To encourage diverse exploration in reinforcement learning (RL) for large language models (LLMs) without compromising accuracy, we propose Policy Split, a novel paradigm that bifurcates the policy into normal and high-entropy modes with a high-entropy prompt. While sharing model parameters, the two modes undergo collaborative dual-mode entropy regularization tailored to distinct objectives. Specifically, the normal mode optimizes for task correctness, while the high-entropy mode incorporates a preference for exploration, and the two modes learn collaboratively. Extensive experiments demonstrate that our approach consistently outperforms established entropy-guided RL baselines across various model sizes in general and creative tasks. Further analysis reveals that Policy Split facilitates dual-mode exploration, where the high-entropy mode generates distinct behavioral patterns to the normal mode, providing unique learning signals.

While Graphical User Interface (GUI) agents have shown promising performance in automated device interaction, they primarily depend on static parametric knowledge from pre-training or instruction tuning. This reliance fundamentally limits their ability to handle long-tailed tasks that require explicit procedural knowledge absent from model parameters, often forcing agents to resort to inefficient and brittle trial-and-error exploration. To mitigate this limitation, we introduce Proactive Document-Guided Action for GUI agents in dynamic, open-web environments, a novel paradigm that mirrors human problem-solving by enabling agents to autonomously search for relevant documentation to resolve long-tailed tasks. To evaluate agents' capability in this paradigm, we propose DocOS, a benchmark designed to assess document-guided problem solving in fully interactive environments. DocOS requires agents to autonomously navigate a web browser, locate relevant online documentation, comprehend procedural instructions, and faithfully ground them into executable GUI actions. Extensive experiments reveal that progress is strictly constrained by dual bottlenecks: agents struggle to reliably locate relevant information during proactive search and frequently fail to faithfully ground retrieved instructions into precise actions, pointing toward document-guided interaction as a crucial pathway for enabling self-evolving GUI agents in dynamic environments.

Repository-level code generation constitutes a fundamental building block for automated software development. Consequently, numerous benchmarks have been proposed to evaluate the capabilities of large language mod els (LLMs) in this domain. However, existing benchmarks are largely limited to a single programming language and a fixed granularity level. To address these two challenges, we introduce ReCode-bench, a multi-language benchmark covering 7 widely-used programming languages and comprising three repository-level code gener ation tasks. These tasks are designed around varying levels of requirement granularity and include full project creation from scratch as well as guided development based on structural or functional specifications. In the lat ter, intentionally introduced requirements are treated as positive noise to better reflect realistic development scenarios. To enhance LLM robustness across these tasks, we propose RepoGenesis, a GRPO-based reinforcement learning framework that incorporates 3 distinct reward signals: structural similarity to human-written repositories, syntactic correctness verified through abstract syntax tree analysis, and functional validity confirmed via unit test execution. We evaluated 8 LLMs on ReCode-bench and found that even Claude-Sonnet-4-currently among the strongest code generation models-achieves an average pass@1 score of less than 4% across the three tasks. However, after training with RepoGenesis, Qwen2.5-coder-7B-Instruct attains performance comparable to Claude-Sonnet-4 (>100B).(2)

Large Language Models (LLMs) have significantly advanced Machine Translation (MT), applying them to linguistically complex domains-such as Social Network Services, literature etc. In these scenarios, translations often require handling non-literal expressions, leading to the inaccuracy of MT metrics. To systematically investigate the reliability of MT metrics, we first curate a meta-evaluation dataset focused on non-literal translations, namely MENT. MENT encompasses four non-literal translation domains and features source sentences paired with translations from diverse MT systems, with 7,530 human-annotated scores on translation quality. Experimental results reveal the inaccuracies of traditional MT metrics and the limitations of LLM-as-a-Judge, particularly the knowledge cutoff and score inconsistency problem. To mitigate these limitations, we propose RATE, a novel agentic translation evaluation framework, centered by a reflective Core Agent that dynamically invokes specialized sub-agents. Experimental results indicate the efficacy of RATE, achieving an improvement of at least 3.2 points in combined system- and segment-level correlation with human judgments compared with current methods. Further experiments demonstrate the robustness of RATE to general-domain MT evaluation. Code and dataset are available at: https://github.com/BITHLP/RATE.

