Changlong Yu

Changlong Yu, Weiqi Wang, Xin Liu et al. · amazon-science

Understanding users' intentions in e-commerce platforms requires commonsense knowledge. In this paper, we present FolkScope, an intention knowledge graph construction framework to reveal the structure of humans' minds about purchasing items. As commonsense knowledge is usually ineffable and not expressed explicitly, it is challenging to perform information extraction. Thus, we propose a new approach that leverages the generation power of large language models~(LLMs) and human-in-the-loop annotation to semi-automatically construct the knowledge graph. LLMs first generate intention assertions via e-commerce-specific prompts to explain shopping behaviors, where the intention can be an open reason or a predicate falling into one of 18 categories aligning with ConceptNet, e.g., IsA, MadeOf, UsedFor, etc. Then we annotate plausibility and typicality labels of sampled intentions as training data in order to populate human judgments to all automatic generations. Last, to structurize the assertions, we propose pattern mining and conceptualization to form more condensed and abstract knowledge. Extensive evaluations and studies demonstrate that our constructed knowledge graph can well model e-commerce knowledge and have many potential applications.

Fangqi Zhu, Jun Gao, Changlong Yu et al.

Script event prediction aims to predict the subsequent event given the context. This requires the capability to infer the correlations between events. Recent works have attempted to improve event correlation reasoning by using pretrained language models and incorporating external knowledge~(e.g., discourse relations). Though promising results have been achieved, some challenges still remain. First, the pretrained language models adopted by current works ignore event-level knowledge, resulting in an inability to capture the correlations between events well. Second, modeling correlations between events with discourse relations is limited because it can only capture explicit correlations between events with discourse markers, and cannot capture many implicit correlations. To this end, we propose a novel generative approach for this task, in which a pretrained language model is fine-tuned with an event-centric pretraining objective and predicts the next event within a generative paradigm. Specifically, we first introduce a novel event-level blank infilling strategy as the learning objective to inject event-level knowledge into the pretrained language model, and then design a likelihood-based contrastive loss for fine-tuning the generative model. Instead of using an additional prediction layer, we perform prediction by using sequence likelihoods generated by the generative model. Our approach models correlations between events in a soft way without any external knowledge. The likelihood-based prediction eliminates the need to use additional networks to make predictions and is somewhat interpretable since it scores each word in the event. Experimental results on the multi-choice narrative cloze~(MCNC) task demonstrate that our approach achieves better results than other state-of-the-art baselines. Our code will be available at https://github.com/zhufq00/mcnc.

Jun Gao, Wei Wang, Changlong Yu et al.

Representations of events described in text are important for various tasks. In this work, we present SWCC: a Simultaneous Weakly supervised Contrastive learning and Clustering framework for event representation learning. SWCC learns event representations by making better use of co-occurrence information of events. Specifically, we introduce a weakly supervised contrastive learning method that allows us to consider multiple positives and multiple negatives, and a prototype-based clustering method that avoids semantically related events being pulled apart. For model training, SWCC learns representations by simultaneously performing weakly supervised contrastive learning and prototype-based clustering. Experimental results show that SWCC outperforms other baselines on Hard Similarity and Transitive Sentence Similarity tasks. In addition, a thorough analysis of the prototype-based clustering method demonstrates that the learned prototype vectors are able to implicitly capture various relations between events.

Yuwei Zhang, Chengyu Dong, Shuowei Jin et al.

Context management enables agentic models to solve long-horizon tasks through iterative summarization of previous interaction histories. However, this process typically incurs substantial decoding overhead for the extra summarization tokens, which significantly affect the end-to-end response latency at deployment. In this paper, we introduce CoMem, a novel framework that decouples memory management from the primary agent workflow, enabling these processes to execute in parallel. We propose a $k$-step-off asynchronous pipeline that overlaps the memory model's summarization with the agent's inference, effectively masking the latency of context processing. To ensure robustness under this asynchronous setting, we introduce a reward-driven training strategy that aligns the memory model to capture sufficient statistics for the agent's decision-making. Theoretical analysis confirms that CoMem offers a superior efficiency-effectiveness trade-off compared to coupled architectures. Our extensive experimental results on SWE-Bench-Verified show that CoMem provides 1.4x latency improvements upon vanilla long-context solutions while preserving most of the performance. Furthermore, we demonstrate that these latency gains scale favorably with increased system throughput, offering a modular path forward for the independent optimization of agent reasoning and memory compression.

Jun Gao, Huan Zhao, Changlong Yu et al.

