Multimodal modeling represents a vital step from modality-agnostic reasoning toward world modeling. While early approaches predominantly rely on late-fusion that assembles encoders and frozen language backbones with output heads, recent efforts have shifted the paradigm toward native multimodal modeling (NMM) with the intrinsic integration of modalities for superior multimodal performance. Despite its potential, the design space of native architectures remains insufficiently defined. In this paper, we present the community with a formalized roadmap for this transition. Specifically, we formally define the architectural nativity, distinguishing mid-fusion and early-fusion from non-native paradigms. We further organize the existing native models through the lens of input-output duality into three categories: (i) Multi-to-Text for cross-modal comprehension with text-only output; (ii) Multi-to-Target for scenario-oriented generation e.g., image, audio and video generation, and (iii) Multi-to-Multi for unified modeling with symmetric input-output. We systematically unpack the end-to-end pipeline from industrial perspectives from: architectural coordination, massive data curation, to full-stack training recipes, inference deployment, and evaluation. Finally, based on our assessment of native multimodal development trends, we provide a forward-looking perspective on its future trajectory.

Trajectory data is getting more vital for training large language models for boosting the agentic abilities. Unlike the verifiable domains such as coding or mathematics, scaling trajectory data for open-ended tasks are much more difficult because these tasks lack singular ground truth and are costly to annotate or verify. In this paper, we propose RetroGen, a self-improving framework of retrospective process supervision. Our key observation is that although expert trajectories are scarce, high-quality final artifacts such as literature reviews, analyst reports and legal judgments, are abundant in pre-training data and can be viewed as compressed traces of the evidence-seeking processes that produced them. RetroGen reconstructs candidate latent trajectories from expert artifacts, verifies them against both the artifact and supporting evidence, and trains models on their own successful reconstruction data, without requiring trajectory data from stronger models. Experiments show that RetroGen improves grounding, faithful synthesis, and long-form evidence-seeking agent tasks.

Evaluating the exhaustive search capabilities of large language models (LLMs) is plagued by a fundamental paradox: verifying completeness requires complete ground truth, yet high-entropy enumeration tasks make such ground truth impossible for humans to create. This causes benchmarks to systematically penalize models for outperforming their human annotators. Despite rapid progress in web-search and deep research agents—which now issue hundreds of queries, traverse diverse sites, and synthesize long reports—evaluation still largely relies on partially annotated answer sets, LLM-based judges, or single-answer questions that avoid genuinely exhaustive search scenarios.We break this paradox by shifting the evaluation paradigm from simulating a messy reality to constructing computationally pure challenges. We introduce VERITAS (Verifiable Traversal Assessment for Search), a framework built on the principle of computationally irreducible constraints. By introducing novel, non-optimizable constraints, we create verifiable, sparse-answer search tasks that are computationally equivalent to exhaustive enumeration. These constraints are easy to verify but impossible for LLMs or search engines to optimize, forcing agents to genuinely traverse the entire search space. VERITAS can automatically generate a virtually infinite number of test cases with perfect ground truth and precise difficulty control, with marginal instance cost dominated by hash computations. This provides not only a robust benchmark for evaluating systematic exploration under uncertainty but also a scalable method for generating training data to improve these crucial, yet underdeveloped, capabilities.

Large language model (LLM) agents have demonstrated strong capability in sequential decision-making, yet they remain fundamentally reactive in long-horizon tasks. Unlike humans who employ "what-if" reasoning to evaluate potential plans before commitment, standard agents lack an internal world model to simulate future outcomes. Therefore, we propose to internalize future-aware planning by training a single autoregressive model to verbalize both a prospective state rollout and a plan-conditioned success estimate—a textual analogue of the Q-value. Crucially, we identify a format-capability gap: simply fine-tuning agents on look-ahead traces during post-training leads to superficial mimicry of foresight without genuine predictive grounding. To bridge this gap, we introduce a three-stage training paradigm: (i) World Model Agentic Mid-Training (WM-AMT) to inject latent predictive capabilities into the policy; (ii) Format-Eliciting SFT (FE-SFT) to structure this injected capability; and (iii) Foresight-Conditioned Reinforcement Learning (FC-RL) to refine the calibration and utility of the generated simulations. Evaluated on search and mathematical reasoning tasks, our approach consistently outperforms other training baselines. Our results demonstrate that effective internal world modeling in LLM agents requires a capability-first training pipeline to achieve grounded and calibrated foresight.

