Weizhi Ma

Yuanqing Yu, Yifan Wang, Weizhi Ma et al.

Generative recommendation has recently emerged as a promising paradigm for sequential recommendation. It formulates the task as an autoregressive generation process, predicting tokens of the next item conditioned on user interaction histories. Existing generative recommendation models are typically trained with token-level likelihood objectives such as cross-entropy loss, while employing beam search during inference to generate ranked candidates. However, this leads to a fundamental training-inference inconsistency: standard training assumes ground-truth tokens are always available, while beam search prunes low-probability branches during inference, causing the correct item to be prematurely discarded when its prefixes receive low scores. To address this issue, we propose the Adaptive Prefix-Aware Optimization (APAO) framework, which introduces prefix-level optimization losses to better align the training objective with the inference setting. Furthermore, we design an adaptive worst-prefix optimization strategy that dynamically focuses on the most vulnerable prefixes during training, thereby enhancing the model's ability to retain correct candidates under beam search constraints. We provide theoretical analyses to demonstrate the effectiveness and efficiency of our framework. Extensive experiments show that APAO consistently alleviates the training-inference inconsistency and improves performance across generative recommendation backbones. Our codes are publicly available at https://github.com/yuyq18/APAO.

Qingyao Ai, Ting Bai, Zhao Cao et al. · pku, tsinghua

The research field of Information Retrieval (IR) has evolved significantly, expanding beyond traditional search to meet diverse user information needs. Recently, Large Language Models (LLMs) have demonstrated exceptional capabilities in text understanding, generation, and knowledge inference, opening up exciting avenues for IR research. LLMs not only facilitate generative retrieval but also offer improved solutions for user understanding, model evaluation, and user-system interactions. More importantly, the synergistic relationship among IR models, LLMs, and humans forms a new technical paradigm that is more powerful for information seeking. IR models provide real-time and relevant information, LLMs contribute internal knowledge, and humans play a central role of demanders and evaluators to the reliability of information services. Nevertheless, significant challenges exist, including computational costs, credibility concerns, domain-specific limitations, and ethical considerations. To thoroughly discuss the transformative impact of LLMs on IR research, the Chinese IR community conducted a strategic workshop in April 2023, yielding valuable insights. This paper provides a summary of the workshop's outcomes, including the rethinking of IR's core values, the mutual enhancement of LLMs and IR, the proposal of a novel IR technical paradigm, and open challenges.

Weitao Li, Hao Zhou, Xuanyu Lei et al.

Reinforcement learning with verifiable rewards (RLVR) has become central to LLM reasoning, but its outcome-level rewards can make models more willing to give confident answers when evidence or reasoning is unreliable. Existing SFT or RL methods mainly teach LLMs to refuse or express uncertainty at the response level, which can overfit abstention behavior rather than improve reasoning reliability. To address this limitation, we propose Cognitive Pairwise Training (CPT), a cognitive mid-training alignment stage that turns pairwise comparisons over reasoning traces into a reusable alignment signal. By learning to distinguish trustworthy from flawed reasoning, CPT encourages the model to internalize a reasoning-quality discrimination boundary rather than memorize surface refusal patterns. Across five model scales and three model families, CPT improves the reasoning--metacognition trade-off. At 14B, CPT+RL outperforms the standard SFT+RL pipeline by +2.2 math-average points and +5.2 abstention-F1 points. Further analyses show that CPT improves trace quality and exhibits strong robustness and scalability across evaluation and training settings. Code and models are released at https://github.com/Tsinghua-dhy/CPT.

Yangkun Li, Weizhi Ma, Chong Chen et al.

Overfitting is a common problem in machine learning, which means the model too closely fits the training data while performing poorly in the test data. Among various methods of coping with overfitting, dropout is one of the representative ways. From randomly dropping neurons to dropping neural structures, dropout has achieved great success in improving model performances. Although various dropout methods have been designed and widely applied in past years, their effectiveness, application scenarios, and contributions have not been comprehensively summarized and empirically compared by far. It is the right time to make a comprehensive survey. In this paper, we systematically review previous dropout methods and classify them into three major categories according to the stage where dropout operation is performed. Specifically, more than seventy dropout methods published in top AI conferences or journals (e.g., TKDE, KDD, TheWebConf, SIGIR) are involved. The designed taxonomy is easy to understand and capable of including new dropout methods. Then, we further discuss their application scenarios, connections, and contributions. To verify the effectiveness of distinct dropout methods, extensive experiments are conducted on recommendation scenarios with abundant heterogeneous information. Finally, we propose some open problems and potential research directions about dropout that worth to be further explored.