Event extraction is a fundamental task in natural language processing that involves identifying and extracting information about events mentioned in text. However, it is a challenging task due to the lack of annotated data, which is expensive and time-consuming to obtain. The emergence of large language models (LLMs) such as ChatGPT provides an opportunity to solve language tasks with simple prompts without the need for task-specific datasets and fine-tuning. While ChatGPT has demonstrated impressive results in tasks like machine translation, text summarization, and question answering, it presents challenges when used for complex tasks like event extraction. Unlike other tasks, event extraction requires the model to be provided with a complex set of instructions defining all event types and their schemas. To explore the feasibility of ChatGPT for event extraction and the challenges it poses, we conducted a series of experiments. Our results show that ChatGPT has, on average, only 51.04% of the performance of a task-specific model such as EEQA in long-tail and complex scenarios. Our usability testing experiments indicate that ChatGPT is not robust enough, and continuous refinement of the prompt does not lead to stable performance improvements, which can result in a poor user experience. Besides, ChatGPT is highly sensitive to different prompt styles.

Weiqi Wang, Xin Liu, Binxuan Huang et al.

RLVR is now a standard way to train LLMs on reasoning tasks with verifiable outcomes, but when rollout generation dominates the cost, efficiency depends heavily on which prompts you sample and when. In practice, prompt pools are often static or only loosely tied to the model's learning progress, so uniform sampling can't keep up with the shifting capability frontier and ends up wasting rollouts on prompts that are already solved or still out of reach. Existing approaches improve efficiency through filtering, curricula, adaptive rollout allocation, or teacher guidance, but they typically assume a fixed pool-which makes it hard to support stable on-policy pool growth-or they add extra teacher cost and latency. We introduce HeaPA (Heap Sampling and On-Policy Query Augmentation), which maintains a bounded, evolving pool, tracks the frontier using heap-based boundary sampling, expands the pool via on-policy augmentation with lightweight asynchronous validation, and stabilizes correlated queries through topology-aware re-estimation of pool statistics and controlled reinsertion. Across two training corpora, two training recipes, and seven benchmarks, HeaPA consistently improves accuracy and reaches target performance with fewer computations while keeping wall-clock time comparable. Our analyses suggest these gains come from frontier-focused sampling and on-policy pool growth, with the benefits becoming larger as model scale increases. Our code is available at https://github.com/horizon-rl/HeaPA.

Jun Gao, Changlong Yu, Wei Wang et al.

We present Mask-then-Fill, a flexible and effective data augmentation framework for event extraction. Our approach allows for more flexible manipulation of text and thus can generate more diverse data while keeping the original event structure unchanged as much as possible. Specifically, it first randomly masks out an adjunct sentence fragment and then infills a variable-length text span with a fine-tuned infilling model. The main advantage lies in that it can replace a fragment of arbitrary length in the text with another fragment of variable length, compared to the existing methods which can only replace a single word or a fixed-length fragment. On trigger and argument extraction tasks, the proposed framework is more effective than baseline methods and it demonstrates particularly strong results in the low-resource setting. Our further analysis shows that it achieves a good balance between diversity and distributional similarity.

Zheye Deng, Weixiang Yan, Changlong Yu et al.

While Large Language Model (LLM) agents show promise in automated trading, they still face critical limitations. Prominent multi-agent frameworks often suffer from inefficiency, produce inconsistent signals, and lack the end-to-end optimization required to learn a coherent strategy from market feedback. To address this, we introduce AlphaQuanter, a single-agent framework that uses reinforcement learning (RL) to learn a dynamic policy over a transparent, tool-augmented decision workflow, which empowers a single agent to autonomously orchestrate tools and proactively acquire information on demand, establishing a transparent reasoning process. Extensive experiments demonstrate that AlphaQuanter achieves state-of-the-art performance on key financial metrics. Moreover, its interpretable reasoning reveals sophisticated strategies, offering novel and valuable insights for human traders. Our code and data can be found at https://github.com/horizon-llm/AlphaQuanter.

Changlong Yu, Tianyi Xiao, Lingpeng Kong et al.

Though linguistic knowledge emerges during large-scale language model pretraining, recent work attempt to explicitly incorporate human-defined linguistic priors into task-specific fine-tuning. Infusing language models with syntactic or semantic knowledge from parsers has shown improvements on many language understanding tasks. To further investigate the effectiveness of structural linguistic priors, we conduct empirical study of replacing parsed graphs or trees with trivial ones (rarely carrying linguistic knowledge e.g., balanced tree) for tasks in the GLUE benchmark. Encoding with trivial graphs achieves competitive or even better performance in fully-supervised and few-shot settings. It reveals that the gains might not be significantly attributed to explicit linguistic priors but rather to more feature interactions brought by fusion layers. Hence we call for attention to using trivial graphs as necessary baselines to design advanced knowledge fusion methods in the future.

Haolin Deng, Yanan Zhang, Yangfan Zhang et al.

