Searcharxiv⌕ Search

arXiv subjects

Weizhi Ma

Publications and source records attributed to Weizhi Ma.

At least 19 recordsLinked to original sources
arXiv2026

Beyond Utility: Evaluating LLM as Recommender

With the rapid development of Large Language Models (LLMs), recent studies employed LLMs as recommenders to provide personalized information services for distinct users. Despite efforts to improve the accuracy of LLM-based recommendation models, relatively little attention is paid to beyond-utility dimensions. Moreover, there are unique evaluation aspects of LLM-based recommendation models, which have been largely ignored. To bridge this gap, we explore four new evaluation dimensions and propose a multidimensional evaluation framework. The new evaluation dimensions include: 1) history length sensitivity, 2) candidate position bias, 3) generation-involved performance, and 4) hallucinations. All four dimensions have the potential to impact performance, but are largely unnecessary for consideration in traditional systems. Using this multidimensional evaluation framework, along with traditional aspects, we evaluate the performance of seven LLM-based recommenders, with three prompting strategies, comparing them with six traditional models on both ranking and re-ranking tasks on four datasets. We find that LLMs excel at handling tasks with prior knowledge and shorter input histories in the ranking setting, and perform better in the re-ranking setting, beating traditional models across multiple dimensions. However, LLMs exhibit substantial candidate position bias issues, and some models hallucinate nonexistent items much more often than others. We intend our evaluation framework and observations to benefit future research on the use of LLMs as recommenders. The code and data are available at https://github.com/JiangDeccc/EvaLLMasRecommender.

cs.IR↗
arXiv2026

Integrating LLM and Diffusion-Based Agents for Social Simulation

Large language models (LLMs) offer strong semantic reasoning capabilities for user modeling, but applying LLM-based simulation to an entire social network is computationally expensive and often unreliable for users with sparse behavioral histories. Meanwhile, conventional information diffusion models efficiently exploit historical propagation patterns and social structures, but provide limited understanding of item content and user-item semantic compatibility. We propose HySID, a hybrid framework for individual-level information adoption prediction that combines semantic reasoning with structural diffusion. HySID first analyzes the historical user-relation graph to adaptively select a small set of structurally informative core users. It then applies LLM-based simulation to estimate the engagement of these users and converts the judgments into a diffusion-compatible seed. Finally, a plug-in diffusion backbone propagates this seed through historical interaction structures to rank potential adopters across the full user population. This design enables LLMs to focus on users for whom semantic reasoning is most informative while allowing neural diffusion models to generalize the evidence to users that are not explicitly simulated. Experiments on three real-world datasets from Weibo, Zhihu, and KuaiRand show that HySID consistently improves four diffusion backbones in Recall and NDCG, while outperforming full-population LLM simulation baselines. At the same time, selective simulation reduces LLM inference cost by approximately 83\% to 99.9\%, demonstrating that HySID provides an effective and computationally practical approach to scalable information adoption prediction.

cs.CY↗
arXiv2026

Enhancing LLM Metacognition via Cognitive Pairwise Training

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 +5.6 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.

cs.LG↗
arXiv2026

Don't Wait to Reply: Towards Responsive yet Thoughtful Dialogue through Proactive Thinking

Thinking has emerged as a critical capability for Large Language Models (LLMs) tackling complex tasks. However, its reactive nature, where reasoning is passively triggered only upon receiving a user response, inevitably introduces latency that compromises conversational fluidity. This stands in sharp contrast to human dialogue, where speakers proactively anticipate and plan future content during natural pauses to ensure seamless interaction. To bridge this gap, we propose Proactive Thinking, a framework that empowers models to pre-compute potential response elements during conversational downtime instead of waiting idly for the next input. We then introduce a training-free baseline that can think ahead by anticipating future states, balancing efficiency and quality through speculative continual thinking. To evaluate this approach in practice, we adapt three benchmarks of varying complexity into time-aware environments that simulate real-time conversational flow. We demonstrate that proactive thinking effectively improves interaction efficiency without compromising performance. Ultimately, this work advocates for a fundamental shift toward more intelligent, anticipatory, and real-time conversational AI.

cs.CL↗
arXiv2026

APAO: Bridging the Training-Inference Gap in Generative Recommendation via Adaptive Prefix-Aware Optimization

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. The source code is publicly available at https://github.com/yuyq18/APAO.