Yuancheng Sun, Yimeng Chen, Weizhi Ma et al.

Molecular property prediction is essential for drug discovery. In recent years, deep learning methods have been introduced to this area and achieved state-of-the-art performances. However, most of existing methods ignore the intrinsic relations between molecular properties which can be utilized to improve the performances of corresponding prediction tasks. In this paper, we propose a new approach, namely Physics properties Enhanced Molecular Property prediction (PEMP), to utilize relations between molecular properties revealed by previous physics theory and physical chemistry studies. Specifically, we enhance the training of the chemical and physiological property predictors with related physics property prediction tasks. We design two different methods for PEMP, respectively based on multi-task learning and transfer learning. Both methods include a model-agnostic molecule representation module and a property prediction module. In our implementation, we adopt both the state-of-the-art molecule embedding models under the supervised learning paradigm and the pretraining paradigm as the molecule representation module of PEMP, respectively. Experimental results on public benchmark MoleculeNet show that the proposed methods have the ability to outperform corresponding state-of-the-art models.

Zhefan Wang, Zhiqiang Guo, Weizhi Ma et al.

User satisfaction with AI assistants is highly personalized: the same response may satisfy one user but disappoint another depending on what each user expects and what they have asked for before. Existing automatic evaluation methods mostly measure generic response quality, making it difficult to judge whether a response satisfies a user at a specific turn. We study this problem as personalized turn-level user conversation satisfaction evaluation. We build a conversation satisfaction evaluator that combines compact user memories with target-turn context to produce satisfaction scores and dissatisfaction-oriented rationales. Meta-evaluation against human satisfaction annotations shows that personalized memory and post-hoc score calibration improve ordinal agreement and dissatisfied-turn detection over supervised, retrieval-based, and generic LLM-as-a-judge baselines. We further introduce PersTurnBench, a personalized turn-level user conversation satisfaction benchmark that uses the verified evaluator to assess generation models via replay. By holding the replay state fixed, PersTurnBench enables controlled comparison of generic generation models and memory-augmented personalized systems without new human labels for every candidate model. The evaluator and benchmark let researchers compare candidate generation models on personalized satisfaction without collecting new user feedback for every model.

Heng Er Metilda Chee, Jiayin Wang, Zhiqiang Guo et al.

Stickers have become a popular form of visual communication, yet understanding their semantic relationships remains challenging due to their highly diverse and symbolic content. In this work, we formally {define the Sticker Semantic Similarity task} and introduce {Triple-S}, the first benchmark for this task, consisting of 905 human-annotated positive and negative sticker pairs. Through extensive evaluation, we show that existing pretrained vision and multimodal models struggle to capture nuanced sticker semantics. To address this, we propose the {General Sticker Encoder (GSE)}, a lightweight and versatile model that learns robust sticker embeddings using both Triple-S and additional datasets. GSE achieves superior performance on unseen stickers, and demonstrates strong results on downstream tasks such as emotion classification and sticker-to-sticker retrieval. By releasing both Triple-S and GSE, we provide standardized evaluation tools and robust embeddings, enabling future research in sticker understanding, retrieval, and multimodal content generation. The Triple-S benchmark and GSE have been publicly released and are available here.

Weitao Li, Junkai Li, Weizhi Ma et al.

Large language models (LLMs) exhibit powerful general intelligence across diverse scenarios, including their integration into chatbots. However, a vital challenge of LLM-based chatbots is that they may produce hallucinated content in responses, which significantly limits their applicability. Various efforts have been made to alleviate hallucination, such as retrieval augmented generation and reinforcement learning with human feedback, but most of them require additional training and data annotation. In this paper, we propose a novel post-hoc Citation-Enhanced Generation (CEG) approach combined with retrieval argumentation. Unlike previous studies that focus on preventing hallucinations during generation, our method addresses this issue in a post-hoc way. It incorporates a retrieval module to search for supporting documents relevant to the generated content, and employs a natural language inference-based citation generation module. Once the statements in the generated content lack of reference, our model can regenerate responses until all statements are supported by citations. Note that our method is a training-free plug-and-play plugin that is capable of various LLMs. Experiments on various hallucination-related datasets show our framework outperforms state-of-the-art methods in both hallucination detection and response regeneration on three benchmarks. Our codes and dataset will be publicly available.