Event extraction (EE) is crucial to downstream tasks such as new aggregation and event knowledge graph construction. Most existing EE datasets manually define fixed event types and design specific schema for each of them, failing to cover diverse events emerging from the online text. Moreover, news titles, an important source of event mentions, have not gained enough attention in current EE research. In this paper, We present Title2Event, a large-scale sentence-level dataset benchmarking Open Event Extraction without restricting event types. Title2Event contains more than 42,000 news titles in 34 topics collected from Chinese web pages. To the best of our knowledge, it is currently the largest manually-annotated Chinese dataset for open event extraction. We further conduct experiments on Title2Event with different models and show that the characteristics of titles make it challenging for event extraction, addressing the significance of advanced study on this problem. The dataset and baseline codes are available at https://open-event-hub.github.io/title2event.

Jun Gao, Huan Zhao, Yice Zhang et al.

Information Extraction (IE) is an essential task in Natural Language Processing. Traditional methods have relied on coarse-grained extraction with simple instructions. However, with the emergence of Large Language Models (LLMs), there is a need to adapt IE techniques to leverage the capabilities of these models. This paper introduces a fine-grained IE benchmark dataset tailored for LLMs, employing augmented instructions for each information type, which includes task descriptions, extraction rules, output formats, and examples. Through extensive evaluations, we observe that encoder-decoder models, particularly T5 and FLAN-T5, perform well in generalizing to unseen information types, while ChatGPT exhibits greater adaptability to new task forms. Our results also indicate that performance is not solely dictated by model scale, and highlight the significance of architecture, data diversity, and learning techniques. This work paves the way for a more refined and versatile utilization of LLMs in Information Extraction.

Yuheng Zhang, Chenlu Ye, Shuowei Jin et al.

Reinforcement learning, including reinforcement learning with verifiable rewards (RLVR), has emerged as a powerful approach for LLM post-training. Central to these approaches is the design of the importance sampling (IS) ratio used in off-policy policy-gradient estimation. Existing methods face a fundamental bias-variance dilemma: token-level IS ratios, as adopted by PPO (Schulman et al., 2017) and GRPO (Shao et al., 2024), introduce bias by ignoring prefix state distribution mismatch; full sequence ratios provide exact trajectory-level correction but suffer from high variance due to the multiplicative accumulation of per-token ratios, while GSPO (Zheng et al., 2025) improves numerical stability via length normalization at the cost of deviating from the exact full-sequence IS correction. In this work, we identify the cumulative token IS ratio, the product of per-token ratios up to position $t$, as a theoretically principled solution to this dilemma. We prove that, under the token-level policy-gradient formulation, this ratio provides an unbiased prefix correction for each token-level gradient term and has strictly lower variance than the full sequence ratio. Building on this insight, we propose CTPO (Cumulative Token Policy Optimization), which combines the cumulative token IS ratio with position-adaptive clipping that scales log-space clip bounds according to the natural $\sqrt{t}$ growth of the cumulative log-ratio. This yields more consistent regularization across token positions. We implement and evaluate CTPO in the tool-integrated reasoning setting on several challenging mathematical reasoning benchmarks, achieving the best average performance across both model scales compared with strong GRPO and GSPO baselines. Code will be available at https://github.com/horizon-llm/CTPO.

Zhenghao Xu, Qin Lu, Changlong Yu et al.

Policy mirror descent (PMD) provides a principled framework for reinforcement learning (RL) by iteratively solving KL-regularized policy improvement subproblems. While this approach has been adopted in training advanced LLMs such as Kimi K1.5/K2, the ideal closed-form PMD updates require reliable partition function estimation, a significant challenge when working with limited rollouts in the vast action spaces of LLMs. We investigate a practical algorithm, termed PMD-mean, that approximates the log-partition term with the mean reward under the sampling policy and performs regression in log-policy space. Specifically, we characterize the population solution of PMD-mean and demonstrate that it implicitly optimizes mirror descent subproblems with an adaptive mixed KL--$χ^2$ regularizer. This additional $χ^2$ regularization constrains large probability changes, producing more conservative updates when expected rewards are low and enhancing robustness against finite-sample estimation errors. Experiments on math reasoning tasks show that PMD-mean achieves superior performance with improved stability and time efficiency. These findings deepen our understanding of PMD-mean and illuminate pathways toward principled improvements in RL algorithms for LLMs. Code is available at https://github.com/horizon-rl/OpenKimi.

Peihao Wang, Shan Yang, Xijun Wang et al.

Associative memory has long underpinned the design of sequential models. Beyond recall, humans reason by projecting future states and selecting goal-directed actions, a capability that modern language models increasingly require but do not natively encode. While prior work uses reinforcement learning or test-time training, planning remains external to the model architecture. We formulate reasoning as optimal control and introduce the Test-Time Control (TTC) layer, which performs finite-horizon LQR planning over latent states at inference time, represents a value function within neural architectures, and leverages it as the nested objective to enable planning before prediction. To ensure scalability, we derive a hardware-efficient LQR solver based on a symplectic formulation and implement it as a fused CUDA kernel, enabling parallel execution with minimal overhead. Integrated as an adapter into pretrained LLMs, TTC layers improve mathematical reasoning performance by up to +27.8% on MATH-500 and 2-3x Pass@8 improvements on AMC and AIME, demonstrating that embedding optimal control as an architectural component provides an effective and scalable mechanism for reasoning beyond test-time training.

