We investigate how embedding dimension affects the emergence of an internal "world model" in a transformer trained with reinforcement learning to perform bubble-sort-style adjacent swaps. Models achieve high accuracy even with very small embedding dimensions, but larger dimensions yield more faithful, consistent, and robust internal representations. In particular, higher embedding dimensions strengthen the formation of structured internal representation and lead to better interpretability. After hundreds of experiments, we observe two consistent mechanisms: (1) the last row of the attention weight matrix monotonically encodes the global ordering of tokens; and (2) the selected transposition aligns with the largest adjacent difference of these encoded values. Our results provide quantitative evidence that transformers build structured internal world models and that model size improves representation quality in addition to end performance. We release our metrics and analyses, which can be used to probe similar algorithmic tasks.

As large language model (LLM) agents increasingly automate complex web tasks, they boost productivity while simultaneously introducing new security risks. However, relevant studies on web agent attacks remain limited. Existing red-teaming approaches mainly rely on manually crafted attack strategies or static models trained offline. Such methods fail to capture the underlying behavioral patterns of web agents, making it difficult to generalize across diverse environments. In web agent attacks, success requires the continuous discovery and evolution of attack strategies. To this end, we propose Genesis, a novel agentic framework composed of three modules: Attacker, Scorer, and Strategist. The Attacker generates adversarial injections by integrating the genetic algorithm with a hybrid strategy representation. The Scorer evaluates the target web agent's responses to provide feedback. The Strategist dynamically uncovers effective strategies from interaction logs and compiles them into a continuously growing strategy library, which is then re-deployed to enhance the Attacker's effectiveness. Extensive experiments across various web tasks show that our framework discovers novel strategies and consistently outperforms existing attack baselines.

Imaging technologies are fundamental to biomedical research and modern medicine, requiring analysis of high-resolution images across various modalities. While multimodal large language models (MLLMs) show promise for biomedical image analysis, most are designed for low-resolution images from general-purpose datasets, risking critical information loss. We investigate how image resolution affects MLLM performance in biomedical applications and demonstrate that: (1) native-resolution training and inference significantly improve performance across multiple tasks, (2) misalignment between training and inference resolutions severely degrades performance, and (3) mixed-resolution training effectively mitigates misalignment and balances computational constraints with performance requirements. Based on these findings, we recommend prioritizing native-resolution inference and mixed-resolution datasets to optimize biomedical MLLMs for transformative impact in scientific research and clinical applications.

Medical question answering (QA) requires extensive access to domain-specific knowledge. A promising direction is to enhance large language models (LLMs) with external knowledge retrieved from medical corpora or parametric knowledge stored in model parameters. Existing approaches typically fall into two categories: Retrieval-Augmented Generation (RAG), which grounds model reasoning on externally retrieved evidence, and Generation-Augmented Generation (GAG), which depends solely on the models internal knowledge to generate contextual documents. However, RAG often suffers from noisy or incomplete retrieval, while GAG is vulnerable to hallucinated or inaccurate information due to unconstrained generation. Both issues can mislead reasoning and undermine answer reliability. To address these challenges, we propose MedRGAG, a unified retrieval-generation augmented framework that seamlessly integrates external and parametric knowledge for medical QA. MedRGAG comprises two key modules: Knowledge-Guided Context Completion (KGCC), which directs the generator to produce background documents that complement the missing knowledge revealed by retrieval; and Knowledge-Aware Document Selection (KADS), which adaptively selects an optimal combination of retrieved and generated documents to form concise yet comprehensive evidence for answer generation. Extensive experiments on five medical QA benchmarks demonstrate that MedRGAG achieves a 12.5% improvement over MedRAG and a 4.5% gain over MedGENIE, highlighting the effectiveness of unifying retrieval and generation for knowledge-intensive reasoning. Our code and data are publicly available at https://anonymous.4open.science/r/MedRGAG

Food insecurity remains a persistent public health emergency in the United States, tightly interwoven with chronic disease, mental illness, and opioid misuse. Yet despite the existence of thousands of food banks and pantries, access remains fragmented: 1) current retrieval systems depend on static directories or generic search engines, which provide incomplete and geographically irrelevant results; 2) LLM-based chatbots offer only vague nutritional suggestions and fail to adapt to real-world constraints such as time, mobility, and transportation; and 3) existing food recommendation systems optimize for culinary diversity but overlook survival-critical needs of food-insecure populations, including immediate proximity, verified availability, and contextual barriers. These limitations risk leaving the most vulnerable individuals, those experiencing homelessness, addiction, or digital illiteracy, unable to access urgently needed resources. To address this, we introduce Food4All, the first multi-agent framework explicitly designed for real-time, context-aware free food retrieval. Food4All unifies three innovations: 1) heterogeneous data aggregation across official databases, community platforms, and social media to provide a continuously updated pool of food resources; 2) a lightweight reinforcement learning algorithm trained on curated cases to optimize for both geographic accessibility and nutritional correctness; and 3) an online feedback loop that dynamically adapts retrieval policies to evolving user needs. By bridging information acquisition, semantic analysis, and decision support, Food4All delivers nutritionally annotated and guidance at the point of need. This framework establishes an urgent step toward scalable, equitable, and intelligent systems that directly support populations facing food insecurity and its compounding health risks.