Jiayin Wang, Fengran Mo, Weizhi Ma et al.

Large language models (LLMs) are essential tools that users employ across various scenarios, so evaluating their performance and guiding users in selecting the suitable service is important. Although many benchmarks exist, they mainly focus on specific predefined model abilities, such as world knowledge, reasoning, etc. Based on these ability scores, it is hard for users to determine which LLM best suits their particular needs. To address these issues, we propose to evaluate LLMs from a user-centric perspective and design this benchmark to measure their efficacy in satisfying user needs under distinct intents. Firstly, we collect 1,846 real-world use cases from a user study with 712 participants from 23 countries. This first-hand data helps us understand actual user intents and needs in LLM interactions, forming the User Reported Scenarios (URS) dataset, which is categorized with six types of user intents. Secondly, based on this authentic dataset, we benchmark 10 LLM services with GPT-4-as-Judge. Thirdly, we show that benchmark scores align well with human preference in both real-world experience and pair-wise annotations, achieving Pearson correlations of 0.95 and 0.94, respectively. This alignment confirms that the URS dataset and our evaluation method establish an effective user-centric benchmark. The dataset, code, and process data are available at https://github.com/Alice1998/URS.

Jingyi Ren, Ante Wang, Yunghwei Lai et al.

Reliable Large Language Models (LLMs) should abstain when confidence is insufficient. However, prior studies often treat refusal as a generic "I don't know'', failing to distinguish input-level ambiguity (data uncertainty) from capability limitations (model uncertainty). This lack of distinction limits downstream action decisions like requesting clarification or invoking external tools. In this work, we introduce UA-Bench, a benchmark of over 3,500 questions drawn from six datasets spanning knowledge-intensive and reasoning-intensive tasks, designed to evaluate explicit uncertainty attribution. An evaluation of 18 frontier LLMs shows that even state-of-the-art models struggle to reliably discriminate between data uncertainty and model uncertainty, and that high answer accuracy does not necessarily imply strong uncertainty attribution ability. To narrow this gap, we propose a lightweight data synthesis and reinforcement learning strategy. Experiments on both Qwen3-4B-Instruct-2507 and Qwen3-8B in thinking mode show that the proposed method improves uncertainty attribution while preserving answer accuracy. Our code and data are publicly available now.

Weitao Li, Boran Xiang, Xiaolong Wang et al.

Large Language Models (LLMs) have shown remarkable capabilities through two complementary paradigms: Retrieval-Augmented Generation (RAG), which enhances knowledge grounding, and Reinforcement Learning from Verifiable Rewards (RLVR), which optimizes complex reasoning abilities. However, these two capabilities are often developed in isolation, and existing efforts to unify them remain narrow in scope -- typically limited to open-domain QA with fixed retrieval settings and task-specific constraints. This lack of integration constrains generalization and limits the applicability of RAG-RL methods to broader domains. To bridge this gap, we propose UR2 (Unified RAG and Reasoning), a general framework that unifies retrieval and reasoning through reinforcement learning. UR2 introduces two key contributions: a difficulty-aware curriculum training that selectively invokes retrieval only for challenging problems, and a hybrid knowledge access strategy combining domain-specific offline corpora with LLM-generated summaries. These components are designed to enable dynamic coordination between retrieval and reasoning, improving adaptability across a diverse range of tasks. Experiments across open-domain QA, MMLU-Pro, medical, and mathematical reasoning tasks demonstrate that UR$^2$ (built on Qwen-2.5-3/7B and LLaMA-3.1-8B) significantly outperforms existing RAG and RL methods, achieving comparable performance to GPT-4o-mini and GPT-4.1-mini on several benchmarks. We have released all code, models, and data at https://github.com/Tsinghua-dhy/UR2.

Junkai Li, Yunghwei Lai, Weitao Li et al.