Chi Han, Xin Liu, Haodong Wang et al.

Despite significant achievements in improving the instruction-following capabilities of large language models (LLMs), the ability to process multiple potentially entangled or conflicting instructions remains a considerable challenge. Real-world scenarios often require consistency across multiple instructions over time, such as secret privacy, personal preferences, and prioritization, which demand sophisticated abilities to integrate multiple turns and carefully balance competing objectives when instructions intersect or conflict. This work presents a systematic investigation of LLMs' capabilities in handling multiple turns of instructions, covering three levels of difficulty: (1) retrieving information from instructions, (2) tracking and reasoning across turns, and (3) resolving conflicts among instructions. We construct MultiTurnInstruct~with $\sim$1.1K high-quality multi-turn conversations through the human-in-the-loop approach and result in nine capability categories, including statics and dynamics, reasoning, and multitasking. Our finding reveals an intriguing trade-off between different capabilities. While GPT models demonstrate superior memorization, they show reduced effectiveness in privacy-protection tasks requiring selective information withholding. Larger models exhibit stronger reasoning capabilities but still struggle with resolving conflicting instructions. Importantly, these performance gaps cannot be attributed solely to information loss, as models demonstrate strong BLEU scores on memorization tasks. Still, their attention mechanisms fail to integrate multiple related instructions effectively. These findings highlight critical areas for improvement in complex real-world tasks involving multi-turn instructions. Data and codes are released at https://github.com/Glaciohound/Multi-Turn-Instruct.

Yuwei Zhang, Sha Li, Changlong Yu et al.

Enabling Large Language Models (LLMs) to continuously improve from environmental interactions is a central challenge in post-training. While on-policy self-distillation offers a promising paradigm, existing methods predominantly treat environmental feedback as a passive conditioning signal. Consequently, they heavily rely on successful demonstrations and struggle to learn in rare-success regimes. To bridge this gap, we introduce Reflection-Enhanced Self-Distillation (RESD), a framework that transforms raw failure feedback into an active source of corrective supervision. Instead of passively appending feedback, RESD interprets failed trajectories by generating retrospective reflections to diagnose local errors, and curates a persistent global playbook to preserve reusable lessons across training steps. The enriched context enables the self-teacher to provide actionable token-level supervision even in the absence of successful rollouts. Empirical evaluations on multiple continual learning tasks demonstrate that RESD substantially outperforms standard self-distillation baselines. Furthermore, RESD achieves significantly faster early-stage improvement than GRPO with $8\times$ samples using only a single rollout per prompt, highlighting its superior interaction efficiency.

Wenxuan Ding, Weiqi Wang, Sze Heng Douglas Kwok et al.

Enhancing Language Models' (LMs) ability to understand purchase intentions in E-commerce scenarios is crucial for their effective assistance in various downstream tasks. However, previous approaches that distill intentions from LMs often fail to generate meaningful and human-centric intentions applicable in real-world E-commerce contexts. This raises concerns about the true comprehension and utilization of purchase intentions by LMs. In this paper, we present IntentionQA, a double-task multiple-choice question answering benchmark to evaluate LMs' comprehension of purchase intentions in E-commerce. Specifically, LMs are tasked to infer intentions based on purchased products and utilize them to predict additional purchases. IntentionQA consists of 4,360 carefully curated problems across three difficulty levels, constructed using an automated pipeline to ensure scalability on large E-commerce platforms. Human evaluations demonstrate the high quality and low false-negative rate of our benchmark. Extensive experiments across 19 language models show that they still struggle with certain scenarios, such as understanding products and intentions accurately, jointly reasoning with products and intentions, and more, in which they fall far behind human performances. Our code and data are publicly available at https://github.com/HKUST-KnowComp/IntentionQA.

Siqi Guo, Ilgee Hong, Vicente Balmaseda et al.

