Large language models (LLMs) excel at solving complex tasks by executing agentic workflows composed of detailed instructions and structured operations. Yet, building general-purpose agents by manually embedding foundation models into agentic systems such as Chain-of-Thought, Self-Reflection, and ReACT through text interfaces limits scalability and efficiency. Recently, many researchers have sought to automate the generation and optimization of these workflows through code-based representations. However, existing methods often rely on labeled datasets to train and optimize workflows, making them ineffective and inflexible for solving real-world, dynamic problems where labeled data is unavailable. To address this challenge, we introduce Polymath, a self-optimizing agent with dynamic hierarchical workflow that leverages the flexibility of task flow graphs and the expressiveness of code-represented workflows to solve a wide range of real-world, dynamic problems. The proposed optimization methodology integrates multi-grid-inspired graph optimization with a self-reflection-guided evolutionary algorithm to refine workflows without labeled data. Experimental results on six benchmark datasets across coding, math, and multi-turn QA tasks show that Polymath achieves 8.1% average improvement over state-of-the-art baselines.

Recent advances in large language models (LLMs) have driven impressive progress in omni-modal understanding and generation. However, training omni-modal LLMs remains a significant challenge due to the heterogeneous model architectures required to process diverse modalities, necessitating sophisticated system design for efficient large-scale training. Existing frameworks typically entangle model definition with parallel logic, incurring limited scalability and substantial engineering overhead for end-to-end omni-modal training. We present VeOmni, a modular and efficient training framework to accelerate the development of omni-modal LLMs. VeOmni introduces model-centric distributed recipes that decouples communication from computation, enabling efficient 3D parallelism on omni-modal LLMs. VeOmni also features a flexible configuration interface supporting seamless integration of new modalities with minimal code change. Using VeOmni, a omni-modal mixture-of-experts (MoE) model with 30B parameters can be trained with over 2,800 tokens/sec/GPU throughput and scale to 160K context lengths via 3D parallelism on 128 GPUs, showcasing its superior efficiency and scalability for training large omni-modal LLMs.

The internet is rife with unattributed, deliberately misleading, or otherwise untrustworthy content. Though large language models (LLMs) are often tasked with autonomous web browsing, the extent to which they have learned the simple heuristics human researchers use to navigate this noisy environment is not currently known. In this paper, we introduce the Synthetic Media Literacy Test (SMeL Test), a minimal benchmark that tests the ability of language models to actively filter out untrustworthy information in context. We benchmark a variety of commonly used instruction-tuned LLMs, including reasoning models, and find that no model consistently succeeds; while reasoning in particular is associated with higher scores, even the best API model we test hallucinates up to 70% of the time. Remarkably, larger and more capable models do not necessarily outperform their smaller counterparts. We hope our work sheds more light on this important form of hallucination and guides the development of new methods to combat it.

Personalization of Large Language Models (LLMs) often assumes users hold static preferences that reflect globally in all tasks. In reality, humans hold dynamic preferences that change depending on the context. As users interact with an LLM in various contexts, they naturally reveal their contextual preferences, which a model must infer and apply in future contexts to ensure alignment. To assess this, we introduce CUPID, a benchmark of 756 human-curated interaction session histories between users and LLM-based chat assistants. In each interaction session, the user provides a request in a specific context and expresses their preference through multi-turn feedback. Given a new user request and prior interaction sessions, our benchmark assesses whether LLMs can infer the preference relevant to this request and generate a response that satisfies this preference. With CUPID, we evaluated 10 open and proprietary LLMs, revealing that state-of-the-art LLMs struggle to infer preferences from multi-turn interactions and fail to discern what previous context is relevant to a new request -- under 50% precision and 65% recall. Our work highlights the need to advance LLM capabilities for more contextually personalized interactions and proposes CUPID as a resource to drive these improvements.