The recent rapid development of large language models (LLMs) has sparked a new wave of technological revolution in medical artificial intelligence (AI). While LLMs are designed to understand and generate text like a human, autonomous agents that utilize LLMs as their "brain" have exhibited capabilities beyond text processing such as planning, reflection, and using tools by enabling their "bodies" to interact with the environment. We introduce a simulacrum of hospital called Agent Hospital that simulates the entire process of treating illness, in which all patients, nurses, and doctors are LLM-powered autonomous agents. Within the simulacrum, doctor agents are able to evolve by treating a large number of patient agents without the need to label training data manually. After treating tens of thousands of patient agents in the simulacrum (human doctors may take several years in the real world), the evolved doctor agents outperform state-of-the-art medical agent methods on the MedQA benchmark comprising US Medical Licensing Examination (USMLE) test questions. Our methods of simulacrum construction and agent evolution have the potential in benefiting a broad range of applications beyond medical AI.

Yunghwei Lai, Kaiming Liu, Ziyue Wang et al.

The professionalism of a human doctor in outpatient service depends on two core abilities: the ability to make accurate medical decisions and the medical consultation skill to conduct strategic, empathetic patient inquiry. Existing Large Language Models (LLMs) have achieved remarkable accuracy on medical decision-making benchmarks. However, they often lack the ability to conduct the strategic and empathetic consultation, which is essential for real-world clinical scenarios. To address this gap, we propose Doctor-R1, an AI doctor agent trained to master both of the capabilities by ask high-yield questions and conduct strategic multi-turn inquiry to guide decision-making. Our framework introduces three key components: a multi-agent interactive environment, a two-tiered reward architecture that separately optimizes clinical decision-making and communicative inquiry skills, and an experience repository to ground policy learning in high-quality prior trajectories. We evaluate Doctor-R1 on OpenAI's HealthBench and MAQuE, assessed across multi-facet metrics, such as communication quality, user experience, and task accuracy. Remarkably, Doctor-R1 surpasses state-of-the-art open-source specialized LLMs by a substantial margin with higher parameter efficiency and outperforms powerful proprietary models. Furthermore, the human evaluations show a strong preference for Doctor-R1 to generate human-preferred clinical dialogue, demonstrating the effectiveness of the framework.

Xiaoyu Zhang, Yishan Li, Jiayin Wang et al.

The explainability of recommender systems has attracted significant attention in academia and industry. Many efforts have been made for explainable recommendations, yet evaluating the quality of the explanations remains a challenging and unresolved issue. In recent years, leveraging LLMs as evaluators presents a promising avenue in Natural Language Processing tasks (e.g., sentiment classification, information extraction), as they perform strong capabilities in instruction following and common-sense reasoning. However, evaluating recommendation explanatory texts is different from these NLG tasks, as its criteria are related to human perceptions and are usually subjective. In this paper, we investigate whether LLMs can serve as evaluators of recommendation explanations. To answer the question, we utilize real user feedback on explanations given from previous work and additionally collect third-party annotations and LLM evaluations. We design and apply a 3-level meta evaluation strategy to measure the correlation between evaluator labels and the ground truth provided by users. Our experiments reveal that LLMs, such as GPT4, can provide comparable evaluations with appropriate prompts and settings. We also provide further insights into combining human labels with the LLM evaluation process and utilizing ensembles of multiple heterogeneous LLM evaluators to enhance the accuracy and stability of evaluations. Our study verifies that utilizing LLMs as evaluators can be an accurate, reproducible and cost-effective solution for evaluating recommendation explanation texts. Our code is available at https://github.com/Xiaoyu-SZ/LLMasEvaluator.

Weitao Li, Kaiming Liu, Xiangyu Zhang et al.

Retrieval-Augmented Generation (RAG) has emerged as a widely adopted approach for knowledge injection during large language model (LLM) inference in recent years. However, due to their limited ability to exploit fine-grained inter-document relationships, current RAG implementations face challenges in effectively addressing the retrieved noise and redundancy content, which may cause error in the generation results. To address these limitations, we propose an Efficient Dynamic Clustering-based document Compression framework (EDC2-RAG) that utilizes latent inter-document relationships while simultaneously removing irrelevant information and redundant content. We validate our approach, built upon GPT-3.5-Turbo and GPT-4o-mini, on widely used knowledge-QA and Hallucination-Detection datasets. Experimental results show that our method achieves consistent performance improvements across various scenarios and experimental settings, demonstrating strong robustness and applicability. Our code and datasets are available at https://github.com/Tsinghua-dhy/EDC-2-RAG.

