The reasoning, writing, text-editing, and retrieval capabilities of proprietary large language models (LLMs) have advanced rapidly, providing users with an ever-expanding set of functionalities. However, this growing utility has also led to a serious societal concern: the over-reliance on LLMs. In particular, users increasingly delegate tasks such as homework, assignments, or the processing of sensitive documents to LLMs without meaningful engagement. This form of over-reliance and misuse is emerging as a significant social issue. In order to mitigate these issues, we propose a method injecting imperceptible phantom tokens into documents, which causes LLMs to generate outputs that appear plausible to users but are in fact incorrect. Based on this technique, we introduce TRAPDOC, a framework designed to deceive over-reliant LLM users. Through empirical evaluation, we demonstrate the effectiveness of our framework on proprietary LLMs, comparing its impact against several baselines. TRAPDOC serves as a strong foundation for promoting more responsible and thoughtful engagement with language models. Our code is available at https://github.com/jindong22/TrapDoc.

Recent advancements in reinforcement learning with verifiable rewards have pushed the boundaries of the visual reasoning capabilities in large vision-language models (L VLMs). However, training L VLMs with reinforcement fine-tuning (RFT) is computationally expensive, posing a significant challenge to scaling model size. Recent advances in large language models have led to the development of systems capable of extended reasoning and deliberation, often referred to as "slow-thinking" models, such as OpenAI-o1 (Jaech et al., 2024), DeepSeek-R1 (Guo et al., 2025), and QwQ (Team, 2025). Unlike "fast-thinking" models such as GPT -4o (Hurst et al., 2024), "slow-thinking" models usually engage in multi-step self-reflection to produce an answer that resembles the thorough thinking process that humans make before producing a final answer for non-trivial problems. These models have achieved remarkable success in complex problem-solving benchmarks, particularly in mathematical and scientific reasoning domains (Shao et al., 2024b; Zeng et al., 2025; Y u et al., 2025). Recent research has also extended such reflective reasoning to multimodal tasks (Huang et al., 2025; Deng et al., 2025; Y ang et al., 2025; Zhou et al., 2025; Wang et al., 2025a), pushing large vision-language models (L VLMs) toward greater performance in scenarios that require structured and contextual understanding across modalities. Many of the most effective "slow-thinking" models rely on reinforcement learning with verifiable rewards (RL VR) (Face, 2025; Su et al., 2025; Wei et al., 2025a), a reinforcement fine-tuning (RFT) framework that encourages the model to generate intermediate reasoning steps that lead to a correct answer for automatically verifiable tasks.

Large language models (LLMs) have demonstrated remarkable performance, but their long-context reasoning remains constrained by the excessive memory required for the Key-Value (KV) cache. This makes KV cache compression a critical step toward efficient long-context inference. Recent methods have explored low-rank techniques to reduce the hidden size of the KV cache. However, they neglect the distinct roles and varying importance of Keys and Values, leading to significant performance drops under high compression. To address this, we propose ReCalKV, a post-training low-rank KV cache compression approach with tailored strategies for Keys and Values. For Keys, we propose Head-wise Similarity aware Reordering (HSR), which clusters structurally similar heads into groups, enabling more accurate low-rank approximation via grouped SVD. For Values, we propose Offline Value Calibration (OVC), which efficiently calibrates the value projection matrix using calibration data without training, ensuring an accurate representation of contextual information. Extensive experiments show that ReCalKV consistently outperforms existing low-rank compression methods, achieving high compression ratios with minimal performance loss. The code and models will be available at:https://github.com/XIANGLONGYAN/ReCalKV.

Weak-to-strong generalization (W2SG) refers to the phenomenon where a strong student model, trained on a dataset labeled by a weak teacher, ultimately outperforms the teacher on the target task. Recent studies attribute this performance gain to the prediction misfit between the student and teacher models. In this work, we theoretically investigate the emergence of W2SG through a generalized bias-variance decomposition of Bregman divergence. Specifically, we show that the expected population risk gap between the student and teacher is quantified by the expected misfit between the two models. While this aligns with previous results, our analysis removes several restrictive assumptions, most notably, the convexity of the student's hypothesis class, required in earlier works. Moreover, we show that W2SG is more likely to emerge when the student model approximates its posterior mean teacher, rather than mimicking an individual teacher. Using a concrete example, we demonstrate that if the student model size is sufficiently large, it can indeed converge to the posterior mean teacher in expectation. Our analysis also suggests that avoiding overfitting to the teacher's supervision and reducing the entropy of student's prediction further facilitate W2SG. In addition, we show that the reverse cross-entropy loss, unlike the standard forward cross-entropy, is less sensitive to the predictive uncertainty of the teacher. Finally, we empirically verify our theoretical insights and demonstrate that incorporating the reverse cross-entropy loss consistently improves student performance.

Modern language models generate chain-of-thought traces by autoregressively sampling tokens from a finite vocabulary. While this discrete sampling has achieved remarkable success, conducting chain-of-thought with continuously-valued tokens (CoT2) offers a richer and more expressive alternative. Our work provides new theoretical guarantees and algorithms for CoT2, motivated by logical reasoning tasks that inherently require search capabilities. Theoretically, we establish how CoT2 facilitates the model to track multiple discrete traces in parallel; and quantify the level of achievable parallelism and its benefits for inference efficiency. We also provide a CoT2-based one-layer transformer construction that solves the combinatorial "subset sum problem" given a sufficient embedding dimension. These insights arise from a novel and effective supervision strategy where we match the language model outputs to the empirical token distributions of a set of target traces. Complementing this, we introduce sampling strategies that unlock policy optimization methods for CoT2. Our primary strategy samples and composes $K$ discrete tokens at each decoding step to control the level of parallelism. Experiments confirm that (i) the optimal level of parallelism is governed by the embedding dimension, (ii) our continuous supervision strategy can outperform alternative methods, and (iii) policy optimization with CoT2 indeed improves the performance of the model beyond its initial discrete or continuous supervision.

