-tuning, providing a strong initialization for subsequent RL. Across seven multi-hop and open-web benchmarks, \framework improves performance by up to 6.0% on Qwen3-8B and 5.8% on Qwen3-4B backbones respectively compared to SOTA baselines. Further analyses of tool-use patterns and training dynamics compared to baselines shed light on the factors underlying these gains. Our code is publicly available at https://github.com/HansiZeng/syn-plan-research.

Jiacheng Lin, Kun Qian, Arvind Srinivasan et al.

Personalization today is fundamentally platform-centric: services build user representations from the behavioral fragments they observe. Yet no platform can construct a complete picture of the user, as competitive incentives, legal constraints, user privacy concerns, and epistemic limits create persistent data barriers. This paper argues for a shift from platform-centric personalization to user-governed personalization, where only the user can integrate fragmented contexts across platforms and the offline world. The key asymmetry lies in data access: only users can aggregate their own cross-platform and offline information. Large language model (LLM) agents make such integration practically feasible for the first time by enabling reasoning over heterogeneous personal data and transforming users' cross-context information into actionable personalization capabilities. We provide proof-of-concept evidence that users equipped with cross-platform data exports and an off-the-shelf LLM agent can outperform single-platform personalization baselines. We conclude by outlining a research agenda for building scalable user-governed personalization systems.

Tianxin Wei, Bowen Jin, Ruirui Li et al.

Developing a universal model that can effectively harness heterogeneous resources and respond to a wide range of personalized needs has been a longstanding community aspiration. Our daily choices, especially in domains like fashion and retail, are substantially shaped by multi-modal data, such as pictures and textual descriptions. These modalities not only offer intuitive guidance but also cater to personalized user preferences. However, the predominant personalization approaches mainly focus on the ID or text-based recommendation problem, failing to comprehend the information spanning various tasks or modalities. In this paper, our goal is to establish a Unified paradigm for Multi-modal Personalization systems (UniMP), which effectively leverages multi-modal data while eliminating the complexities associated with task- and modality-specific customization. We argue that the advancements in foundational generative modeling have provided the flexibility and effectiveness necessary to achieve the objective. In light of this, we develop a generic and extensible personalization generative framework, that can handle a wide range of personalized needs including item recommendation, product search, preference prediction, explanation generation, and further user-guided image generation. Our methodology enhances the capabilities of foundational language models for personalized tasks by seamlessly ingesting interleaved cross-modal user history information, ensuring a more precise and customized experience for users. To train and evaluate the proposed multi-modal personalized tasks, we also introduce a novel and comprehensive benchmark covering a variety of user requirements. Our experiments on the real-world benchmark showcase the model's potential, outperforming competitive methods specialized for each task.

Hansi Zeng, Chen Luo, Hamed Zamani

This paper introduces PAG-a novel optimization and decoding approach that guides autoregressive generation of document identifiers in generative retrieval models through simultaneous decoding. To this aim, PAG constructs a set-based and sequential identifier for each document. Motivated by the bag-of-words assumption in information retrieval, the set-based identifier is built on lexical tokens. The sequential identifier, on the other hand, is obtained via quantizing relevance-based representations of documents. Extensive experiments on MSMARCO and TREC Deep Learning Track data reveal that PAG outperforms the state-of-the-art generative retrieval model by a large margin (e.g., 15.6% MRR improvements on MS MARCO), while achieving 22x speed up in terms of query latency.

Julian Killingback, Hansi Zeng, Hamed Zamani

Existing information retrieval systems are largely constrained by their reliance on vector inner products to assess query-document relevance, which naturally limits the expressiveness of the relevance score they can produce. We propose a new paradigm; instead of representing a query as a vector, we use a small neural network that acts as a learned query-specific relevance function. This small neural network takes a document representation as input (in this work we use a single vector) and produces a scalar relevance score. To produce the small neural network we use a hypernetwork, a network that produces the weights of other networks, as our query encoder. We name this category of encoder models Hypencoders. Experiments on in-domain search tasks show that Hypencoders significantly outperform strong dense retrieval models and even surpass reranking models and retrieval models with an order of magnitude more parameters. To assess the extent of Hypencoders' capabilities, we evaluate on a set of hard retrieval tasks including tip-of-the-tongue and instruction-following retrieval tasks. On harder tasks, we find that the performance gap widens substantially compared to standard retrieval tasks. Furthermore, to demonstrate the practicality of our method, we implement an approximate search algorithm and show that our model is able to retrieve from a corpus of 8.8M documents in under 60 milliseconds.