Weizhi Ma, Min Zhang, Yue Cao et al.

Explainability and effectiveness are two key aspects for building recommender systems. Prior efforts mostly focus on incorporating side information to achieve better recommendation performance. However, these methods have some weaknesses: (1) prediction of neural network-based embedding methods are hard to explain and debug; (2) symbolic, graph-based approaches (e.g., meta path-based models) require manual efforts and domain knowledge to define patterns and rules, and ignore the item association types (e.g. substitutable and complementary). In this paper, we propose a novel joint learning framework to integrate \textit{induction of explainable rules from knowledge graph} with \textit{construction of a rule-guided neural recommendation model}. The framework encourages two modules to complement each other in generating effective and explainable recommendation: 1) inductive rules, mined from item-centric knowledge graphs, summarize common multi-hop relational patterns for inferring different item associations and provide human-readable explanation for model prediction; 2) recommendation module can be augmented by induced rules and thus have better generalization ability dealing with the cold-start issue. Extensive experiments\footnote{Code and data can be found at: \url{https://github.com/THUIR/RuleRec}} show that our proposed method has achieved significant improvements in item recommendation over baselines on real-world datasets. Our model demonstrates robust performance over "noisy" item knowledge graphs, generated by linking item names to related entities.

Junkai Li, Yunghwei Lai, Tianyi Zhu et al.

Formulating a treatment plan is inherently a complex reasoning and refinement task rather than a simple generation problem. However, existing large language models (LLMs) mainly rely on one-shot output without explicit verification, which may result in rough, incomplete, and potentially unsafe treatment plans. To address these limitations, we propose TheraAgent, an agentic framework that replaces one-shot generation with an iterative generate-judge-refine pipeline. By mirroring the actual reasoning process of human experts who iteratively revise treatment plans, our framework progressively transforms coarse and incomplete drafts into precise, comprehensive, and safer therapeutic regimens. To facilitate the critical judge component, we introduce TheraJudge, a treatment-specific evaluation module integrated into the inference loop to enforce clinical standards. Experiments show TheraAgent achieves state-of-the-art results on HealthBench, leading in Accuracy and Completeness. In expert evaluations, it attains an 86% win rate against physicians, with superior Targeting and Harm Control. Moreover, the highly agreement between TheraJudge and HealthBench evaluations confirms the reliability of our framework.

Xun Jiang, Feng Li, Han Zhao et al.

Large language models (LLMs) like GPTs, trained on vast datasets, have demonstrated impressive capabilities in language understanding, reasoning, and planning, achieving human-level performance in various tasks. Most studies focus on enhancing these models by training on ever-larger datasets to build more powerful foundation models. While training stronger models is important, enabling models to evolve during inference is equally crucial, a process we refer to as AI self-evolution. Unlike large-scale training, self-evolution may rely on limited data or interactions. Inspired by the columnar organization of the human cerebral cortex, we hypothesize that AI models could develop cognitive abilities and build internal representations through iterative interactions with their environment. To achieve this, models need long-term memory (LTM) to store and manage processed interaction data. LTM supports self-evolution by representing diverse experiences across environments and agents. In this report, we explore AI self-evolution and its potential to enhance models during inference. We examine LTM's role in lifelong learning, allowing models to evolve based on accumulated interactions. We outline the structure of LTM and the systems needed for effective data retention and representation. We also classify approaches for building personalized models with LTM data and show how these models achieve self-evolution through interaction. Using LTM, our multi-agent framework OMNE achieved first place on the GAIA benchmark, demonstrating LTM's potential for AI self-evolution. Finally, we present a roadmap for future research, emphasizing the importance of LTM for advancing AI technology and its practical applications.

Yifan Wang, Peijie Sun, Weizhi Ma et al.