Current role-playing agents (RPAs) are typically constructed by imitating surface-level behaviors, but this approach lacks internal cognitive consistency, often causing out-of-character errors in complex situations. To address this, we propose Character-R1, a framework designed to provide comprehensive verifiable reward signals for effective role-aware reasoning, which are missing in recent studies. Specifically, our framework comprises three core designs: (1) Cognitive Focus Reward, which enforces explicit label-based analysis of 10 character elements (e.g., worldview) to structure internal cognition; (2) Reference-Guided Reward, which utilizes overlap-based metrics with reference responses as optimization anchors to enhance exploration and performance; and (3) Character-Conditioned Reward Normalization, which adjusts reward distributions based on character categories to ensure robust optimization across heterogeneous roles. Extensive experiments demonstrate that Character-R1 significantly outperforms existing methods in knowledge, memory and others.

Video–based world models have emerged along two dominant paradigms: video generation and 3D reconstruction. However, existing evaluation benchmarks either focus narrowly on visual fidelity and text–video alignment for generative models, or rely on static 3D reconstruction metrics that fundamentally neglect temporal dynamics. We argue that the future of world modeling lies in 4D generation, which jointly models spatial structure and temporal evolution. In this paradigm, the core capability is interactive response: the ability to faithfully reflect how interaction actions drive state transitions across space and time. Yet no existing benchmark systematically evaluates this critical dimension. To address this gap, we propose Omni–WorldBench, a comprehensive benchmark specifically designed to evaluate the interactive response capabilities of world models in 4D settings. Omni–WorldBench comprises two key components: Omni–WorldSuite, a systematic prompt suite spanning diverse interaction levels and scene types; and Omni–Metrics, an agent-based evaluation framework that quantifies world modeling capabilities by measuring the causal impact of interaction actions on both final outcomes and intermediate state evolution trajectories. We conduct extensive evaluations of 18 representative world models across multiple paradigms. Our analysis reveals critical limitations of current world models in interactive response, providing actionable insights for future research. Omni-WorldBench will be publicly released to foster progress in interactive 4D world modeling.

Reinforcement learning with verifiable rewards (RLVR) has driven substantial progress in training LLMs for reasoning tasks, but representative methods such as GRPO assign uniform credit across all tokens, wasting gradient on routine tokens while under-crediting pivotal reasoning steps. Existing token-level credit assignment methods require resources beyond the model's own rollouts. GRPO variants rely on process reward models or ground-truth answers. Knowledge distillation assigns credit through per-token divergence but requires external teachers (On-Policy Distillation) or privileged information (On-Policy Self Distillation). However, these dependencies limit applicability in the pure RLVR setting. We observe that conditioning the model on its own verified trajectories induces a measurable per-token KL divergence between the original and conditioned distributions, and prove that distilling from a self-teacher constructed by verified trajectories leads to infeasible weighted-average solutions when multiple verified trajectories exist. We propose SC-GRPO (Self-Conditioned GRPO), which uses KL divergence mentioned before as a multiplicative weight on GRPO gradients. Across five benchmarks spanning math, code, and agentic tasks, SC-GRPO consistently outperforms 8.1

Real-world GUI usage frequently involves workflows that span multiple devices and platforms, requiring the transfer of intermediate results, maintenance of shared state, and coordination across heterogeneous environments. However, existing GUI benchmarks overwhelmingly evaluate agents on single-device, statically defined tasks, thus leaving such cross-device capabilities largely unexamined, resulting in an overly optimistic assessment of agents' readiness for real-world usage. We introduce JarvisGUI, a dynamic benchmark that evaluates GUI agents on cross-device workflows requiring coordinated interaction across heterogeneous platforms, including Android, Windows, and Ubuntu. Specifically, JarvisGUI formulates GUI tasks as input-output transformations under a lightweight type system, which allows us to automatically compose multi-step, cross-device workflows and dynamically evaluate agent performance within a unified framework. By evaluating agents in virtual environments spanning multiple operating systems, JarvisGUI reveals that state-of-the-art open-source GUI agents struggle with the state-transfer awareness, cross-platform contextual reasoning, and long-horizon dependency management required for real-world workflows, exposing a critical capability gap invisible to existing benchmarks.

