论文检索

Tazeek Bin Abdur Rakib, Ambuj Mehrish, Lay-Ki Soon, Wern Han Lim, Soujanya Poria

Large-language-model (LLM) agents excel at reactive dialogue but struggle with proactive, goal-driven interactions due to myopic decoding and costly planning. We introduce DialogXpert, which leverages a frozen LLM to propose a small, high-quality set of candidate actions per turn and employs a compact Q-network over fixed BERT embeddings trained via temporal-difference learning to select optimal moves within this reduced space. By tracking the user's emotions DialogXpert tailors each decision to advance the task while nurturing a genuine, empathetic connection. Across negotiation, emotional support, and tutoring benchmarks, DialogXpert drives conversations to under 3 turns with success rates exceeding 94% and, with a larger LLM prior, pushes success above 97% while markedly improving negotiation outcomes. This framework delivers real-time, strategic, and emotionally intelligent dialogue planning at scale.

Lixiu Wu, Yuanrong Tang, Qisen Pan, Xianyang Zhan, Yuchen Han, Lanxi Xiao, Tianhong Wang, Chen Zhong, Jiangtao Gong

Due to privacy concerns, open dialogue datasets for mental health are primarily generated through human or AI synthesis methods. However, the inherent implicit nature of psychological processes, particularly those of clients, poses challenges to the authenticity and diversity of synthetic data. In this paper, we propose ECAs (short for Embodied Conversational Agents), a framework for embodied agent simulation based on Large Language Models (LLMs) that incorporates multiple psychological theoretical principles. Using simulation, we expand real counseling case data into a nuanced embodied cognitive memory space and generate dialogue data based on high-frequency counseling questions. We validated our framework using the D4 dataset. First, we created a public ECAs dataset through batch simulations based on D4. Licensed counselors evaluated our method, demonstrating that it significantly outperforms baselines in simulation authenticity and necessity. Additionally, two LLM-based automated evaluation methods were employed to confirm the higher quality of the generated dialogues compared to the baselines.

Heyang Ma, Qirui Mi, Qipeng Yang, Zijun Fan, Bo Li, Haifeng Zhang

Economic decision‑making depends not only on structured signals—such as prices and taxes—but also on unstructured language, including peer dialogue and media narratives. While multi‑agent reinforcement learning (MARL) has shown promise in optimizing economic decisions, it struggles with the semantic ambiguity and contextual richness of language. We propose LAMP (Language‑Augmented Multi‑Agent Policy), the first framework to integrate language into economic decision‑making, narrowing the gap to real‑world settings. LAMP follows a Think–Speak–Decide pipeline: (1) Think interprets numerical observations to extract short‑term shocks and long‑term trends, caching high‑value reasoning trajectories. (2) Speak crafts and exchanges strategic messages based on the reasoning, updating beliefs by parsing peer communications. (3) Decide fuses numerical data, reasoning, and reflections into a MARL policy to optimize language‑augmented decision‑making. Experiments in economic simulation show that LAMP outperforms both MARL and LLM‑only baselines in cumulative return (+63.5%, +34.0%), robustness (+18.8%, +59.4%), and interpretability. These results demonstrate the potential of language‑augmented policies to deliver more effective and robust economic strategies.

Cheng Tang, Guochong Sui, Wenqi Lou, Zihan Wang, Jiayi Tuo, Wenqian Xie, Yinkang Gao, Yixuan Zhu, Lei Gong, Chao Wang 等

Hardware accelerators such as GPUs, NPUs, and FPGAs are essential to meeting AI’s computational demands. With the proliferation of heterogeneous devices across cloud and edge, various model optimization techniques adapt to diverse hardware characteristics through operator transformations and structural modifications. Accurate, efficient latency prediction enables rapid selection of optimal strategies across hardware backends. Many existing methods treat hardware as a black-box executor, directly regressing latency without explicitly modeling the intricate interactions between neural network (NN) structures and device-specific execution behaviors. To address these challenges, we introduce a new modeling perspective that captures the interaction between neural architectures and hardware execution. To capture device-specific characteristics, we propose two complementary modeling strategies. The Device Behavior Signature Selector (DBSel) characterizes hardware execution behavior by selectively probing a small set of representative architectures, forming a compact, workload-driven profile. In parallel, we construct capability vectors that capture the hierarchical memory of each device and compute characteristics, providing a structured abstraction of its architectural capacity. To unify both behavioral and structural views, we introduce the Hardware–Operation Dialogue Module (HODM), which models fine-grained interactions between neural operators and hardware properties. Together, these components empower CloserToMe to deliver accurate and transferable latency predictions across unseen and diverse platforms.