Large Language Models (LLMs) are increasingly deployed in autonomous decision-making roles across high-stakes domains. However, since models are trained on human-generated data, they may inherit cognitive biases that systematically distort human judgment, including escalation of commitment, where decision-makers continue investing in failing courses of action due to prior investment. Understanding when LLMs exhibit such biases presents a unique challenge. While these biases are well-documented in humans, it remains unclear whether they manifest consistently in LLMs or require specific triggering conditions. This paper investigates this question using a two-stage investment task across four experimental conditions: model as investor, model as advisor, multi-agent deliberation, and compound pressure scenario. Across N = 6,500 trials, we find that bias manifestation in LLMs is highly context-dependent. In individual decision-making contexts (Studies 1-2, N = 4,000), LLMs demonstrate strong rational cost-benefit logic with minimal escalation of commitment. However, multi-agent deliberation reveals a striking hierarchy effect (Study 3, N = 500): while asymmetrical hierarchies show moderate escalation rates (46.2%), symmetrical peer-based decision-making produces near-universal escalation (99.2%). Similarly, when subjected to compound organizational and personal pressures (Study 4, N = 2,000), models exhibit high degrees of escalation of commitment (68.95% average allocation to failing divisions). These findings reveal that LLM bias manifestation depends critically on social and organizational context rather than being inherent, with significant implications for the deployment of multi-agent systems and unsupervised operations where such conditions may emerge naturally.

The rapid advancement of Large Language Model (LLM)-driven Graphical User Interface (GUI) agents is significantly hampered by the profound limitations of existing evaluation benchmarks in terms of accuracy, reproducibility, and scalability. To address this critical gap, we introduce NaturalGAIA, a novel benchmark engineered on the principle of Causal Pathways. This design paradigm structures complex tasks into a series of programmatically verifiable atomic steps, ensuring a rigorous, fully automated, and reproducible standard for assessment. Concurrently, to mitigate the inherent capability deficits of agents, we developed LightManus, a hierarchical agent architecture specifically optimized for long-horizon tasks. We leveraged this agent to generate a high-quality, human-verified trajectory dataset that uniquely captures diverse and even self-correcting interaction patterns of LLMs. We then utilized this dataset to perform Reinforcement Fine-Tuning (RFT) on the Qwen2.5-VL-7B model. Our experiments reveal that NaturalGAIA presents a formidable challenge to current state-of-the-art LLMs; even the top-performing Claude-sonnet-4 achieved a Weighted Pathway Success Rate (WPSR) of only 34.6%. Moreover, while RFT substantially improved the smaller model's GUI execution capabilities (WPSR increased from 3.3% to 10.8%), its performance degraded sharply when handling complex scenarios. This outcome highlights the inherent capability ceiling of smaller models when faced with comprehensive tasks that integrate perception, decision-making, and execution. This research contributes a rigorous evaluation standard and a high-quality dataset to the community, aiming to guide the future development of GUI agents.

Transformer-based architectures have become the prevailing backbone of large language models. However, the quadratic time and memory complexity of self-attention remains a fundamental obstacle to efficient long-context modeling. To address this limitation, recent research has introduced two principal categories of efficient attention mechanisms. Linear attention methods achieve linear complexity through kernel approximations, recurrent formulations, or fastweight dynamics, thereby enabling scalable inference with reduced computational overhead. Sparse attention techniques, in contrast, limit attention computation to selected subsets of tokens based on fixed patterns, block-wise routing, or clustering strategies, enhancing efficiency while preserving contextual coverage. This survey provides a systematic and comprehensive overview of these developments, integrating both algorithmic innovations and hardware-level considerations. In addition, we analyze the incorporation of efficient attention into largescale pre-trained language models, including both architectures built entirely on efficient attention and hybrid designs that combine local and global components. By aligning theoretical foundations with practical deployment strategies, this work aims to serve as a foundational reference for advancing the design of scalable and efficient language models.