Recent advancements in Large Language Models (LLMs) have empowered home assistants with natural language interaction capabilities. However, current assistants overlook the progressive omission that occurs in human dialogue as shared context accumulates, leading to more elliptical expressions for efficient communication. Thus, current assistants still struggle to interpret such elliptical expressions accurately, which limits their effectiveness in real-world applications. In practical smart home scenarios, assistants face two major challenges caused by elliptical commands: (1) referential ambiguity caused by different environmental expectations among multiple users; and (2) intention ambiguity resulting from user preferences that evolve over time or change with the environment. To address these challenges, we introduce PEC-Home, the first simulated home dataset specifically designed for interpreting progressively elliptical commands in smart homes. Extensive experiments on various LLMs, including GPT-4o, show that existing home assistants struggle to execute user-intended operations based solely on elliptical commands. Even when equipped with tools for storing and retrieving user dialogue history, execution accuracy remains below that achieved with complete commands.}.

A large language model (LLM) agent is a model coupled to a harness: parametric weights that hold knowledge and reasoning ability, wrapped in an external apparatus of memory, tools, and orchestration through which the model perceives and acts. Progress over the past years has come from pushing on both, with stronger models and richer harnesses, but the design decision that quietly governs an agent’s behavior is where competence should live: which knowledge, skills, and control to internalize into the model’s parameters, and which to externalize onto the surrounding harness. These decisions increasingly determine whether an agent is reliable, efficient, and able to keep improving. We investigate this core problem through the lens of the Theory of Agent (ToA), which views an LLM agent not as a task executor but as an epistemic decision-maker in terms of reasoning or acting at every step. The two means it has for that epistemic uncertainty reduction, namely reasoning over what it already holds and acting on the world to obtain what it does not, are the run-time faces of internalization and externalization: reasoning draws on competence that has been absorbed into the model, while acting exploits competence that remains in the harness. We review this internalization and externalization across five chapters. Foundations sets up the basic picture: an agent that, at each step, decides whether to answer from what it already knows or reach outside for what it does not, and what separates a good version of that decision from a wasteful one. Learning examines how training places competence on either side of the boundary, spanning internalization of memory, reasoning, tool-use policy, and world models, and externalization of memory, skills, and orchestration. Aligning treats calibration as the connecting thread across an agent’s cognitive boundaries, efficiency constraints, human preferences, environment constraints, and other agents. Evolving studies how agents revise both sides of the boundary over time, from model-side and harness-side updates to agent–environment co-evolution. Evaluation reinterprets metrics and benchmarks by the epistemic effort they expose and argues for measuring the boundary decision rather than task completion alone. We conclude with the open problems that follow once internalization and externalization are treated as the governing axis of agent design.

Terminal agents extend Large Language Models with the ability to execute tasks directly in command-line environments, but their progress is bottlenecked by the scarcity of high-quality training data. Existing approaches bootstrap from partial sources such as human-defined seeds or GitHub repositories to instantiate one component and then complete the rest, producing tasks confined to narrow seed distributions, environments misaligned with task semantics, and inefficient trajectories from unguided exploration. To address these limitations, we introduce Terminal-World, a fully automated pipeline that uses agent skills as the central synthesis primitive, which jointly encode what to accomplish, when to apply (preconditions and environment state), and how to execute, enabling task instructions, environments, and teacher trajectories to be co-derived. To further broaden the synthesis space, Terminal-World composes skills into skill teams and skill graphs for multi-role and cross-domain task synthesis. Using this pipeline, we construct 5,723 training environments and train Terminal-World-8B/14B/32B, evaluated across 6 benchmarks where the Terminal-World series consistently outperforms terminal-agent baselines. Notably, using the same teacher model and only 1.2