Supervised fine-tuning (SFT) has become a crucial step for aligning pretrained large language models (LLMs) using supervised datasets of input-output pairs. However, despite being supervised, SFT is inherently limited by its generative training objective. To address its limitations, the existing common strategy is to follow SFT with a separate phase of preference optimization (PO), which relies on either human-labeled preference data or a strong reward model to guide the learning process. In this paper, we address the limitations of SFT by exploring one of the most successful techniques in conventional supervised learning: discriminative learning. We introduce Discriminative Fine-Tuning (DFT), an improved variant of SFT, which mitigates the burden of collecting human-labeled preference data or training strong reward models. Unlike SFT that employs a generative approach and overlooks negative data, DFT adopts a discriminative paradigm that increases the probability of positive answers while suppressing potentially negative ones, aiming for data prediction instead of token prediction. Our contributions include: (i) a discriminative probabilistic framework for fine-tuning LLMs by explicitly modeling the discriminative likelihood of an answer among all possible outputs given an input; (ii) efficient algorithms to optimize this discriminative likelihood; and (iii) extensive experiments demonstrating DFT's effectiveness, achieving performance better than SFT and comparable to if not better than SFT$\rightarrow$PO. The code can be found at https://github.com/Optimization-AI/DFT.

Changlong Yu, Hongming Zhang, Yangqiu Song et al.

Computational and cognitive studies suggest that the abstraction of eventualities (activities, states, and events) is crucial for humans to understand daily eventualities. In this paper, we propose a scalable approach to model the entailment relations between eventualities ("eat an apple'' entails ''eat fruit''). As a result, we construct a large-scale eventuality entailment graph (EEG), which has 10 million eventuality nodes and 103 million entailment edges. Detailed experiments and analysis demonstrate the effectiveness of the proposed approach and quality of the resulting knowledge graph. Our datasets and code are available at https://github.com/HKUST-KnowComp/ASER-EEG.

Zhepei Wei, Wenlin Yao, Yao Liu et al.

While reinforcement learning (RL) has demonstrated remarkable success in enhancing large language models (LLMs), it has primarily focused on single-turn tasks such as solving math problems. Training effective web agents for multi-turn interactions remains challenging due to the complexity of long-horizon decision-making across dynamic web interfaces. In this work, we present WebAgent-R1, a simple yet effective end-to-end multi-turn RL framework for training web agents. It learns directly from online interactions with web environments by asynchronously generating diverse trajectories, entirely guided by binary rewards depending on task success. Experiments on the WebArena-Lite benchmark demonstrate the effectiveness of WebAgent-R1, boosting the task success rate of Qwen-2.5-3B from 6.1% to 33.9% and Llama-3.1-8B from 8.5% to 44.8%, significantly outperforming existing state-of-the-art methods and strong proprietary models such as OpenAI o3. In-depth analyses reveal the effectiveness of the thinking-based prompting strategy and test-time scaling through increased interactions for web tasks. We further investigate different RL initialization policies by introducing two variants, namely WebAgent-R1-Zero and WebAgent-R1-CoT, which highlight the importance of the warm-up training stage (i.e., behavior cloning) and provide insights on incorporating long chain-of-thought (CoT) reasoning in web agents.

Weiqi Wang, Limeng Cui, Xin Liu et al.

Goal-oriented script planning, or the ability to devise coherent sequences of actions toward specific goals, is commonly employed by humans to plan for typical activities. In e-commerce, customers increasingly seek LLM-based assistants to generate scripts and recommend products at each step, thereby facilitating convenient and efficient shopping experiences. However, this capability remains underexplored due to several challenges, including the inability of LLMs to simultaneously conduct script planning and product retrieval, difficulties in matching products caused by semantic discrepancies between planned actions and search queries, and a lack of methods and benchmark data for evaluation. In this paper, we step forward by formally defining the task of E-commerce Script Planning (EcomScript) as three sequential subtasks. We propose a novel framework that enables the scalable generation of product-enriched scripts by associating products with each step based on the semantic similarity between the actions and their purchase intentions. By applying our framework to real-world e-commerce data, we construct the very first large-scale EcomScript dataset, EcomScriptBench, which includes 605,229 scripts sourced from 2.4 million products. Human annotations are then conducted to provide gold labels for a sampled subset, forming an evaluation benchmark. Extensive experiments reveal that current (L)LMs face significant challenges with EcomScript tasks, even after fine-tuning, while injecting product purchase intentions improves their performance.

Jun Gao, Huan Zhao, Wei Wang et al.

In this study, we present EventRL, a reinforcement learning approach developed to enhance event extraction for large language models (LLMs). EventRL utilizes outcome supervision with specific reward functions to tackle prevalent challenges in LLMs, such as instruction following and hallucination, manifested as the mismatch of event structure and the generation of undefined event types. We evaluate EventRL against existing methods like Few-Shot Prompting (FSP) (based on GPT4) and Supervised Fine-Tuning (SFT) across various LLMs, including GPT-4, LLaMa, and CodeLLaMa models. Our findings show that EventRL significantly outperforms these conventional approaches by improving the performance in identifying and structuring events, particularly in handling novel event types. The study emphasizes the critical role of reward function selection and demonstrates the benefits of incorporating code data for better event extraction. While increasing model size leads to higher accuracy, maintaining the ability to generalize is essential to avoid overfitting.