Chengtao Lv, Bilang Zhang, Yang Yong, Ruihao Gong, Yushi Huang, Shiqiao Gu, Jiajun Wu, Yumeng Shi, Jinyang Guo, Wenya Wang

Large Vision-Language Models (VLMs) exhibit impressive multi-modal capabilities but suffer from prohibitive computational and memory demands, due to their long visual token sequences and massive parameter sizes. To address these issues, recent works have proposed training-free compression methods. However, existing efforts often suffer from three major limitations: (1) Current approaches do not decompose techniques into comparable modules, hindering fair evaluation across spatial and temporal redundancy. (2) Evaluation confined to simple single-turn tasks, failing to reflect performance in realistic scenarios. (3) Isolated use of individual compression techniques, without exploring their joint potential. To overcome these gaps, we introduce LLMC+, a comprehensive VLM compression benchmark with a versatile, plug-and-play toolkit. LLMC+ supports over 20 algorithms across five representative VLM families and enables systematic study of token-level and model-level compression. Our benchmark reveals that: (1) Spatial and temporal redundancies demand distinct technical strategies. (2) Token reduction methods degrade significantly in multi-turn dialogue and detail-sensitive tasks. (3) Combining token and model compression achieves extreme compression with minimal performance loss. We believe LLMC+ will facilitate fair evaluation and inspire future research in efficient VLM.

Xun Liang, Xin Guo, Zhongming Jin, Weihang Pan, Penghui Shang, Deng Cai, Binbin Lin, Jieping Ye

The spatial reasoning task aims to reason about the spatial relationships in 2D and 3D space, which is a fundamental capability for Visual Question Answering (VQA) and robotics. Although vision language models (VLMs) have developed rapidly in recent years, they are still struggling with the spatial reasoning task. In this paper, we introduce a method that can enhance Spatial reasoning through Visual and Textual thinking Simultaneously (SpatialVTS). In the spatial visual thinking phase, our model is trained to generate location-related specific tokens of important targets automatically. Not only are the objects mentioned in the problem addressed, but also the potential objects related to the reasoning are considered. During the spatial textual thinking phase, our model conducts long-term thinking based on visual cues and dialogues and gradually inferences the answers to spatial reasoning problems. To effectively support the model's training, we made manual corrections to the existing spatial reasoning dataset, eliminating numerous incorrect labels resulting from automatic annotation, restructuring the data input format to enhance generalization, and developing a reasoning framework for model thinking. Without introducing any additional information (such as masks or depth), our model's overall average level in several spatial understanding tasks has significantly improved compared with other models.

Heyan Huang, Yizhe Yang, Huashan Sun, Jiawei Li, Yang Gao

Large language models have enabled sophisticated dialogue planning policy, but their reliance on LLM-generated simulation and feedback for policy optimization may introduce systematic preference bias. We present the first comprehensive analysis of preference bias in LLM-based dialogue planners, evaluating four state-of-the-art planning policies across three dialogue domains using multiple LLM families at varying scales. Our investigation reveals that all tested planners exhibit significant preference bias, systematically favoring narrow strategy sets rather than maintaining balanced distributions. User simulation emerges as the primary bias driver, while diverse persona simulation fails as an effective mitigation strategy. Most concerning, preference bias drives planners toward ethically problematic strategies that achieve short-term success while undermining real-world effectiveness and ethical standards. Our findings establish fundamental challenges for responsible deployment of LLM-based dialogue systems and provide crucial insights for developing more reliable and ethically-aligned planning approaches.