Embodied Action Sequence Planning focuses on the capability of embodied agents to implement action planning via environmental perception. This technology enables diverse intelligent assistance for real-world scenarios such as home and office environments. To address the limitations of existing embodied agents in meeting the requirement for proactivity and achieving joint understanding of visual and audio information, this study investigates the ability of embodied agents to proactively provide assistance through action sequence planning based on joint understanding of vision and audio perception without explicit human instructions. Correspondingly, we propose PEAP, the first multimodal proactive embodied action sequence planning dataset. We evaluate the performance of multiple Large Language Models on the PEAP dataset. The results demonstrate that these models still exhibit significant deficiencies on this task particularly lacking accurate environmental perception capabilities. Furthermore, ablation experiment and replacement experiment further corroborate that the joint understanding of multimodal information can significantly improve the models’ performance on proactive embodied action sequence planning task.

GUI agents today assume a static screen, where the world is frozen between two actions. However, real interfaces such as short-video applications violate this assumption, as their content keeps playing, and a competent user must decide what to watch and for how long. We formalize this task as Living-Screen-Native GUI agents and introduce LivingScreen, the first benchmark instantiating it on short-video platforms, with a faithful browser-based environment, a three-tier task suite, and metrics that jointly score accuracy and information efficiency. Evaluating extensive frontier models, we find that none reaches the human cost-accuracy performance, and that their dominant failure mode is over- and under-observation, pointing to observation control as a missing capability axis for future GUI agents. All data and code will be available at https://github.com/BITHLP/LivingScreen.

Grounded Multimodal Named Entity Recognition (GMNER) aims to extract text-based entities, assign them semantic categories, and ground them to corresponding visual regions. In this work, we explore the potential of Multimodal Large Language Models (MLLMs) to perform GMNER in an end-to-end manner, moving beyond their typical role as auxiliary tools within cascaded pipelines. Crucially, our investigation reveals a fundamental challenge: MLLMs exhibit modality bias, including visual bias and textual bias, which stems from their tendency to take unimodal shortcuts rather than rigorous cross-modal verification. To address this, we propose Modality-aware Consistency Reasoning (MCR), which enforces structured cross-modal reasoning through Multi-style Reasoning Schema Injection (MRSI) and Constraint-guided Verifiable Optimization (CVO). MRSI transforms abstract constraints into executable reasoning chains, while CVO empowers the model to dynamically align its reasoning trajectories with Group Relative Policy Optimization (GRPO). Experiments on GMNER and visual grounding tasks demonstrate that MCR effectively mitigates modality bias and achieves superior performance compared to existing baselines.

While large language model–powered agents can self-evolve by accumulating experience or by dynamically creating new assets (i.e., tools or expert agents), existing frameworks typically treat these two evolutionary processes in isolation. This separation overlooks their intrinsic interdependence: the former is inherently bounded by a manually predefined static toolset, while the latter generates new assets from scratch without experiential guidance, leading to limited capability growth and unstable evolution. To address this limitation, we introduce a novel paradigm of co-evolutionary Capability Expansion and Experience Distillation. Guided by this paradigm, we propose the **Mem 2 Evolve**, which integrates two core components: **Experience Memory** and **Asset Memory**. Specifically, Mem 2 Evolve leverages accumulated experience to guide the dynamic creation of assets, thereby expanding the agent’s capability space while simultaneously acquiring new experience to achieve co-evolution. Extensive experiments across 6 task categories and 8 benchmarks demonstrate that Mem 2 Evolve achieves improvement of 18.53% over standard LLMs, 11.80% over agents evolving solely through experience, and 6.46% over those evolving solely through asset creation, establishing it as a substantially more effective and stable self-evolving agent framework.