Ilgee Hong, Changlong Yu, Liang Qiu et al. · gatech

Reinforcement learning from human feedback (RLHF) has become a powerful post-training paradigm for aligning large language models with human preferences. A core challenge in RLHF is constructing accurate reward signals, where the conventional Bradley-Terry reward models (BT RMs) often suffer from sensitivity to data size and coverage, as well as vulnerability to reward hacking. Generative reward models (GenRMs) offer a more robust alternative by generating chain-of-thought (CoT) rationales followed by a final reward. However, existing GenRMs rely on shallow, vertically scaled reasoning, limiting their capacity to handle nuanced or complex (e.g., reasoning-intensive) tasks. Moreover, their pairwise preference outputs are incompatible with standard RLHF algorithms that require pointwise reward signals. In this work, we introduce Think-RM, a training framework that enables long-horizon reasoning in GenRMs by modeling an internal thinking process. Rather than producing structured, externally provided rationales, Think-RM generates flexible, self-guided reasoning traces that support advanced capabilities such as self-reflection, hypothetical reasoning, and divergent reasoning. To elicit these reasoning abilities, we first warm-up the models by supervised fine-tuning (SFT) over long CoT data. We then further improve the model's long-horizon abilities by rule-based reinforcement learning (RL). In addition, we propose a novel pairwise RLHF pipeline that directly optimizes policies using pairwise preference rewards, eliminating the need for pointwise reward conversion and enabling more effective use of Think-RM outputs. Experiments show that Think-RM achieves state-of-the-art results on RM-Bench, outperforming both BT RM and vertically scaled GenRM by 8%. When combined with our pairwise RLHF pipeline, it demonstrates superior end-policy performance compared to traditional approaches.

Chen Xu, Tian Lan, Yu Ji et al.

Constrained decoding approaches aim to control the meaning or style of text generated by the pre-trained large language models (LLMs or also PLMs) for various tasks at inference time. However, these methods often guide plausible continuations by greedily and explicitly selecting targets. Though fulfilling the task requirements, these methods may overlook certain general and natural logics that humans would implicitly follow towards such targets. Inspired by cognitive dual-process theory, in this work, we propose a novel decoding framework DECIDER where the base LLMs are equipped with a First-Order Logic (FOL) reasoner to express and evaluate the rules, along with a decision function that merges the outputs of both systems to guide the generation. Unlike previous constrained decodings, DECIDER transforms the encouragement of target-specific words into all words that satisfy several high-level rules, enabling us to programmatically integrate our logic into LLMs. Experiments on CommonGen and PersonaChat demonstrate that DECIDER effectively follows given FOL rules to guide LLMs in a more human-like and logic-controlled manner.

Yining Lu, Zilong Wang, Shiyang Li et al.

Prior works in multi-objective reinforcement learning typically use linear reward scalarization with fixed weights, which provably fail to capture non-convex Pareto fronts and thus yield suboptimal results. This limitation becomes especially critical in online preference alignment for large language models. Here, stochastic trajectories generated by parameterized policies create highly non-linear and non-convex mappings from parameters to objectives that no single static weighting scheme can find optimal trade-offs. We address this limitation by introducing dynamic reward weighting, which adaptively adjusts reward weights during the online reinforcement learning process. Unlike existing approaches that rely on fixed-weight interpolation, our dynamic weighting continuously balances and prioritizes objectives in training, facilitating effective exploration of Pareto fronts in objective space. We introduce two approaches of increasing sophistication and generalizability: (1) hypervolume-guided weight adaptation and (2) gradient-based weight optimization, offering a versatile toolkit for online multi-objective alignment. Our extensive experiments demonstrate their compatibility with commonly used online reinforcement learning algorithms (including GRPO, REINFORCE, and RLOO), effectiveness across multiple mathematical reasoning datasets, and applicability to different model families, consistently achieving Pareto dominant solutions with fewer training steps than fixed-weight linear scalarization baselines.

Yuqi Yang, Weiqi Wang, Baixuan Xu et al.

Session history is a common way of recording user interacting behaviors throughout a browsing activity with multiple products. For example, if an user clicks a product webpage and then leaves, it might because there are certain features that don't satisfy the user, which serve as an important indicator of on-the-spot user preferences. However, all prior works fail to capture and model customer intention effectively because insufficient information exploitation and only apparent information like descriptions and titles are used. There is also a lack of data and corresponding benchmark for explicitly modeling intention in E-commerce product purchase sessions. To address these issues, we introduce the concept of an intention tree and propose a dataset curation pipeline. Together, we construct a sibling multimodal benchmark, SessionIntentBench, that evaluates L(V)LMs' capability on understanding inter-session intention shift with four subtasks. With 1,952,177 intention entries, 1,132,145 session intention trajectories, and 13,003,664 available tasks mined using 10,905 sessions, we provide a scalable way to exploit the existing session data for customer intention understanding. We conduct human annotations to collect ground-truth label for a subset of collected data to form an evaluation gold set. Extensive experiments on the annotated data further confirm that current L(V)LMs fail to capture and utilize the intention across the complex session setting. Further analysis show injecting intention enhances LLMs' performances.