Fairness of recommender systems (RS) has attracted increasing attention recently. Based on the involved stakeholders, the fairness of RS can be divided into user fairness, item fairness, and two-sided fairness which considers both user and item fairness simultaneously. However, we argue that the intersectional two-sided unfairness may still exist even if the RS is two-sided fair, which is observed and shown by empirical studies on real-world data in this paper, and has not been well-studied previously. To mitigate this problem, we propose a novel approach called Intersectional Two-sided Fairness Recommendation (ITFR). Our method utilizes a sharpness-aware loss to perceive disadvantaged groups, and then uses collaborative loss balance to develop consistent distinguishing abilities for different intersectional groups. Additionally, predicted score normalization is leveraged to align positive predicted scores to fairly treat positives in different intersectional groups. Extensive experiments and analyses on three public datasets show that our proposed approach effectively alleviates the intersectional two-sided unfairness and consistently outperforms previous state-of-the-art methods.

Yuanqing Yu, Chongming Gao, Jiawei Chen et al.

Reinforcement Learning (RL)-Based Recommender Systems (RSs) have gained rising attention for their potential to enhance long-term user engagement. However, research in this field faces challenges, including the lack of user-friendly frameworks, inconsistent evaluation metrics, and difficulties in reproducing existing studies. To tackle these issues, we introduce EasyRL4Rec, an easy-to-use code library designed specifically for RL-based RSs. This library provides lightweight and diverse RL environments based on five public datasets and includes core modules with rich options, simplifying model development. It provides unified evaluation standards focusing on long-term outcomes and offers tailored designs for state modeling and action representation for recommendation scenarios. Furthermore, we share our findings from insightful experiments with current methods. EasyRL4Rec seeks to facilitate the model development and experimental process in the domain of RL-based RSs. The library is available for public use.

Jingyi Ren, Yekun Xu, Xiaolong Wang et al.

Retrieval-Augmented Generation (RAG) delivers substantial value in knowledge-intensive applications. Many recent works use reinforcement learning (RL) to elicit strong reasoning in RAG generators. However, two key challenges remain unresolved: (1) Transparency: most prior methods do not explicitly indicate which references are actually used during the reasoning that leads to the final answer, limiting interpretability and visibility; (2) Stability: the KL divergence estimator used in existing RL-based approaches may cause gradient spikes, leading to unstable training. To address these challenges, we propose Adaptive-Rewarded Evidence Navigation Agent (ARENA), a transparent and robust RAG generator framework trained via RL with designed rewards. Based on our structured protocol, KL divergence stabilization, and adaptive reward calculation modules, ARENA enables the RAG generator to identify key evidence, perform structured reasoning, and generate answers with interpretable decision traces. Applied to Qwen2.5-7B-Instruct and Llama3.1-8B-Instruct, extensive experiments across multiple baselines show 10-30% accuracy improvements on three multi-hop QA datasets, comparable to advanced closed-source LLMs (e.g., OpenAI o1, DeepSeek R1). Further analyses show that ARENA generalizes well to unseen datasets and tasks. Our models and codes are publicly released.

Linlu Gong, Ante Wang, Yunghwei Lai et al.

An effective physician should possess a combination of empathy, expertise, patience, and clear communication when treating a patient. Recent advances have successfully endowed AI doctors with expert diagnostic skills, particularly the ability to actively seek information through inquiry. However, other essential qualities of a good doctor remain overlooked. To bridge this gap, we present MAQuE(Medical Agent Questioning Evaluation), the largest-ever benchmark for the automatic and comprehensive evaluation of medical multi-turn questioning. It features 3,000 realistically simulated patient agents that exhibit diverse linguistic patterns, cognitive limitations, emotional responses, and tendencies for passive disclosure. We also introduce a multi-faceted evaluation framework, covering task success, inquiry proficiency, dialogue competence, inquiry efficiency, and patient experience. Experiments on different LLMs reveal substantial challenges across the evaluation aspects. Even state-of-the-art models show significant room for improvement in their inquiry capabilities. These models are highly sensitive to variations in realistic patient behavior, which considerably impacts diagnostic accuracy. Furthermore, our fine-grained metrics expose trade-offs between different evaluation perspectives, highlighting the challenge of balancing performance and practicality in real-world clinical settings.