Long-horizon agentic tasks require large language models (LLMs) to iteratively retrieve, integrate, and maintain dispersed information across multi-turn interactions, but preserving all interaction histories leads to a continuously growing working context. Recent proactive context management methods allow models to edit their own working context with specialized tools, yet they still face three key limitations: (1) a limited toolset restricted to search, deletion, and summarization, with no support for global planning, long-term memory, and adaptive compression; (2) inefficient exploration that treats context management actions uniformly despite their heterogeneous impacts on final outcomes; and (3) coarse-grained credit assignment that assigns the final trajectory-level reward to all intermediate context editing actions during RL. To bridge these gaps, we introduce ContextPilot, a proactive context management framework for long-horizon agentic reasoning. Our approach systematically augments the toolset with planning, long-term memory, and soft context offloading tools. We further propose an RL method tailored for context management, which uses context and entropy variation to identify critical context control actions for branch sampling and estimates action-level advantages from all branched trajectories that pass through these actions. Experiments on long-context QA and deep search tasks show that ContextPilot achieves stronger performance with a more compact working context, outperforming existing baselines across various base models and benchmarks.

Large Language Models (LLMs) are evolving from performing end-to-end mathematical reasoning to integrating agentic intelligence. However, most existing math benchmarks evaluate only final answers. This outcome-oriented evaluation provides limited diagnostic value for identifying process-level failures or rigorous logic, failing to guide the transformation of LLMs into robust agents. To bridge this gap, we present a process-level benchmark designed to evaluate the inherent agentic mathematical reasoning abilities of LLMs. Our framework aligns problem-solving agentic behaviors with a structured taxonomy of reusable mathematical atomic capabilities. We design a comprehensive suite of planning, action, and feedback tasks across both textual and multimodal contexts, supported by an automated pipeline that synthesizes high-quality trajectories and produces fine-grained annotations via controlled LLM rewriting. Experiments reveal that models with stronger end-to-end accuracy can exhibit markedly different agentic capability profiles. This demonstrates that process-level evaluation is crucial for interpreting the true potential of LLMs and guiding the development of next-generation mathematical agents.

Despite the success of Large Language Models (LLMs) in long-context tasks, their reliability in recognizing unanswerable questions remains a critical challenge. In information-seeking scenarios, models often suffer from overconfidence and hallucinations when evidence is missing or misleading. In this paper, we propose FactGuard, a systematic multi-task workflow for generating evidence-grounded answerable and unanswerable questions. We introduce FactGuard-Bench, a benchmark comprising 25,220 examples with context lengths up to 128K, specifically targeting Lack of Evidence and Misleading Evidence scenarios. Evaluations of eleven representative open-source and proprietary LLMs reveal a significant reliability gap: models struggle to reject unanswerable questions as context scales, with the leading model achieving only 71.62% accuracy. By training with FactGuard-Bench, we improve overall accuracy to 81.17% and significantly enhance models’ reasoned rejection capabilities. Our work provides a scalable path toward building more faithful and trustworthy long-context LLMs. Our source is publicly available at https://github.com/biaoliu-kiritsugu/FactGuard.

Retrieval-Augmented Generation (RAG) has demonstrated significant potential in enhancing large language models (LLMs) by supplementing external knowledge. However, existing approaches focus primarily on retrieving isolated factual knowledge entities while neglecting the critical reasoning relationships. To address this limitation, Graph-Augmented Generation (GraphRAG) has emerged as an effective solution, which explicitly integrates structured knowledge graphs to support complex reasoning tasks. Although diverse graph construction methods have been explored, they typically rely on static, query-agnostic graphs constructed via fixed heuristics. We are thereby motivated to propose a query-centric retrieval framework that adaptively constructs a graph tailored to each query. However, it is challenging to accurately identify these latent relationships from queries to the corpus. Moreover, unifying multiple local-perspective connections into a globally coherent structured corpus introduces additional complexity. To this end, we introduce HyperRAG, a novel framework in the Hyperbolic space that captures both explicit entity-based links and implicit query-aware connections. Extensive experiments on three benchmark datasets demonstrate that HyperRAG consistently outperforms existing baselines.