Zhenghao Xu, Qin Lu, Qingru Zhang et al. · gatech

Reward model (RM) plays a pivotal role in reinforcement learning with human feedback (RLHF) for aligning large language models (LLMs). However, classical RMs trained on human preferences are vulnerable to reward hacking and generalize poorly to out-of-distribution (OOD) inputs. By contrast, strong LLM judges equipped with reasoning capabilities demonstrate superior generalization, even without additional training, but incur significantly higher inference costs, limiting their applicability in online RLHF. In this work, we propose an uncertainty-based routing framework that efficiently complements a fast RM with a strong but costly LLM judge. Our approach formulates advantage estimation in policy gradient (PG) methods as pairwise preference classification, enabling principled uncertainty quantification to guide routing. Uncertain pairs are forwarded to the LLM judge, while confident ones are evaluated by the RM. Experiments on RM benchmarks demonstrate that our uncertainty-based routing strategy significantly outperforms random judge calling at the same cost, and downstream alignment results showcase its effectiveness in improving online RLHF.

Yizhuo Chen, Xin Liu, Ruijie Wang et al.

Large language models (LLMs) achieve strong benchmark performance, yet user experiences remain inconsistent due to diverse preferences in style, tone, and reasoning mode. Nevertheless, existing alignment techniques such as reinforcement learning from human feedback (RLHF) or Direct Preference Optimization (DPO) largely optimize toward population-level averages and overlook individual variation. Naive personalization strategies like per-user fine-tuning are computationally prohibitive, and in-context approaches that prepend raw user signals often suffer from inefficiency and noise. To address these challenges, we propose POPI, a general framework that introduces a preference inference model to distill heterogeneous user signals into concise natural language summaries. These summaries act as transparent, compact, and transferable personalization representations that condition a shared generation model to produce personalized responses. POPI jointly optimizes both preference inference and personalized generation under a unified objective using reinforcement learning, ensuring summaries maximally encode useful preference information. Extensive experiments across four personalization benchmarks demonstrate that POPI consistently improves personalization accuracy while reducing context overhead by a large margin. Moreover, optimized summaries seamlessly transfer to frozen off-the-shelf LLMs, enabling plug-and-play personalization without weight updates.

Yuheng Zhang, Wenlin Yao, Changlong Yu et al.

Large language models (LLMs) have achieved impressive reasoning performance, with reinforcement learning with verifiable rewards (RLVR) emerging as a standard paradigm for post-training. A representative algorithm, group relative policy optimization (GRPO) (Shao et al., 2024), computes advantages by normalizing outcome rewards within response groups, but suffers from a vanishing advantage issue when all responses in a group receive identical rewards. To address this issue, we propose Adaptive Rollout and Response Reuse Policy Optimization (AR3PO), a sampling efficient RLVR algorithm that introduces two novel techniques: adaptive rollout, which dynamically allocates more responses to difficult prompts while saving computation on easier ones, and response reuse, which leverages previously generated correct responses to provide useful training signals. We compare AR3PO with strong RLVR baselines on multiple representative benchmarks using two different families of base models. Across the 7B and 8B models, AR3PO consistently outperforms GRPO and matches or surpasses DAPO (Yu et al., 2025), reducing rollout cost by up to 4.2x. On the larger 32B model, AR3PO achieves comparable performance to DAPO at similar training steps while maintaining substantially lower rollout cost.

Baixuan Xu, Weiqi Wang, Haochen Shi et al.

Improving user experience and providing personalized search results in E-commerce platforms heavily rely on understanding purchase intention. However, existing methods for acquiring large-scale intentions bank on distilling large language models with human annotation for verification. Such an approach tends to generate product-centric intentions, overlook valuable visual information from product images, and incurs high costs for scalability. To address these issues, we introduce MIND, a multimodal framework that allows Large Vision-Language Models (LVLMs) to infer purchase intentions from multimodal product metadata and prioritize human-centric ones. Using Amazon Review data, we apply MIND and create a multimodal intention knowledge base, which contains 1,264,441 million intentions derived from 126,142 co-buy shopping records across 107,215 products. Extensive human evaluations demonstrate the high plausibility and typicality of our obtained intentions and validate the effectiveness of our distillation framework and filtering mechanism. Additional experiments reveal that our obtained intentions significantly enhance large language models in two intention comprehension tasks.

Changlong Yu, Hongming Zhang, Yangqiu Song et al.