Chris Samarinas, Pracha Promthaw, Atharva Nijasure et al.

This paper explores SynTOD, a new synthetic data generation approach for developing end-to-end Task-Oriented Dialogue (TOD) Systems capable of handling complex tasks such as intent classification, slot filling, conversational question-answering, and retrieval-augmented response generation, without relying on crowdsourcing or real-world data. SynTOD utilizes a state transition graph to define the desired behavior of a TOD system and generates diverse, structured conversations through random walks and response simulation using large language models (LLMs). In our experiments, using graph-guided response simulations leads to significant improvements in intent classification, slot filling and response relevance compared to naive single-prompt simulated conversations. We also investigate the end-to-end TOD effectiveness of different base and instruction-tuned LLMs, with and without the constructed synthetic conversations. Finally, we explore how various LLMs can evaluate responses in a TOD system and how well they are correlated with human judgments. Our findings pave the path towards quick development and evaluation of domain-specific TOD systems. We release our datasets, models, and code for research purposes.

Jiacheng Lin, Zhongruo Wang, Kun Qian et al.

Supervised Fine-Tuning (SFT) on domain-specific datasets is a common approach to adapt Large Language Models (LLMs) to specialized tasks but is often believed to degrade their general capabilities. In this work, we revisit this trade-off and present both empirical and theoretical insights. First, we show that SFT does not always hurt: using a smaller learning rate can substantially mitigate general performance degradation while preserving comparable target-domain performance. We then provide a theoretical analysis that explains these phenomena and further motivates a new method, Token-Adaptive Loss Reweighting (TALR). Building on this, and recognizing that smaller learning rates alone do not fully eliminate general-performance degradation in all cases, we evaluate a range of strategies for reducing general capability loss, including L2 regularization, LoRA, model averaging, FLOW, and our proposed TALR. Experimental results demonstrate that while no method completely eliminates the trade-off, TALR consistently outperforms these baselines in balancing domain-specific gains and general capabilities. Finally, we distill our findings into practical guidelines for adapting LLMs to new domains: (i) using a small learning rate to achieve a favorable trade-off, and (ii) when a stronger balance is further desired, adopt TALR as an effective strategy.

Fengran Mo, Yifan Gao, Sha Li et al.

Large Language Models (LLMs) have become a popular interface for human-AI interaction, supporting information seeking and task assistance through natural, multi-turn dialogue. To respond to users within multi-turn dialogues, the context-dependent user intent evolves across interactions, requiring contextual interpretation, query reformulation, and dynamic coordination between retrieval and generation. Existing studies usually follow static rewrite, retrieve, and generate pipelines, which optimize different procedures separately and overlook the mixed-initiative action optimization simultaneously. Although the recent developments in deep search agents demonstrate the effectiveness in jointly optimizing retrieval and generation via reasoning, these approaches focus on single-turn scenarios, which might lack the ability to handle multi-turn interactions. We introduce a conversational agent that interleaves search and reasoning across turns, enabling exploratory and adaptive behaviors learned through reinforcement learning (RL) training with tailored rewards towards evolving user goals. The experimental results across four widely used conversational benchmarks demonstrate the effectiveness of our methods by surpassing several existing strong baselines.

Hansi Zeng, Kai Hui, Honglei Zhuang et al. · deepmind

While metrics available during pre-training, such as perplexity, correlate well with model performance at scaling-laws studies, their predictive capacities at a fixed model size remain unclear, hindering effective model selection and development. To address this gap, we formulate the task of selecting pre-training checkpoints to maximize downstream fine-tuning performance as a pairwise classification problem: predicting which of two LLMs, differing in their pre-training, will perform better after supervised fine-tuning (SFT). We construct a dataset using 50 1B parameter LLM variants with systematically varied pre-training configurations, e.g., objectives or data, and evaluate them on diverse downstream tasks after SFT. We first conduct a study and demonstrate that the conventional perplexity is a misleading indicator. As such, we introduce novel unsupervised and supervised proxy metrics derived from pre-training that successfully reduce the relative performance prediction error rate by over 50%. Despite the inherent complexity of this task, we demonstrate the practical utility of our proposed proxies in specific scenarios, paving the way for more efficient design of pre-training schemes optimized for various downstream tasks.