Retrieval-Augmented Generation (RAG) enhances the response capabilities of language models by integrating external knowledge sources. However, document chunking as an important part of RAG system often lacks effective evaluation tools. This paper first analyzes why existing RAG evaluation benchmarks are inadequate for assessing document chunking quality, specifically due to evidence sparsity. Based on this conclusion, we propose HiCBench, which includes manually annotated multi-level document chunking points, synthesized evidence-dense question answer(QA) pairs, and their corresponding evidence sources. We also propose HiChunk, a hierarchical document structuring framework using fine-tuned LLMs and the Auto-Merge retrieval algorithm to enhance retrieval quality. Experiments demonstrate that HiCBench effectively evaluates the impact of different chunking methods across the entire RAG pipeline. Moreover, HiChunk achieves better chunking quality within reasonable time consumption, thereby enhancing the overall performance of RAG systems. Source code is available at https://github.com/TencentCloudADP/hichunk.

Retrieval-Augmented Generation (RAG) has long been a promising paradigm for enhancing large language models (LLMs) with external knowledge. Traditional embedding-based methods for graph construction can capture semantic similarity but struggle to establish fine-grained, interpretable logical relationships. Recently, Graph-enhanced RAG (GraphRAG) has gained increasing popularity for its capability in modeling logical relationships. However, graph construction requires extensive token consumption for triple extraction and summarization, making it costly and slow. Accordingly, we propose MeshRAG, a novel framework for mining efficient structures via hashing to enhance RAG. We adopt an inductive paradigm in which global graph structure emerges from local hash collisions rather than explicit symbolic extraction. By replacing neural embedding search with lightweight and bitwise operations, MeshRAG automates a simple and rapid graph construction process. Furthermore, the hash collision mechanism provides transparent evidence for logical connections and retrieval decisions. Experimental results show that MeshRAG outperforms existing baselines, while its graph construction requires no GPU resources or token budget and can structure over ten thousand chunks in a few minutes.

Role-Playing Agents (RPAs) have shown remarkable performance in various applications, yet they often struggle to recognize and appropriately respond to hard queries that conflict with their role-play knowledge. To investigate RPAs’ performance when faced with different types of conflicting requests, we develop an evaluation benchmark that includes contextual knowledge conflicting requests, parametric knowledge conflicting requests, and non-conflicting requests to assess RPAs’ ability to identify conflicts and refuse to answer appropriately without over-refusing. Through extensive evaluation, we find that most RPAs behave significant performance gaps toward different conflict requests. To elucidate the reasons, we conduct an in-depth representation-level analysis of RPAs under various conflict scenarios. Our findings reveal the existence of rejection regions and direct response regions within the model’s forwarding representation, and thus influence the RPA’s final response behavior. Therefore, we introduce a lightweight representation editing approach that conveniently shifts conflicting requests to the rejection region, thereby enhancing the model’s refusal accuracy. The extensive experiments validate the effectiveness of our editing method, improving RPAs’ refusal ability of conflicting requests while maintaining their general role-playing capabilities.

Role-playing is important for Large Language Models (LLMs) to follow diverse instructions while maintaining role identity and the role’s pre-defined ability limits. Existing role-playing datasets mostly contribute to controlling role style and knowledge boundaries, but overlook role-playing in instruction-following scenarios. We introduce a fine-grained role-playing and instruction-following composite benchmark, named RoleMRC, including: (1) Multi-turn dialogues between ideal roles and humans, including free chats or discussions upon given passages; (2) Role-playing machine reading comprehension, involving response, refusal, and attempts according to passage answerability and role ability; (3) More complex scenarios with nested, multi-turn and prioritized instructions. The final RoleMRC features a 10.2k role profile meta-pool, 37.9k well-synthesized role-playing instructions, and 1.4k testing samples. We develop a pipeline to quantitatively evaluate the fine-grained role-playing and instruction-following capabilities of several mainstream LLMs, as well as models that are fine-tuned on our data. Moreover, cross-evaluation on external role-playing datasets confirms that models fine-tuned on RoleMRC enhances instruction-following without compromising general role-playing and reasoning capabilities. We also probe the neural-level activation maps of different capabilities over post-tuned LLMs. Access to our RoleMRC, RoleMRC-mix and Codes: https://github.com/LuJunru/RoleMRC.