We introduce SafeWork-R1, a cutting-edge multimodal reasoning model that demonstrates the coevolution of capabilities and safety. It is developed by our proposed SafeLadder framework, which incorporates large-scale, progressive, safety-oriented reinforcement learning post-training, supported by a suite of multi-principled verifiers. Unlike previous alignment methods such as RLHF that simply learn human preferences, SafeLadder enables SafeWork-R1 to develop intrinsic safety reasoning and self-reflection abilities, giving rise to safety `aha' moments. Notably, SafeWork-R1 achieves an average improvement of $46.54\%$ over its base model Qwen2.5-VL-72B on safety-related benchmarks without compromising general capabilities, and delivers state-of-the-art safety performance compared to leading proprietary models such as GPT-4.1 and Claude Opus 4. To further bolster its reliability, we implement two distinct inference-time intervention methods and a deliberative search mechanism, enforcing step-level verification. Finally, we further develop SafeWork-R1-InternVL3-78B, SafeWork-R1-DeepSeek-70B, and SafeWork-R1-Qwen2.5VL-7B. All resulting models demonstrate that safety and capability can co-evolve synergistically, highlighting the generalizability of our framework in building robust, reliable, and trustworthy general-purpose AI.

Large reasoning models achieve remarkable performance through extensive chain-of-thought generation, yet they suffer from a critical inefficiency: applying uniformly extensive reasoning regardless of problem complexity. We present Hierarchical Budget Policy Optimization (HBPO), a reinforcement learning framework that enables models to learn problem-specific reasoning depths without sacrificing capability. Unlike existing approaches that impose rigid constraints or rely on discrete mode selection, HBPO partitions the exploration space into budget-constrained hierarchies (512-2560 tokens), each with differentiated reward structures that preserve both efficiency incentives and reasoning capabilities. This design addresses a fundamental challenge in efficient reasoning training: traditional length penalties systematically bias models away from necessary long reasoning paths, causing exploration space collapse. Through hierarchical sampling and budget-aware rewards, HBPO maintains exploration diversity while teaching models to recognize when extended deliberation is warranted. Extensive experiments demonstrate that HBPO reduces average token usage by up to 60.6% while improving accuracy by 3.14% across four reasoning benchmarks. Most notably, HBPO exhibits emergent adaptive behavior where models automatically adjust reasoning depth based on problem complexity. Our results suggest that reasoning efficiency and capability are not inherently conflicting, and can be simultaneously optimized through appropriately structured hierarchical training that preserves exploration diversity. Advances in large reasoning models have led to impressive performance on complex reasoning tasks through chain-of-thought methodologies (OpenAI, 2024; DeepSeek-AI, 2025).

Recently, inspired by OpenAI-o1/o3 and Deepseek-R1, the R1-Style method based on reinforcement learning fine-tuning has received widespread attention from the community. Previous R1-Style methods mainly focus on mathematical reasoning and code intelligence. It is of great research significance to verify their advantages on more general multimodal data. Chart is an important multimodal data type with rich information, which brings important research challenges in complex reasoning. In this work, we introduce Chart-R1, a chart-domain vision-language model with reinforcement learning fine-tuning to enable complex chart reasoning. To support Chart-R1, we first propose a novel programmatic data synthesis technology to generate high-quality step-by-step chart reasoning data covering single- and multi-subcharts, which makes up for the lack of reasoning data in the chart domain. Then we develop a two-stage training strategy: Chart-COT with step-by-step chain-of-thought supervision, and Chart-RFT with numerically sensitive reinforcement fine-tuning. Chart-COT aims to decompose complex chart reasoning tasks into fine-grained, understandable subtasks through step-by-step supervision, which lays a good foundation for improving the reasoning level of reinforcement learning. Chart-RFT utilize the typical group relative policy optimization strategy, in which a relatively soft reward is adopted for numerical response to emphasize the numerical sensitivity in the chart domain. We conduct extensive experiments on open-source benchmarks and self-built chart reasoning dataset (\emph{i.e., ChartRQA}). Experimental results show that Chart-R1 has significant advantages compared to chart-domain methods, even comparable to open/closed source large-scale models (\emph{e.g., GPT-4o, Claude-3.5}).