cs.IR↗
arXiv2026

UR$^2$: Unify RAG and Reasoning through Reinforcement Learning

Large Language Models (LLMs) have shown strong capabilities through two complementary paradigms: Retrieval-Augmented Generation (RAG) for knowledge grounding and Reinforcement Learning from Verifiable Rewards (RLVR) for complex reasoning. However, existing attempts to unify these paradigms remain narrow in scope, typically limited to open-domain QA with fixed retrieval settings, which constrains generalization to broader domains. To address this limitation, we propose UR$^2$ (Unified RAG and Reasoning)), a general reinforcement learning framework that dynamically coordinates retrieval and reasoning. UR$^2$ introduces two key designs: a difficulty-aware curriculum that selectively invokes retrieval only for challenging instances, and a hybrid knowledge access strategy that combines domain-specific offline corpora with on-the-fly LLM-generated summaries. Together, these components mitigate the imbalance between retrieval and reasoning and improve robustness to noisy information. Experiments on open-domain QA, MMLU-Pro, medical, and mathematical reasoning tasks show that UR$^2$, built on Qwen-2.5-3/7B and LLaMA-3.1-8B, consistently outperforms existing RAG and RL baselines, and achieves performance comparable to GPT-4o-mini and GPT-4.1-mini on several benchmarks. Our code is available at https://github.com/Tsinghua-dhy/UR2.

cs.CL↗
arXiv2026

Personalized Turn-Level User Conversation Satisfaction Benchmark

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.

cs.CL↗
arXiv2026

TheraAgent: Self-Improving Therapeutic Agent for Precise and Comprehensive Treatment Planning

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.

cs.AI↗
arXiv2026

Beyond "I Don't Know": Evaluating LLM Self-Awareness in Discriminating Data and Model Uncertainty

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.

cs.CL↗
arXiv2026

Let the Model Distribute Its Doubt: Confidence Estimation through Verbalized Probability Distribution

Knowing the reliability of a model's response is essential in practical applications. Given the strong generation capabilities of large language models (LLMs), research has focused on generating verbalized confidence. This approach is further enhanced by integrating chain-of-thought reasoning, which provides logical and transparent estimates. However, how reasoning strategies affect the estimated confidence remains under-explored. In this work, we demonstrate that predicting a verbalized probability distribution effectively promotes reasoning for confidence estimation. It requires an LLM to consider all possible answers rather than relying on a single guess, and the requirement of producing a distribution elicits more careful confidence assignment. We conduct systematic experiments comparing different verbalization-based methods across multiple LLMs and tasks. Our method consistently shows advantages, whether in the simple prompting setup or after optimization via reinforcement learning (RL). Notably, it achieves higher reasoning efficacy during inference-time scaling, saving nearly 6$\times$ the computation to reach the best Brier score of the strongest baseline on MMLU-Pro. Additionally, we reveal its limitations on specific tasks and discuss possible solutions for broader applicability.

cs.CL↗
arXiv2026

Doctor-R1: Mastering Clinical Inquiry with Experiential Agentic Reinforcement Learning

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 expert evaluations show that Doctor-R1 achieves superior clinical capability and patient-centric performance, demonstrating the effectiveness of the framework.

cs.AI↗
arXiv2026

Towards Transparent RAG: Fostering Evidence Traceability in LLM Generation via Reinforcement Learning

Retrieval-Augmented Generation (RAG) delivers substantial value in knowledge-intensive applications. However, its generated responses often lack transparent reasoning paths that trace back to source evidence from retrieved documents. This opacity not only compromises the interpretability of the output but also limits the model's ability to fully exploit the provided context. To address this, we propose TRACE (Transparent RAG with evidenCE tracing), a framework designed to enhance evidence traceability in Large Language Models (LLMs) through reinforcement learning (RL). TRACE guides LLMs to produce structured outputs with explicit evidence citations by prompting and rewarding evidence relevance and proper formatting, alongside accuracy, to optimize structured traceability. To ensure training stability with multiple reward signals, we further introduce an adaptive strategy for merging rewards and adopt a stabilized KL-divergence estimator. Experiments on three multi-hop QA datasets using Qwen2.5-7B-Instruct and Llama-3.1-8B-Instruct show that TRACE achieves both transparent, evidence-attributed outputs and accuracy improvements of 10-30%. The resulting performance is comparable to advanced commercial LLMs (e.g., OpenAI o1, DeepSeek-R1). Further analyses demonstrate strong generalization capabilities to unseen tasks. Our code is publicly available now.

cs.CL↗
arXiv2026

Understanding Internal Representations of Recommendation Models with Sparse Autoencoders