To overcome the sparse reward challenge in reinforcement learning (RL) for agents based on large language models (LLMs), we propose Mutual Information Self-Evaluation (MISE), an RL paradigm that utilizes hindsight generative self-evaluation as dense reward signals while simultaneously calibrating them against the environmental feedbacks. Empirically, MISE enables an agent to learn autonomously from dense internal rewards supplementing sparse extrinsic signals. Theoretically, our work provides the first formal foundation for the paradigm of generative self-rewarding. We prove that utilizing hindsight self-evaluation rewards is equivalent to minimizing an objective that combines mutual information with a KL divergence term between the policy and a proxy reward policy. This theoretical insight then informs and justifies our calibration step, which actively aligns these rewards with the optimal policy. Extensive experiments show that MISE outperforms strong baselines, enabling open-source LLMs about 7B parameters to achieve performance comparable to GPT-4o on validation without expert supervision.

LLM hidden states are ordinary vectors, but the distances among those vectors may still show hierarchical structure. To our knowledge, this paper is the first systematic study of whether prompt-token hidden states in contemporary LLMs exhibit Gromov Hyperbolicity (GH), a distance-based measure of tree-likeness. Using 818,904 sample-layer measurements from ten open-weight models across MATH500, HumanEval, WinoGrande, and TruthfulQA, we build a GH map over four axes: parameter scale, layer depth, model family, and input domain. The clearest pattern is depth, not scale: middle layers usually form a high-relative-hyperbolicity plateau, while final layers often become substantially more tree-like. Scale effects are weak and non-monotonic, matched 7/8B model families differ strongly, and domains interact with model specialization. These findings make GH useful as a practical diagnostic: it shows where hierarchical distance structure appears, how specialization changes it, and which model-layer-domain comparisons deserve closer analysis.

Multimodal Large Language Models (MLLMs) have made remarkable progress in video understanding and consistently perform well on vision-centric benchmarks. However, existing benchmarks primarily evaluate factual or event-based comprehension, while neglecting audience insights. It is a critical yet underexplored dimension of video understanding, reflecting a deep comprehension of cognitive processes from the audience’s perspective. As a result, MLLMs, shaped by such benchmarks, often produce responses that are factually correct but misaligned with audience’s interests. To bridge this gap, we leverage audience insights derived from video comments as a direct proxy to guide the annotation process and introduce A3Bench, an audience-aligned benchmark for evaluating video audience insights with large-scale videos and high-quality multilingual comments. Furthermore, inspired by neuro-imaging studies, we propose Cognition Interaction of Thought (CIoT), a structured reasoning framework that emulates key aspects of cognitive processes. Extensive experiments on A3Bench reveal that current MLLMs struggle to understand audience insights, particularly compared to human-level understanding. In contrast, CIoT can improve the performance of these models, highlighting its potential to enhance the MLLMs’ capability of understanding audience insights in future research.

Despite large language models (LLMs) have achieved impressive achievements across numerous tasks, supervised fine-tuning (SFT) remains essential for adapting these models to specialized domains. However, SFT for domain specialization can be resource-intensive and sometimes leads to a deterioration in performance over general capabilities due to catastrophic forgetting (CF). To address these issues, we propose a self-adaptive gradient-aware data selection approach (GrADS) for supervised fine-tuning of LLMs, which identifies effective subsets of training data by analyzing gradients distribution obtained from a preliminary training phase. The purpose of GrADS is not to eliminate noisy data from the training set, but to obtain a diverse subset that is most conducive to efficient training. Through extensive experimentation with various LLMs across diverse domains such as medicine, law, and finance, GrADS has demonstrated significant efficiency and cost-effectiveness. Remarkably, utilizing merely 5% of the selected GrADS data, LLMs already surpass the performance of those fine-tuned on the entire dataset, and increasing to 50% of the data results in significant improvements! With catastrophic forgetting substantially mitigated simultaneously.2