Large-scale pre-trained language models have demonstrated strong knowledge representation ability. However, recent studies suggest that even though these giant models contains rich simple commonsense knowledge (e.g., bird can fly and fish can swim.), they often struggle with the complex commonsense knowledge that involves multiple eventualities (verb-centric phrases, e.g., identifying the relationship between ``Jim yells at Bob'' and ``Bob is upset'').To address this problem, in this paper, we propose to help pre-trained language models better incorporate complex commonsense knowledge. Different from existing fine-tuning approaches, we do not focus on a specific task and propose a general language model named CoCoLM. Through the careful training over a large-scale eventuality knowledge graphs ASER, we successfully teach pre-trained language models (i.e., BERT and RoBERTa) rich complex commonsense knowledge among eventualities. Experiments on multiple downstream commonsense tasks that requires the correct understanding of eventualities demonstrate the effectiveness of CoCoLM.

Fuyu Lv, Mengxue Li, Tonglei Guo et al.

Deep learning-based sequential recommender systems have recently attracted increasing attention from both academia and industry. Most of industrial Embedding-Based Retrieval (EBR) system for recommendation share the similar ideas with sequential recommenders. Among them, how to comprehensively capture sequential user interest is a fundamental problem. However, most existing sequential recommendation models take as input clicked or purchased behavior sequences from user-item interactions. This leads to incomprehensive user representation and sub-optimal model performance, since they ignore the complete user behavior exposure data, i.e., items impressed yet unclicked by users. In this work, we attempt to incorporate and model those unclicked item sequences using a new learning approach in order to explore better sequential recommendation technique. An efficient triplet metric learning algorithm is proposed to appropriately learn the representation of unclicked items. Our method can be simply integrated with existing sequential recommendation models by a confidence fusion network and further gain better user representation. The offline experimental results based on real-world E-commerce data demonstrate the effectiveness and verify the importance of unclicked items in sequential recommendation. Moreover we deploy our new model (named XDM) into EBR of recommender system at Taobao, outperforming the deployed previous generation SDM.

Changlong Yu, Jialong Han, Peifeng Wang et al.

We address hypernymy detection, i.e., whether an is-a relationship exists between words (x, y), with the help of large textual corpora. Most conventional approaches to this task have been categorized to be either pattern-based or distributional. Recent studies suggest that pattern-based ones are superior, if large-scale Hearst pairs are extracted and fed, with the sparsity of unseen (x, y) pairs relieved. However, they become invalid in some specific sparsity cases, where x or y is not involved in any pattern. For the first time, this paper quantifies the non-negligible existence of those specific cases. We also demonstrate that distributional methods are ideal to make up for pattern-based ones in such cases. We devise a complementary framework, under which a pattern-based and a distributional model collaborate seamlessly in cases which they each prefer. On several benchmark datasets, our framework achieves competitive improvements and the case study shows its better interpretability.

Hongming Zhang, Jiaxin Bai, Yan Song et al.

Conventional word embeddings represent words with fixed vectors, which are usually trained based on co-occurrence patterns among words. In doing so, however, the power of such representations is limited, where the same word might be functionalized separately under different syntactic relations. To address this limitation, one solution is to incorporate relational dependencies of different words into their embeddings. Therefore, in this paper, we propose a multiplex word embedding model, which can be easily extended according to various relations among words. As a result, each word has a center embedding to represent its overall semantics, and several relational embeddings to represent its relational dependencies. Compared to existing models, our model can effectively distinguish words with respect to different relations without introducing unnecessary sparseness. Moreover, to accommodate various relations, we use a small dimension for relational embeddings and our model is able to keep their effectiveness. Experiments on selectional preference acquisition and word similarity demonstrate the effectiveness of the proposed model, and a further study of scalability also proves that our embeddings only need 1/20 of the original embedding size to achieve better performance.

Fuyu Lv, Taiwei Jin, Changlong Yu et al.

Capturing users' precise preferences is a fundamental problem in large-scale recommender system. Currently, item-based Collaborative Filtering (CF) methods are common matching approaches in industry. However, they are not effective to model dynamic and evolving preferences of users. In this paper, we propose a new sequential deep matching (SDM) model to capture users' dynamic preferences by combining short-term sessions and long-term behaviors. Compared with existing sequence-aware recommendation methods, we tackle the following two inherent problems in real-world applications: (1) there could exist multiple interest tendencies in one session. (2) long-term preferences may not be effectively fused with current session interests. Long-term behaviors are various and complex, hence those highly related to the short-term session should be kept for fusion. We propose to encode behavior sequences with two corresponding components: multi-head self-attention module to capture multiple types of interests and long-short term gated fusion module to incorporate long-term preferences. Successive items are recommended after matching between sequential user behavior vector and item embedding vectors. Offline experiments on real-world datasets show the superior performance of the proposed SDM. Moreover, SDM has been successfully deployed on online large-scale recommender system at Taobao and achieves improvements in terms of a range of commercial metrics.