Zhiqi Huang, Hansi Zeng, Hamed Zamani et al.

In this work, we explore a Multilingual Information Retrieval (MLIR) task, where the collection includes documents in multiple languages. We demonstrate that applying state-of-the-art approaches developed for cross-lingual information retrieval to MLIR tasks leads to sub-optimal performance. This is due to the heterogeneous and imbalanced nature of multilingual collections -- some languages are better represented in the collection and some benefit from large-scale training data. To address this issue, we present KD-SPD, a novel soft prompt decoding approach for MLIR that implicitly "translates" the representation of documents in different languages into the same embedding space. To address the challenges of data scarcity and imbalance, we introduce a knowledge distillation strategy. The teacher model is trained on rich English retrieval data, and by leveraging bi-text data, our distillation framework transfers its retrieval knowledge to the multilingual document encoder. Therefore, our approach does not require any multilingual retrieval training data. Extensive experiments on three MLIR datasets with a total of 15 languages demonstrate that KD-SPD significantly outperforms competitive baselines in all cases. We conduct extensive analyses to show that our method has less language bias and better zero-shot transfer ability towards new languages.

Zhichao Xu, Hansi Zeng, Qingyao Ai

Modern E-commerce websites contain heterogeneous sources of information, such as numerical ratings, textual reviews and images. These information can be utilized to assist recommendation. Through textual reviews, a user explicitly express her affinity towards the item. Previous researchers found that by using the information extracted from these reviews, we can better profile the users' explicit preferences as well as the item features, leading to the improvement of recommendation performance. However, most of the previous algorithms were only utilizing the review information for explicit-feedback problem i.e. rating prediction, and when it comes to implicit-feedback ranking problem such as top-N recommendation, the usage of review information has not been fully explored. Seeing this gap, in this work, we investigate the effectiveness of textual review information for top-N recommendation under E-commerce settings. We adapt several SOTA review-based rating prediction models for top-N recommendation tasks and compare them to existing top-N recommendation models from both performance and efficiency. We find that models utilizing only review information can not achieve better performances than vanilla implicit-feedback matrix factorization method. When utilizing review information as a regularizer or auxiliary information, the performance of implicit-feedback matrix factorization method can be further improved. However, the optimal model structure to utilize textual reviews for E-commerce top-N recommendation is yet to be determined.

Hansi Zeng, Zhichao Xu, Qingyao Ai

User and item reviews are valuable for the construction of recommender systems. In general, existing review-based methods for recommendation can be broadly categorized into two groups: the siamese models that build static user and item representations from their reviews respectively, and the interaction-based models that encode user and item dynamically according to the similarity or relationships of their reviews. Although the interaction-based models have more model capacity and fit human purchasing behavior better, several problematic model designs and assumptions of the existing interaction-based models lead to its suboptimal performance compared to existing siamese models. In this paper, we identify three problems of the existing interaction-based recommendation models and propose a couple of solutions as well as a new interaction-based model to incorporate review data for rating prediction. Our model implements a relevance matching model with regularized training losses to discover user relevant information from long item reviews, and it also adapts a zero attention strategy to dynamically balance the item-dependent and item-independent information extracted from user reviews. Empirical experiments and case studies on Amazon Product Benchmark datasets show that our model can extract effective and interpretable user/item representations from their reviews and outperforms multiple types of state-of-the-art review-based recommendation models.

Hansi Zeng, Qingyao Ai

Using reviews to learn user and item representations is important for recommender system. Current review based methods can be divided into two categories: (1) the Convolution Neural Network (CNN) based models that extract n-gram features from user/item reviews; (2) the Recurrent Neural Network (RNN) based models that learn global contextual representations from reviews for users and items. Despite their success, both CNN and RNN based models in previous studies suffer from their own drawbacks. While CNN based models are weak in modeling long-dependency relation in text, RNN based models are slow in training and inference due to their incapability with parallel computing. To alleviate these problems, we propose a new text encoder module for review modeling in recommendation by combining convolution networks with self-attention networks to model local and global interactions in text together.As different words, sentences, reviews have different importance for modeling user and item representations, we construct review models hierarchically in sentence-level, review-level, and user/item level by encoding words for sentences, encoding sentences for reviews, and encoding reviews for user and item representations. Experiments on Amazon Product Benchmark show that our model can achieve significant better performance comparing to the state-of-the-art review based recommendation models.