Braille plays a vital role in education and information accessibility for visually impaired individuals. However, Braille information processing faces challenges such as data scarcity and ambiguities in mixed-text contexts. We construct English and Chinese Braille Mixed Datasets (EBMD/CBMD) with mathematical formulas to support diverse Braille domain research, and propose a syntax tree-based augmentation method tailored for Braille data. To address the underperformance of traditional fine-tuning methods in Braille-related tasks, we investigate Braille Knowledge-Based Fine-Tuning (BKFT), which reduces the learning difficulty of Braille contextual features. BrailleLLM employs BKFT via instruction tuning to achieve unified Braille translation, formula-to-Braille conversion, and mixed-text translation. Experiments demonstrate that BKFT achieves significant performance improvements over conventional fine-tuning in Braille translation scenarios. Our open-sourced datasets and methodologies establish a foundation for low-resource multilingual Braille research.

Unlike offline processing, streaming video vision-language models face two fundamental constraints: causality and accumulation. Causality prevents access to future frames that offline methods exploit, while accumulation causes tokens to grow unbounded, creating efficiency bottlenecks. However, existing approaches only regulate post-LLM kv-cache, leaving costly pre-LLM prefill unchanged. We introduce StreamingTOM, a training-free, plug-and-play two-stage framework that addresses both pre-LLM and post-LLM bottlenecks with predictable latency. Causal Temporal Reduction imposes a fixed per-frame budget and selects tokens based on adjacent-frame changes and token saliency, drastically reducing per-frame prefill cost by processing only a compact subset of visual tokens per frame instead of all visual tokens. Online Quantized Memory stores tokens in 4-bit format, retrieves relevant groups on demand, and dequantizes them, keeping the active kv-cache bounded regardless of stream length. Experiments demonstrate our method achieves $15.7\times$ kv-cache compression, $1.2\times$ lower peak memory and $2\times$ faster TTFT compared to prior SOTA. StreamingTOM maintains state-of-the-art accuracy among training-free methods with an average of $63.8\%$ on offline benchmarks and $55.8\%/3.7$ on RVS. These results highlight the practical benefits of our two-stage approach for efficient streaming video understanding with bounded growth.

Subject-driven image generation models face a fundamental trade-off between identity preservation (fidelity) and prompt adherence (editability). While online reinforcement learning (RL), specifically GPRO, offers a promising solution, we find that a naive application of GRPO leads to competitive degradation, as the simple linear aggregation of rewards with static weights causes conflicting gradient signals and a misalignment with the temporal dynamics of the diffusion process. To overcome these limitations, we propose Customized-GRPO, a novel framework featuring two key innovations: (i) Synergy-Aware Reward Shaping (SARS), a non-linear mechanism that explicitly penalizes conflicted reward signals and amplifies synergistic ones, providing a sharper and more decisive gradient. (ii) Time-Aware Dynamic Weighting (TDW), which aligns the optimization pressure with the model's temporal dynamics by prioritizing prompt-following in the early, identity preservation in the later. Extensive experiments demonstrate that our method significantly outperforms naive GRPO baselines, successfully mitigating competitive degradation. Our model achieves a superior balance, generating images that both preserve key identity features and accurately adhere to complex textual prompts.

Prompt Optimization has emerged as a crucial approach due to its capabilities in steering Large Language Models to solve various tasks. However, current works mainly rely on the random rewriting ability of LLMs, and the optimization process generally focus on specific influencing factors, which makes it easy to fall into local optimum. Besides, the performance of the optimized prompt is often unstable, which limits its transferability in different tasks. To address the above challenges, we propose DelvePO (Direction-Guided Self-Evolving Framework for Flexible Prompt Optimization), a task-agnostic framework to optimize prompts in self-evolve manner. In our framework, we decouple prompts into different components that can be used to explore the impact that different factors may have on various tasks. On this basis, we introduce working memory, through which LLMs can alleviate the deficiencies caused by their own uncertainties and further obtain key insights to guide the generation of new prompts. Extensive experiments conducted on different tasks covering various domains for both open-and closed-source LLMs, including DeepSeek-R1-Distill-Llama-8B, Qwen2.5-7B-Instruct and GPT -4o-mini. Experimental results show that DelvePO consistently outperforms previous SOT A methods under identical experimental settings, demonstrating its effectiveness and transferability across different tasks. The rapid advancement of Large Language Models (LLMs) (DeepSeek-AI, 2025; Li et al., 2025) has revolutionized various real-world applications (Shao et al., 2024; Zheng et al., 2025) . Prompt, a method that steers LLMs to produce desired results without modifying parameters, has garnered significant interest among non-AI experts from different domains (Wan et al., 2024; Guo et al., 2025; Fernando et al., 2024). Consequently, the rapid growth in users has increased demand for prompt engineering methods. Previous efforts primarily focused on manually designing specialized prompts (Brown et al., 2020; Kojima et al., 2022; Wei et al., 2023).

An important design question is which layer's attention matrix should be used to compute the attention scores. In Appendix B, we conduct ablation studies comparing early, middle, and deep layers, and conclude that using deeper layers generally yields better results. This finding aligns with prior studies (Aljaafari et al., 2024; Zheng et al., 2024; Rocchetti & Ferrara, 2025), which suggest that semantic representations become increasingly abstract across layers: shallow layers (closer to the input) primarily capture syntax, local structures, and surface-level patterns such as positional relations, bracket matching, and syntactic cues, while deeper layers (closer to the output) focus more on semantic abstraction, high-level concepts, and task-relevant global information, which are typically more influential for instruction following. Moreover, prior works in other domains that rely on attention scores for analysis (Chen et al., 2024; Y e et al., 2025) are often incompatible with efficient attention implementations such as FlashAttention (Dao et al., 2022). To avoid this, we design a lightweight solution: during the forward pass, we use a hook to store the hidden states of the target layer and then perform a simple recomputation of that layer to retrieve its attention matrix. This design eliminates the need to output full attention matrices during the complete forward pass, thereby making our algorithm fully compatible with efficient attention mechanisms like FlashAttention and ensuring training efficiency.