Renke Dai, Hebin Hu, Jiahui Zhang, Yilin Kang, Ah-Hwee Tan

Though promising in healthcare consultation applications, large language models (LLMs) face critical limitations in retaining and utilizing long-term memory across multi-turn interactions. In particular, existing memory enhancing paradigms are constrained by limited context windows and embedding-based retrieval, often failing to maintain task relevance and still suffering from memory prototype collapse in multi-turn healthcare consultation. To address these challenges, we propose a cognitively-inspired memory framework named MemoryART, which is grounded in Adaptive Resonance Theory (ART)—a cognitive and learning theory of how humans and animals adapt to dynamic environments. MemoryART employs three memory modules—working memory, episodic memory, and semantic memory to support task-aware memory organization and dynamic retrieval. Specifically, episodic memory provides the storage of specific experiences along with contextual clues, which is crucial for managing patient-specific information and perfect for multi-turn healthcare consultation interactions. Building upon this concept, MemoryART leverages multi-channel competitive learning and resonance matching to enable efficient and interpretable episodic memory encoding, alleviating issues of prototype collapse and noisy memory associations. For evaluation, we construct a long-term medical dialogue benchmark called MediLongChat using a LLM-based generation pipeline. The resulting dataset features realistic, multi-disease chat histories, each exceeding 100K tokens across 20–30 dialogues, simulating real-world healthcare interaction patterns. Our experimental results show that MemoryART outperforms mainstream approaches in memory-intensive tasks, achieving SOTA results and significantly reducing token consumption across five popular LLMs, confirming its effectiveness and efficiency in providing scalable, reliable memory for LLMs in healthcare.

Bo Zhang, Hao Yu, Wenjie Dong, Yvhang Yang, Dezhuang Miao, Fengyi Song, Yanhui Gu, Xiaoming Zhang, Junsheng Zhou

Conventional feedback, even when accompanied by brief explanations, rarely uncovers the hidden contradictions that trigger a learner's mistake. We bridge this gap with counterfactual question generation (CFQG): given a learner's answer, generate a follow-up question that deliberately contradicts it, compelling the learner to confront the underlying conflict. CFQG thus transforms assessment from passive scoring into an interactive and contradiction-centered dialogue that supports knowledge repair. To automate CFQG, we propose GapProbe, which probes the knowledge gap between a learner’s belief and curated facts through a knowledge graph (KG), then designs counterfactual questions (CFQs) that negate the belief. Identifying contradiction-aware triples, and more importantly, selecting those most likely to confuse the learner, are highly challenging in large-scale KGs. GapProbe tackles these challenges with an iterative ProConB cycle coupled with a schema-aware KGMap. By caching one- and multi-hop schema patterns of the KG, KGMap provides ``roadmap'' to guide LLMs jump to deep and contradiction-aware triples, beyond traditional step-wise graph traversal. We present the CFQG benchmark and corresponding metrics for evaluating how generated CFQs trigger, focus, and deepen learner reflection through explicit contradictions. Experiments on multiple datasets and LLMs show that GapProbe boosts LLM reasoning over KGs and generates follow-up questions that consistently promote deeper and more focused learner reflection.

Yeming Li, Chenxi Liu, Jie Zou, Cheng Long, Chaoning Zhang, Peng Wang, Yang Yang

Conversational Recommender Systems (CRS) aim to provide personalized recommendations by interacting with users through natural language dialogue. However, in scenarios requiring deep geospatial awareness, existing methods, including those based on Large Language Models (LLMs), still face significant challenges in effectively fusing heterogeneous, multimodal geographic information with dynamic dialogue context. Simple fusion strategies struggle to resolve the asymmetric dependencies between dynamic user intent and static geographic context and fail to bridge the semantic gap between LLMs and structured geospatial data. To address these issues, we propose a framework for geography-aware CRS, named GeoCRS. Our core idea is to empower a frozen LLM with powerful geospatial reasoning capabilities by conditioning it on a dynamic, multimodal guidance signal generated by an external fusion architecture, all without altering the LLM's internal parameters. Specifically, we first design a hierarchical geographical encoder to uniformly represent heterogeneous geographic data. Subsequently, we introduce a contextual feature modulation module that asymmetrically injects the geographic context into the user's dialogue intent via a novel modulation mechanism to improve conversational recommendation via both geographic and dialogue context. Extensive experiments on public benchmark datasets demonstrate that our proposed GeoCRS significantly outperforms state-of-the-art baselines on the geography-aware conversational recommendation task.

Weijie Li, Zhongqing Wang, Guodong Zhou

Most commonsense reasoning models overlook the influence of personality traits, limiting their effectiveness in personalized systems such as dialogue generation. To address this limitation, we introduce the Personality-aware Commonsense Knowledge Graph (PCoKG), a structured dataset comprising 521,316 quadruples. We begin by employing three evaluators to score and filter events from the ATOMIC dataset, selecting those that are likely to elicit diverse reasoning patterns across different personality types. For knowledge graph construction, we leverage the role-playing capabilities of large language models (LLMs) to perform reasoning tasks. To enhance the quality of the generated knowledge, we incorporate a debate mechanism consisting of a proponent, an opponent, and a judge, which iteratively refines the outputs through feedback loops. We evaluate the dataset from multiple perspectives and conduct fine-tuning and ablation experiments using multiple LLM backbones to assess PCoKG's robustness and the effectiveness of its construction pipeline. Our LoRA-based fine-tuning results indicate a positive correlation between model performance and the parameter scale of the base models. Finally, we apply PCoKG to persona-based dialogue generation, where it demonstrates improved consistency between generated responses and reference outputs. This work bridges the gap between commonsense reasoning and individual cognitive differences, enabling the development of more personalized and context-aware AI systems.