Recommendation model interpretation aims to reveal the relationships between inputs, model internal representations and outputs to enhance the transparency, interpretability, and trustworthiness of recommendation systems. However, the inherent complexity and opacity of deep learning models pose challenges for model-level interpretation. Moreover, most existing methods for interpreting recommendation models are tailored to specific architectures or model types, limiting their generalizability across different types of recommenders. In this paper, we propose RecSAE, a generalizable probing framework that interprets Recommendation models with Sparse AutoEncoders. The framework extracts interpretable latents from the internal representations of recommendation models, and links them to semantic concepts for interpretations. It does not alter original models during interpretations and also enables targeted tuning to models. Experiments on three types of recommendation models (general, graph-based, sequential) with four widely used public datasets demonstrate the effectiveness and generalization of RecSAE framework. The interpreted concepts are further validated by human experts, showing strong alignment with human perception. Overall, RecSAE serves as a novel step in both model-level interpretations to various types of recommendation models without affecting their functions and offering potential for targeted tuning of models.

cs.IR↗
arXiv2026

Patient-Zero: Scaling Synthetic Patient Agents to Real-World Distributions without Real Patient Data

Synthetic data generation with Large Language Models (LLMs) has emerged as a promising solution in the medical domain to mitigate data scarcity and privacy constraints. However, existing approaches remain constrained by their derivative nature, relying on real-world records, which pose privacy risks and distribution biases. Furthermore, current patient agents face the Stability-Plasticity Dilemma, struggling to maintain clinical consistency during dynamic inquiries. To address these challenges, we introduce Patient-Zero, a novel framework for ab initio patient simulation that requires no real medical records. Our Medically-Aligned Hierarchical Synthesis framework generates comprehensive and diverse patient records from abstract clinical guidelines via stratified attribute permutation. To support rigorous clinical interaction, we design a Dual-Track Cognitive Memory System to enable agents dynamically update memory while preserving logical consistency and persona adherence. Extensive evaluations show that Patient-Zero establishes a new state-of-the-art in both data quality and interaction fidelity. In human expert evaluations, senior licensed physicians judge our synthetic data to be statistically indistinguishable from real human-authored data and higher in clinical quality. Furthermore, downstream medical reasoning model trained on our synthetic dataset shows substantial performance gains (MedQA +24.0%; MMLU +14.5%), demonstrating the practical utility of our framework.

cs.AI↗
arXiv2025

GSE: Evaluating Sticker Visual Semantic Similarity via a General Sticker Encoder

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.

cs.CV↗
arXiv2025

Human vs. Agent in Task-Oriented Conversations

Task-oriented conversational systems are essential for efficiently addressing diverse user needs, yet their development requires substantial amounts of high-quality conversational data that is challenging and costly to obtain. While large language models (LLMs) have demonstrated potential in generating synthetic conversations, the extent to which these agent-generated interactions can effectively substitute real human conversations remains unclear. This work presents the first systematic comparison between LLM-simulated users and human users in personalized task-oriented conversations. We propose a comprehensive analytical framework encompassing three key aspects (conversation strategy, interaction style, and conversation evaluation) and ten distinct dimensions for evaluating user behaviors, and collect parallel conversational datasets from both human users and LLM agent users across four representative scenarios under identical conditions. Our analysis reveals significant behavioral differences between the two user types in problem-solving approaches, question broadness, user engagement, context dependency, feedback polarity and promise, language style, and hallucination awareness. We found consistency in the agent users and human users across the depth-first or breadth-first dimensions, as well as the usefulness dimensions. These findings provide critical insights for advancing LLM-based user simulation. Our multi-dimensional taxonomy constructed a generalizable framework for analyzing user behavior patterns, offering insights from LLM agent users and human users. By this work, we provide perspectives on rethinking how to use user simulation in conversational systems in the future.

cs.IR↗
arXiv2025

The Dialogue That Heals: A Comprehensive Evaluation of Doctor Agents' Inquiry Capability

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.

cs.CL↗
arXiv2025

Small Stickers, Big Meanings: A Multilingual Sticker Semantic Understanding Dataset with a Gamified Approach

Stickers, though small, are a highly condensed form of visual expression, ubiquitous across messaging platforms and embraced by diverse cultures, genders, and age groups. Despite their popularity, sticker retrieval remains an underexplored task due to the significant human effort and subjectivity involved in constructing high-quality sticker query datasets. Although large language models (LLMs) excel at general NLP tasks, they falter when confronted with the nuanced, intangible, and highly specific nature of sticker query generation. To address this challenge, we propose a threefold solution. First, we introduce Sticktionary, a gamified annotation framework designed to gather diverse, high-quality, and contextually resonant sticker queries. Second, we present StickerQueries, a multilingual sticker query dataset containing 1,115 English and 615 Chinese queries, annotated by over 60 contributors across 60+ hours. Lastly, Through extensive quantitative and qualitative evaluation, we demonstrate that our approach significantly enhances query generation quality, retrieval accuracy, and semantic understanding in the sticker domain. To support future research, we publicly release our multilingual dataset along with two fine-tuned query generation models.

cs.MM↗