Yunghwei Lai, Weizhi Ma, Yang Liu

Synthetic data generation using large language models (LLMs) has emerged as a promising solution across various domains, particularly in medical field, to mitigate data collection challenges. However, existing studies mainly utilize LLMs to rewrite and complete existing medical records, where the limitations in data privacy, accuracy, and diversity sill exist, and additionally lack the ability to interact like real patients. To address these issues, we propose a realistic patient generation framework, Patient-Zero, which requires no real medical records. Patient-Zero first introduces a medically-aligned multi-step generation architecture, which builds comprehensive patient records through hierarchical medical knowledge injection without real medical records. Then, to optimize the virtual patient's interaction abilities with humans, Patient-Zero designs a dynamic updating mechanism to improve the consistency and conversational performance. Our framework enables the generation of contextually diverse patient records while maintaining strict medical coherence, supported by adaptive dialogue strategies and real-time clinical plausibility verification. Experimental results demonstrate that our model achieves good performance in accuracy, diversity, and consistency. After training with our generated virtual patients, existing models show significant improvements on the MedQA dataset.

Luanbo Wan, Weizhi Ma

Long-term memory (LTM) is essential for large language models (LLMs) to achieve autonomous intelligence in complex, evolving environments. Despite increasing efforts in memory-augmented and retrieval-based architectures, there remains a lack of standardized benchmarks to systematically evaluate LLMs' long-term memory abilities. Existing benchmarks still face challenges in evaluating knowledge retention and dynamic sequential reasoning, and in their own flexibility, all of which limit their effectiveness in assessing models' LTM capabilities. To address these gaps, we propose a novel benchmark framework based on interactive fiction games, featuring dynamically branching storylines with complex reasoning structures. These structures simulate real-world scenarios by requiring LLMs to navigate hierarchical decision trees, where each choice triggers cascading dependencies across multi-turn interactions. Our benchmark emphasizes two distinct settings to test reasoning complexity: one with immediate feedback upon incorrect decisions, and the other requiring models to independently trace back and revise earlier choices after failure. As part of this benchmark, we also construct a new dataset designed to test LLMs' LTM within narrative-driven environments. We further validate the effectiveness of our approach through detailed experiments. Experimental results demonstrate the benchmark's ability to robustly and reliably assess LTM in LLMs.

Jiayin Wang, Zhiquang Guo, Weizhi Ma et al.

As evaluation designs of large language models may shape our trajectory toward artificial general intelligence, comprehensive and forward-looking assessment is essential. Existing benchmarks primarily assess static knowledge, while intelligence also entails the ability to rapidly learn from experience. To this end, we advocate for the evaluation of Test-time Learning, the capacity to improve performance in experience-based, reasoning-intensive tasks during test time. In this work, we propose semantic games as effective testbeds for evaluating test-time learning, due to their resistance to saturation and inherent demand for strategic reasoning. We introduce an objective evaluation framework that compares model performance under both limited and cumulative experience settings, and contains four forms of experience representation. To provide a comparative baseline, we recruit eight human participants to complete the same task. Results show that LLMs exhibit measurable test-time learning capabilities; however, their improvements are less stable under cumulative experience and progress more slowly than those observed in humans. These findings underscore the potential of LLMs as general-purpose learning machines, while also revealing a substantial intellectual gap between models and humans, irrespective of how well LLMs perform on static benchmarks.

Ante Wang, Weizhi Ma, Yang Liu

Knowing the reliability of a model's response is essential in application. With the strong generation capabilities of LLMs, research has focused on generating verbalized confidence. This is further enhanced by combining chain-of-thought reasoning, which provides logical and transparent estimation. However, how reasoning strategies affect the estimated confidence is still under-explored. In this work, we demonstrate that predicting a verbalized probability distribution can effectively encourage in-depth reasoning for confidence estimation. Intuitively, it requires an LLM to consider all candidates within the answer space instead of basing on a single guess, and to carefully assign confidence scores to meet the requirements of a distribution. This method shows an advantage across different models and various tasks, regardless of whether the answer space is known. Its advantage is maintained even after reinforcement learning, and further analysis shows its reasoning patterns are aligned with human expectations.

Jiayu Li, Ziyi Ye, Guohao Jian et al.