Jinyu Chen, Hongyu Li, Zongheng Tang, Xiaoduo Li, Wenjun Wu, Si Liu

Visual Dialogue Navigation (VDN) aims to enable agents to reach target locations through dialogue with humans. The integration of VDN into Unmanned Aerial Vehicle (UAV) systems enhances human-machine interaction by enabling intuitive, hands-free operation, thereby unlocking vast applications. However, existing VDN models for UAVs can only perform navigation based on dialogue history, lacking proactive interaction capabilities to correct trajectories. Moreover, their sequential observation history recording mechanism struggles to accurately localize landmarks observed in the historical context, leading to ineffective utilization of referential information in new user instructions.To address these, we present AerialVLA, an end-to-end UAV navigation framework integrating dialogue comprehension, action decision-making, and navigational question generation. AerialVLA comprises three core components: i) we propose the Progress-Driven Navigation-Query Alternation mechanism to determine optimal questioning timing through navigation progress estimation autonomously. ii) To effectively model long-horizon history observation sequences, we develop the History Spatial-Temporal Fusion module that extracts discriminative spatial-temporal representations from historical observations. iii) Furthermore, to overcome data scarcity in training, we devise the Online Task-Driven Augmentation strategy that enhances learning through action-conditioned data augmentation. Experimental results demonstrate that AerialVLA achieves state-of-the-art navigation performance while exhibiting effective dialogue capabilities.Moreover, to better evaluate the agent's proactive dialogue and navigation abilities, our evaluation benchmark, named UAV Navigation with Online Dialogue (UNOD), incorporates an online dialogue interaction module. The UNOD assesses UAV agents' real-time questioning capabilities by leveraging an Air Commander Large Language Model to simulate human-UAV interactions during testing.

Sicheng Yang, Yukai Huang, Weitong Cai, Shitong Sun, You He, Jiankang Deng, Hang Zhang, Jifei Song, Zhensong Zhang

The performance of egocentric AI agents is fundamentally limited by multimodal intent ambiguity. This challenge arises from a combination of underspecified language, imperfect visual data, and deictic gestures, which frequently leads to task failure. Existing monolithic Vision-Language Models (VLMs) struggle to resolve these multimodal ambiguous inputs, often failing silently or hallucinating responses. To address these ambiguities, we introduce the Plug-and-Play Clarifier, a zero-shot and modular framework that decomposes the problem into discrete, solvable sub-tasks. Specifically, our framework consists of three synergistic modules: (1) a text clarifier that uses dialogue-driven reasoning to interactively disambiguate linguistic intent, (2) a vision clarifier that delivers real-time guidance feedback, instructing users to adjust their positioning for improved capture quality, and (3) a cross-modal clarifier with grounding mechanism that robustly interprets 3D pointing gestures and identifies the specific objects users are pointing to. Extensive experiments demonstrate that our framework improves the intent clarification performance of small language models (4-8B) by approximately 30%, making them competitive with significantly larger counterparts. We also observe consistent gains when applying our framework to these larger models. Furthermore, our vision clarifier increases corrective guidance accuracy by over 20%, and our cross-modal clarifier improves semantic answer accuracy for referential grounding by 5%. Overall, our method provides a plug-and-play framework that effectively resolves multimodal ambiguity and significantly enhances user experience in egocentric interaction.

Yuanchen Shi, Longyin Zhang, Maodong Li, Yibin Zheng, Xiuhong Wang, Fang Kong

The growing demand for psychological support underscores the lack of high-quality counseling dialogue datasets, particularly in non-English contexts. We propose PGSim, a Path-Guided Simulation framework that mirrors real counseling processes—symptom description, problem identification, cause analysis, strategy planning, and iterative adjustment. PGSim models each user scenario as a fine-grained quadruple {Group, Psychological Problem, Problem Cause, Support Focus} and guides dialogue generation through expert-annotated strategy paths. Real counseling dialogues and expert-edited samples are used to fine-tune two language models: a Dialog Generator for strategy-aligned dialogue creation and a Dialog Modifier for expert-level refinement. After automated and human verification, we construct the Chinese Psychological support Dialogue Dataset (CPsDD), containing 68K dialogues across 13 groups, 16 problems, 13 causes, and 12 support focuses. We further present the Comprehensive Agent Dialogue Support System (CADSS), which integrates profiling, summarization, strategy planning, and empathetic response. Experiments on CPsDD and ESConv demonstrate that CADSS achieves state-of-the-art results on Strategy Prediction and Emotional Support Conversation tasks.