Evaluating the ability of large language models (LLMs) to process lengthy contexts is critical, especially for retrieving query-relevant information embedded within them. We introduce Sequential-NIAH, a benchmark specifically designed to evaluate the capability of LLMs to extract sequential information items (known as needles) from long contexts. The benchmark includes three needle generation pipelines: synthetic-temporal, real-temporal, and real-logical orders, with context lengths ranging from 8K to 128K, which comprises 14,000 samples (2,000 for testing). To facilitate the evaluation of this benchmark, we trained an evaluation model that assesses the correctness of LLM responses by comparing their completeness and sequential consistency against the ground truth, which provides a more reliable evaluation metric than GPT-4 or Claude. We conducted experiments on six well-known LLMs, revealing that even the best-performing model achieved a maximum accuracy of only 63.50% on test set of this benchmark. Further analysis highlights the growing challenges posed by increasing the context length or the number of needles, underscoring substantial room for improvement of LLMs. Additionally, noise analysis validates the reliability and challenge of the benchmark, making Sequential-NIAH an important reference for advancing research on long text information extraction capabilities of LLMs.

When carefully optimized by human experts, naive prompts can significantly enhance the task performance of large language models (LLMs). However, such expert-driven prompt optimizations are resource-intensive. To address this, some studies have proposed Automatic Prompt Optimization (APO), which refines naive prompts according to task outputs from in-box testing models, utilizing advanced LLMs (e.g., GPT-4) in an ad-hoc way. Although effective, current approaches face challenges in generalization and privacy risks. To overcome these limitations, we have developed the first large-scale Prompt Optimization Preference (POP) dataset, fine-tuned offline local LLM-based optimizers, and conducted fairly evaluations across various downstream models. Our method, named Free-from Instruction-oriented Prompt Optimization (FIPO), allows precise optimization of the core task instructions in naive prompts in a model-agnostic manner. FIPO uses a modular APO template that dynamically incorporates the naive task instructions, optional instruction responses, and optional ground truth to produce refined prompts. The POP dataset is meticulously constructed using advanced LLMs, undergoing rigorous cross-validation by human experts and analytical models. By leveraging insights from this dataset, along with Tulu2 models and diverse fine-tuning strategies, we validate the efficacy of the FIPO framework across five public benchmarks and six testing models. Our dataset and codes are available at: https://github.com/LuJunru/FIPO_Project.

Recent LLM-based Text-to-SQL methods usually suffer from significant performance degradation on “huge” databases and complex user questions that require multi-step reasoning. Moreover, most existing methods neglect the crucial significance of LLMs utilizing external tools and model collaboration. To address these challenges, we introduce MAC-SQL, a novel LLM-based multi-agent collaborative framework. Our framework comprises a core decomposer agent for Text-to-SQL generation with few-shot chain-of-thought reasoning, accompanied by two auxiliary agents that utilize external tools or models to acquire smaller sub-databases and refine erroneous SQL queries. The decomposer agent collaborates with auxiliary agents, which are activated as needed and can be expanded to accommodate new features or tools for effective Text-to-SQL parsing. In our framework, We initially leverage GPT-4 as the strong backbone LLM for all agent tasks to determine the upper bound of our framework. We then fine-tune an open-sourced instruction-followed model, SQL-Llama, by leveraging Code Llama 7B, to accomplish all tasks as GPT-4 does. Experiments show that SQL-Llama achieves a comparable execution accuracy of 43.94, compared to the baseline accuracy of 46.35 for vanilla GPT-4. At the time of writing, MAC-SQL+GPT-4 achieves an execution accuracy of 59.59 when evaluated on the BIRD benchmark, establishing a new state-of-the-art (SOTA) on its holdout test set.