Can a recommendation model be self-aware? This paper investigates the recommender's self-awareness by quantifying its uncertainty, which provides a label-free estimation of its performance. Such self-assessment can enable more informed understanding and decision-making before the recommender engages with any users. To this end, we propose an intuitive and effective method, probability-based List Distribution uncertainty (LiDu). LiDu measures uncertainty by determining the probability that a recommender will generate a certain ranking list based on the prediction distributions of individual items. We validate LiDu's ability to represent model self-awareness in two settings: (1) with a matrix factorization model on a synthetic dataset, and (2) with popular recommendation algorithms on real-world datasets. Experimental results show that LiDu is more correlated with recommendation performance than a series of label-free performance estimators. Additionally, LiDu provides valuable insights into the dynamic inner states of models throughout training and inference. This work establishes an empirical connection between recommendation uncertainty and performance, framing it as a step towards more transparent and self-evaluating recommender systems.

Jimmy Lin, Junkai Li, Jiasi Gao et al.

Tactile signals collected by wearable electronics are essential in modeling and understanding human behavior. One of the main applications of tactile signals is action classification, especially in healthcare and robotics. However, existing tactile classification methods fail to capture the spatial and temporal features of tactile signals simultaneously, which results in sub-optimal performances. In this paper, we design Spatio-Temporal Aware tactility Transformer (STAT) to utilize continuous tactile signals for action classification. We propose spatial and temporal embeddings along with a new temporal pretraining task in our model, which aims to enhance the transformer in modeling the spatio-temporal features of tactile signals. Specially, the designed temporal pretraining task is to differentiate the time order of tubelet inputs to model the temporal properties explicitly. Experimental results on a public action classification dataset demonstrate that our model outperforms state-of-the-art methods in all metrics.

Bin Hao, Min Zhang, Weizhi Ma et al.

Data plays a vital role in machine learning studies. In the research of recommendation, both user behaviors and side information are helpful to model users. So, large-scale real scenario datasets with abundant user behaviors will contribute a lot. However, it is not easy to get such datasets as most of them are only hold and protected by companies. In this paper, a new large-scale dataset collected from a knowledge-sharing platform is presented, which is composed of around 100M interactions collected within 10 days, 798K users, 165K questions, 554K answers, 240K authors, 70K topics, and more than 501K user query keywords. There are also descriptions of users, answers, questions, authors, and topics, which are anonymous. Note that each user's latest query keywords have not been included in previous open datasets, which reveal users' explicit information needs. We characterize the dataset and demonstrate its potential applications for recommendation study. Multiple experiments show the dataset can be used to evaluate algorithms in general top-N recommendation, sequential recommendation, and context-aware recommendation. This dataset can also be used to integrate search and recommendation and recommendation with negative feedback. Besides, tasks beyond recommendation, such as user gender prediction, most valuable answerer identification, and high-quality answer recognition, can also use this dataset. To the best of our knowledge, this is the largest real-world interaction dataset for personalized recommendation.

Shaoyun Shi, Hanxiong Chen, Weizhi Ma et al.

Recent years have witnessed the success of deep neural networks in many research areas. The fundamental idea behind the design of most neural networks is to learn similarity patterns from data for prediction and inference, which lacks the ability of cognitive reasoning. However, the concrete ability of reasoning is critical to many theoretical and practical problems. On the other hand, traditional symbolic reasoning methods do well in making logical inference, but they are mostly hard rule-based reasoning, which limits their generalization ability to different tasks since difference tasks may require different rules. Both reasoning and generalization ability are important for prediction tasks such as recommender systems, where reasoning provides strong connection between user history and target items for accurate prediction, and generalization helps the model to draw a robust user portrait over noisy inputs. In this paper, we propose Logic-Integrated Neural Network (LINN) to integrate the power of deep learning and logic reasoning. LINN is a dynamic neural architecture that builds the computational graph according to input logical expressions. It learns basic logical operations such as AND, OR, NOT as neural modules, and conducts propositional logical reasoning through the network for inference. Experiments on theoretical task show that LINN achieves significant performance on solving logical equations and variables. Furthermore, we test our approach on the practical task of recommendation by formulating the task into a logical inference problem. Experiments show that LINN significantly outperforms state-of-the-art recommendation models in Top-K recommendation, which verifies the potential of LINN in practice.