Jing Li, Yanxin Luo, Donghong Han, Yimeng Zhan, Xiaoming Fu, Baiyou Qiao, Gang Wu

Emotional Support Conversation (ESC) aims to alleviate individuals’ negative emotions through multi-turn dialogues, where effective strategy planning and response generation are essential. However, existing methods often suffer from limitations in both planning reasonable support strategies and effectively expressing them in responses. To the end, we propose a novel LLM-based Emotional Support Conversation Agent (ESCA) with a plug-in strategy planner and a strategy-aligned prompt generator. The strategy planner cooperates with four aspects of the seeker’s state, including emotion intensity, trust degree, dialogue behavior, and stage of change, to enhance the rationality and effectiveness of the strategy prediction. To ensure that predicted strategies are better conveyed, the prompt generator integrates strategy-aligned instructions, knowledge, and context to generate the soft prompt for guiding the LLM to generate supportive responses. In addition to supervised fine-tuning, the prompt generator is further optimized by reinforcement learning. Experimental results demonstrate that ESCA significantly improves both response quality and the success rate of achieving the ESC task goal.

Jian Zhang, Zhangqi Wang, Haiping Zhu, Kangda Cheng, Kai He, Bo Li, Qika Lin, Jun Liu, Erik Cambria

Large language models (LLMs) typically operate in a question-answering paradigm, where the quality of the input prompt critically affects the response. Automated Prompt Optimization (APO) aims to overcome the cognitive biases of manually crafted prompts and explore a broader prompt design space. However, existing APO methods often suffer from rigid template structures and inefficient exploration in the prompt space. To this end, we propose a Multi-Agent Adaptive Reasoning with Socratic guidance framework (MARS) for APO. MARS consists of five complementary agents and formulates the optimization process as a Partially Observable Markov Decision Process (POMDP), enabling adaptive prompt refinement through explicit state modeling and interactive feedback. Specifically, a Planner agent generates flexible optimization trajectories, a Teacher-Critic-Student triad engages in Socratic-style dialogue to iteratively optimize the prompt based on pseudo-gradient signals in the text space, and a Target agent executes the prompt in downstream tasks to provide performance feedback. MARS integrates reasoning, feedback, and state transition into a unified hidden-state evolution process, improving both the effectiveness and interpretability of optimization. Extensive experiments on multiple datasets demonstrate that MARS outperforms existing APO methods in terms of optimization performance, search efficiency, and interpretability.

Guanghui Ye, Huan Zhao, Yingxue Gao, Zhixue Zhao, Kehan Wang, Xupeng Zha, Zhihua Jiang

Large language model (LLM)-based visual dialogue (VD) systems have made response generation for image-grounded conversations more correct and coherent. However, user engagement - the extent to which a user is interested, emotionally involved, and willing to continue the conversation - remains a challenge. To fully explore engaging VD, we propose: (i) a new task named Audio-enhanced VD (AVD), which introduces additional audio dialogue contexts that can more vividly convey the speaker's emotions as input, with the aim of generating correct but more engaging dialogue responses. Specifically, we employ a text-to-speech model as the modality translator to generate the paired acoustic utterances from the inputting textual utterances; (ii) an accompanying approach named Visually-grounded and Interleaved Text-Audio Dialogue Modeling (VITA-DM), which utilizes both image-grounded information and interleaved text-audio utterances for visual dialogue modeling, differentiating from previous multi-modal LLM (MLLM)-based methods that normally model text and audio modalities separately. We also present three pre-training tasks to better learn multi-modal interactions across language, vision, and audio; (iii) a novel metric named Multi-Modal Engagement (MME), which fills the gap of engagement estimation in VD and can provide a fine-grained assessment along emotional, attentional, and reply engagement dimensions (EE, AE, RE). We experiment on two popular datasets and provide extensive evaluations (automatic, engagement-specific, and human), supporting the validity of our approach. Furthermore, based on empirical results that reveal that emotions contribute the most to engagement, we justify our emphasis on the emotional aspect throughout the definition, solution, and evaluation of our task.