Role-playing agents (RPAs) are garnering increasing interests as a novel form of conversational AI. While previous research has predominantly concentrated on their ability to portray specified characters, we argue from a user-centered perspective that RPAs’ capability to advance the plot requires substantial improvements to deliver more engaging interaction. To bridge this gap, we propose RolePlot, a role-playing framework specifically designed to evaluate and enhance the plot-progression capabilities of RPAs. RolePlot begins by constructing a plot-progression dataset extended from human-written literary scripts and specially designed synthetic data, followed by narrative theory-driven manual annotation and automated labeling validated through human verification. We then exploit the over-parameterized embedding space of LLMs to detect a “trigger subspace” that identifies dialogue segments catalyzing plot transitions. When user’s inputs align with this subspace, we explicitly prompt RPAs to advance the plot. For evaluation, we simulate User-RPA interactions and track both the conversation longevity (measured in dialogue turns before disengagement) and users’ arousal levels across different stages. Empirically, our method improves RPAs’ capability to time plot developments, and more importantly, yielding a significant increase in conversation turns and sustained higher arousal levels, thereby confirming that users experience more immersive engagements.

Knowledge distillation (KD) has been widely adopted to compress large language models (LLMs). Existing KD methods investigate various divergence measures including the Kullback-Leibler (KL), reverse Kullback-Leibler (RKL), and Jensen-Shannon (JS) divergences. However, due to limitations inherent in their assumptions and definitions, these measures fail to deliver effective supervision when few distribution overlap exists between the teacher and the student. In this paper, we show that the aforementioned KL, RKL, and JS divergences respectively suffer from issues of mode-averaging, mode-collapsing, and mode-underestimation, which deteriorates logits-based KD for diverse NLP tasks. We propose the Sinkhorn Knowledge Distillation (SinKD) that exploits the Sinkhorn distance to ensure a nuanced and precise assessment of the disparity between teacher and student distributions. Besides, profit by properties of the Sinkhorn metric, we can get rid of sample-wise KD that restricts the perception of divergence in each teacher-student sample pair. Instead, we propose a batch-wise reformulation to capture geometric intricacies of distributions across samples in the high-dimensional space. Comprehensive evaluation on GLUE and SuperGLUE, in terms of comparability, validity, and generalizability, highlights our superiority over state-of-the-art methods on all kinds of LLMs with encoder-only, encoder-decoder, and decoder-only architectures.

Direct Preference Optimization (DPO) has emerged as a prominent algorithm for the direct and robust alignment of Large Language Models (LLMs) with human preferences, offering a more straightforward alternative to the complex Reinforcement Learning from Human Feedback (RLHF). Despite its promising efficacy, DPO faces a notable drawback: “verbosity”, a common over-optimization phenomenon also observed in RLHF. While previous studies mainly attributed verbosity to biased labels within the data, we propose that the issue also stems from an inherent algorithmic length reliance in DPO. Specifically, we suggest that the discrepancy between sequence-level Kullback–Leibler (KL) divergences between chosen and rejected sequences, used in DPO, results in overestimated or underestimated rewards due to varying token lengths. Empirically, we utilize datasets with different label lengths to demonstrate the presence of biased rewards. We then introduce an effective downsampling approach, named SamPO, to eliminate potential length reliance. Our experimental evaluations, conducted across three LLMs of varying scales and a diverse array of conditional and open-ended benchmarks, highlight the efficacy of SamPO in mitigating verbosity, achieving improvements of 5% to 12% over DPO through debaised rewards. Our code can be accessed at: https://github.com/LuJunru/SamPO/.

In this paper, we propose a novel span-level model for Aspect-Based Sentiment Analysis (ABSA), which aims at identifying the sentiment polarity of the given aspect. In contrast to conventional ABSA models that focus on modeling the word-level dependencies between an aspect and its corresponding opinion expressions, in this paper, we propose Table Filling BERT (TF-BERT), which considers the consistency of multi-word opinion expressions at the span-level. Specially, we learn the span representations with a table filling method, by constructing an upper triangular table for each sentiment polarity, of which the elements represent the sentiment intensity of the specific sentiment polarity for all spans in the sentence. Two methods are then proposed, including table-decoding and table-aggregation, to filter out target spans or aggregate each table for sentiment polarity classification. In addition, we design a sentiment consistency regularizer to guarantee the sentiment consistency of each span for different sentiment polarities. Experimental results on three benchmarks demonstrate the effectiveness of our proposed model.