Mainstream large vision-language models (LVLMs) inherently encode cultural biases, highlighting the need for diverse multimodal datasets. To address this gap, we introduce PEARL, a large-scale Arabic multimodal dataset and benchmark explicitly designed for cultural understanding. Constructed through advanced agentic workflows and extensive human-in-the-loop annotations by 37 annotators from across the Arab world, PEARL comprises over 309K multimodal examples spanning ten culturally significant domains covering all Arab countries. We further provide two robust evaluation benchmarks (PEARL and PEARL-LITE) along with a specialized subset (PEARL-X) explicitly developed to assess nuanced cultural variations. Comprehensive evaluations on state-of-the-art open and proprietary LVLMs demonstrate that reasoning-centric instruction alignment substantially improves models' cultural grounding compared to conventional scaling methods. PEARL establishes a foundational resource for advancing culturally-informed multimodal modeling research. All datasets and benchmarks are publicly available.

Text-to-image generation increasingly demands access to domain-specific, fine-grained, and rapidly evolving knowledge that pretrained models cannot fully capture, necessitating the integration of retrieval methods. Existing Retrieval-Augmented Generation (RAG) methods attempt to address this by retrieving globally relevant images, but they fail when no single image contains all desired elements from a complex user query. We propose Cross-modal RAG, a novel framework that decomposes both queries and images into sub-dimensional components, enabling subquery-aware retrieval and generation. Our method introduces a hybrid retrieval strategy - combining a sub-dimensional sparse retriever with a dense retriever - to identify a Pareto-optimal set of images, each contributing complementary aspects of the query. During generation, a multimodal large language model is guided to selectively condition on relevant visual features aligned to specific subqueries, ensuring subquery-aware image synthesis. Extensive experiments on MS-COCO, Flickr30K, WikiArt, CUB, and ImageNet-LT demonstrate that Cross-modal RAG significantly outperforms existing baselines in the retrieval and further contributes to generation quality, while maintaining high efficiency.

Processing long contexts has become a critical capability for modern large language models (LLMs). Existing works leverage agent-based divide-and-conquer methods for processing long contexts. But these methods face crucial limitations, including prohibitive accumulated latency and amplified information loss from excessive agent invocations, and the disruption of inherent textual dependencies by immoderate partitioning. In this paper, we propose a novel multi-agent framework XpandA (Expand-Agent) coupled with question-driven workflow and dynamic partitioning for robust long-context processing. XpandA overcomes these limitations through: 1) dynamic partitioning of long texts, which adaptively modulates the filling rate of context windows for input sequences of vastly varying lengths; 2) question-guided protocol to update flat information ensembles within centralized shared memory, constructing consistent inter-agent knowledge across partitions; and 3) selectively replaying specific partitions based on the state-tracking of question-information couples to promote the resolution of inverted-order structures across partitions (e.g., flashbacks). We perform a comprehensive evaluation of XpandA on multiple long-context benchmarks with length varying from 1k to 1M, demonstrating XpandA's feasibility for processing ultra-long sequences and its significant effectiveness in enhancing the long-context capabilities of various LLMs by achieving 20\% improvements and 1.5x inference speedup over baselines of full-context, RAG and previous agent-based methods.

Function-guided protein design is a crucial task with significant applications in drug discovery and enzyme engineering. However, the field lacks a unified and comprehensive evaluation framework. Current models are assessed using inconsistent and limited subsets of metrics, which prevents fair comparison and a clear understanding of the relationships between different evaluation criteria. To address this gap, we introduce PDFBench, the first comprehensive benchmark for function-guided denovo protein design. Our benchmark systematically evaluates eight state-of-the-art models on 16 metrics across two key settings: description-guided design, for which we repurpose the Mol-Instructions dataset, originally lacking quantitative benchmarking, and keyword-guided design, for which we introduce a new test set, SwissTest, created with a strict datetime cutoff to ensure data integrity. By benchmarking across a wide array of metrics and analyzing their correlations, PDFBench enables more reliable model comparisons and provides key insights to guide future research.

Spatial Planning is a crucial part in the field of spatial intelligence, which requires the understanding and planning about object arrangements in space perspective. AI agents with the spatial planning ability can better adapt to various real-world applications, including robotic manipulation, automatic assembly, urban planning etc. Recent works have attempted to construct benchmarks for evaluating the spatial intelligence of Multimodal Large Language Models (MLLMs). Nevertheless, these benchmarks primarily focus on spatial reasoning based on typical Visual Question-Answering (VQA) forms, which suffers from the gap between abstract spatial understanding and concrete task execution. In this work, we take a step further to build a comprehensive benchmark called MineAnyBuild, aiming to evaluate the spatial planning ability of open-world AI agents in the Minecraft game. Specifically, MineAnyBuild requires an agent to generate executable architecture building plans based on the given multi-modal human instructions. It involves 4,000 curated spatial planning tasks and also provides a paradigm for infinitely expandable data collection by utilizing rich player-generated content. MineAnyBuild evaluates spatial planning through four core supporting dimensions: spatial understanding, spatial reasoning, creativity, and spatial commonsense. Based on MineAnyBuild, we perform a comprehensive evaluation for existing MLLM-based agents, revealing the severe limitations but enormous potential in their spatial planning abilities. We believe our MineAnyBuild will open new avenues for the evaluation of spatial intelligence and help promote further development for open-world AI agents capable of spatial planning.