Yongwen Ren, Chao Wang, Peng Du, Chuan Qin, Dazhong Shen, Hui Xiong

Recent advances in pretrained language models (PLMs) have significantly improved conversational recommender systems (CRS), enabling more fluent and context-aware interactions. To further enhance accuracy and mitigate hallucination, many methods integrate PLMs with knowledge graphs (KGs), but face key challenges: failing to fully exploit PLM reasoning over graph relationships, indiscriminately incorporating retrieved knowledge without context filtering, and neglecting collaborative preferences in multi-turn dialogues. To this end, we propose PCRS-TKA, a prompt-based framework employing retrieval-augmented generation to integrate PLMs with KGs. PCRS-TKA constructs dialogue-specific knowledge trees from KGs and serializes them into texts, enabling structure-aware reasoning while capturing rich entity semantics. Our approach selectively filters context-relevant knowledge and explicitly models collaborative preferences using specialized supervision signals. A semantic alignment module harmonizes heterogeneous inputs, reducing noise and enhancing accuracy. Extensive experiments demonstrate that PCRS-TKA consistently outperforms all baselines in both recommendation and conversational quality.

Yuxuan Hu, Jian Chen, Yuhao Wang, Zixuan Li, Jing Xiong, Pengyue Jia, Wei Wang, Chengming Li, Xiangyu Zhao

Stickers are widely used in online communication to convey emotions and implicit intentions. The Sticker Response Selection (SRS) task aims to select the most contextually appropriate sticker based on the dialogue. However, existing methods typically rely on semantic matching and model emotional and intentional cues separately, which can lead to mismatches when emotions and intentions are misaligned. To address this issue, we propose Emotion and Intention Guided Multi-Modal Learning (EIGML). This framework is the first to jointly model emotion and intention, effectively reducing the bias caused by isolated modeling and significantly improving selection accuracy. Specifically, we introduce Dual-Level Contrastive Framework to perform both intra-modality and inter-modality alignment, ensuring consistent representation of emotional and intentional features within and across modalities. In addition, we design an Intention-Emotion Guided Multi-Modal Fusion module that integrates emotional and intentional information progressively through three components: Emotion-Guided Intention Knowledge Selection, Intention-Emotion Guided Attention Fusion, and Similarity-Adjusted Matching Mechanism. This design injects rich, effective information into the model and enables a deeper understanding of the dialogue, ultimately enhancing sticker selection performance. Experimental results on two public datasets show that EIGML outperforms state-of-the-art baselines, achieving higher accuracy and a better understanding of emotional and intentional features.

Hanbo Bi, Zhiqiang Yuan, Zexi Jia, Jiapei Zhang, Chongyang Li, Peixiang Luo, Ying Deng, Xiaoyue Duan, Jinchao Zhang

Traditional dialogue retrieval aims to select the most appropriate utterance or image from recent dialogue history. However, they often fail to meet users’ actual needs for revisiting semantically coherent content scattered across long-form conversations. To fill this gap, we define the Fine-grained Fragment Retrieval (FFR) task, requiring models to locate query-relevant fragments, comprising both utterances and images, from multimodal long-form dialogues. As a foundation for FFR, we construct MLDR, the longest-turn multimodal dialogue retrieval dataset to date, averaging 25.45 turns per dialogue, with each naturally spanning three distinct topics. To evaluate generalization in real-world scenarios, we curate and annotate a WeChat-based test set comprising real-world multimodal dialogues with an average of 75.38 turns. Building on these resources, we explore existing generation-based Vision-Language Models (VLMs) on FFR and observe that they often retrieve incoherent utterance-image fragments. While optimized for generating responses from visual-textual inputs, these models lack explicit supervision to ensure semantic coherence within retrieved fragments. To address this, we propose F2RVLM, a generative retrieval model trained in a two-stage paradigm: (1) supervised fine-tuning to inject fragment-level retrieval knowledge, and (2) GRPO-based reinforcement learning with multi-objective rewards to encourage outputs with semantic precision, relevance, and contextual coherence. In addition, to account for difficulty variations arising from differences in intra-fragment element distribution, ranging from locally dense to sparsely scattered, we introduce a difficulty-aware curriculum sampling that ranks training instances by predicted difficulty and gradually incorporates harder examples. This strategy enhances the model’s reasoning ability in long-form, multi-turn dialogue contexts. Experiments on both in-domain and real-domain sets demonstrate that F2RVLM substantially outperforms popular VLMs, achieving superior retrieval performance.