Geometric Problem Solving (GPS) poses a unique challenge for Multimodal Large Language Models (MLLMs), requiring not only the joint interpretation of text and diagrams but also iterative visuospatial reasoning. While existing approaches process diagrams as static images, they lack the capacity for dynamic manipulation--a core aspect of human geometric reasoning involving auxiliary line construction and affine transformations. GeoSketch integrates: (1) a Perception module that abstracts diagrams into structured logic forms, (2) a Symbolic Reasoning module that applies geometric theorems to decide the next deductive step, and (3) a Sketch Action module that executes operations such as drawing auxiliary lines or applying transformations, thereby updating the diagram in a closed loop. To train this agent, we develop a two-stage pipeline: supervised fine-tuning on 2,000 symbolic-curated trajectories followed by reinforcement learning with dense, symbolic rewards to enhance robustness and strategic exploration. To evaluate this paradigm, we introduce the GeoSketch Benchmark, a high-quality set of 390 geometry problems requiring auxiliary construction or affine transformations. Experiments on strong MLLM baselines demonstrate that GeoSketch significantly improves stepwise reasoning accuracy and problem-solving success over static perception methods. Work done during an internship at Hainan University. With the advent of Multimodal Large Language Models (MLLMs) (OpenAI, 2024; Comanici et al., 2025; Hong et al., 2025), Geometric Problem Solving (GPS) presents a unique challenge to MLLMs, demanding not only the joint understanding of text and diagrams but also rigorous, multi-step deductive reasoning (Zhang et al., 2023; Qiao et al., 2024; He et al., 2025). While modern MLLMs can ingest multimodal inputs, their reasoning output remains confined to static text. This limits the use of dynamic and visuospatial strategies in geometric problem solving, which becomes particularly evident in complex tasks requiring multi-stage manipulation.

Graph Retrieval-Augmented Generation (GraphRAG) enhances factual reasoning in LLMs by structurally modeling knowledge through graph-based representations. However, existing GraphRAG approaches face two core limitations: shallow retrieval that fails to surface all critical evidence, and inefficient utilization of pre-constructed structural graph data, which hinders effective reasoning from complex queries. To address these challenges, we propose \textsc{GraphSearch}, a novel agentic deep searching workflow with dual-channel retrieval for GraphRAG. \textsc{GraphSearch} organizes the retrieval process into a modular framework comprising six modules, enabling multi-turn interactions and iterative reasoning. Furthermore, \textsc{GraphSearch} adopts a dual-channel retrieval strategy that issues semantic queries over chunk-based text data and relational queries over structural graph data, enabling comprehensive utilization of both modalities and their complementary strengths. Experimental results across six multi-hop RAG benchmarks demonstrate that \textsc{GraphSearch} consistently improves answer accuracy and generation quality over the traditional strategy, confirming \textsc{GraphSearch} as a promising direction for advancing graph retrieval-augmented generation.

Fine-tuned Large Language Models (LLMs) are vulnerable to backdoor attacks through data poisoning, yet the internal mechanisms governing these attacks remain a black box. Previous research on interpretability for LLM safety tends to focus on alignment, jailbreak, and hallucination, but overlooks backdoor mechanisms, making it difficult to understand and fully eliminate the backdoor threat. In this paper, aiming to bridge this gap, we explore the interpretable mechanisms of LLM backdoors through Backdoor Attribution (BkdAttr), a tripartite causal analysis framework. We first introduce the Backdoor Probe that proves the existence of learnable backdoor features encoded within the representations. Building on this insight, we further develop Backdoor Attention Head Attribution (BAHA), efficiently pinpointing the specific attention heads responsible for processing these features. Our primary experiments reveals these heads are relatively sparse; ablating a minimal \textbf{$\sim$ 3%} of total heads is sufficient to reduce the Attack Success Rate (ASR) by \textbf{over 90%}. More importantly, we further employ these findings to construct the Backdoor Vector derived from these attributed heads as a master controller for the backdoor. Through only \textbf{1-point} intervention on \textbf{single} representation, the vector can either boost ASR up to \textbf{$\sim$ 100% ($\uparrow$)} on clean inputs, or completely neutralize backdoor, suppressing ASR down to \textbf{$\sim$ 0% ($\downarrow$)} on triggered inputs. In conclusion, our work pioneers the exploration of mechanistic interpretability in LLM backdoors, demonstrating a powerful method for backdoor control and revealing actionable insights for the community.

Given the prohibitive cost of pre-training large language models, it is essential to leverage smaller proxy models to optimize datasets before scaling up. However, this approach becomes challenging for reasoning capabilities, which exhibit emergent behavior that only appear reliably at larger model sizes, often exceeding 7B parameters. In response, leveraging smaller models as a proxy for larger model performance has been a key direction through the establishment of empirical scaling laws for prediction (Kaplan et al., 2020; Hoffmann et al., 2022b) or derivation of pre-training dataset rank invariance across scale (Magnusson et al., 2025). 's reasoning trace (Wei et al., 2022b) as the gold label Y Each token i's NLL is weighted by the frontier model's confidence in that token (MinMax normalized). To handle tokenizer mismatches between proxy and frontier models, we compute weights at the letter level and average within tokens. 1 Figure 2: Using MA TH500 as an example benchmark, given the same data source OLMo-Mix-1124 (OLMo et al., 2024), smaller models exhibit more noise and get the direction wrong, making it challenging to use smaller models to proxy larger model performance. Extended visualization across other reasoning benchmarks are available in Appendix A. However, the literature on the emergence of reasoning performance as we scale model size (Wei et al., 2022a; Almazrouei et al., 2023; Du et al., 2024) suggest that there may be a limit to how small theses proxy models can be. They demonstrate that reasoning capabilities only appear when models are sufficiently large in size.

The remarkable zero-shot capabilities of Large Language Models (LLMs) have propelled natural language processing from task-specific models to unified, generalist foundation models. This transformation emerged from simple primitives: large, generative models trained on web-scale data. Curiously, the same primitives apply to today's generative video models. Could video models be on a trajectory towards general-purpose vision understanding, much like LLMs developed general-purpose language understanding? We demonstrate that Veo 3 can solve a broad variety of tasks it wasn't explicitly trained for: segmenting objects, detecting edges, editing images, understanding physical properties, recognizing object affordances, simulating tool use, and more. These abilities to perceive, model, and manipulate the visual world enable early forms of visual reasoning like maze and symmetry solving. Veo's emergent zero-shot capabilities indicate that video models are on a path to becoming unified, generalist vision foundation models.

Current LLM unlearning methods face a critical security vulnerability that undermines their fundamental purpose: while they appear to successfully remove sensitive or harmful knowledge, this ``forgotten" information remains precariously recoverable through relearning attacks. We identify that the root cause is that conventional methods optimizing the forgetting loss at individual data points will drive model parameters toward sharp minima in the loss landscape. In these unstable regions, even minimal parameter perturbations can drastically alter the model's behaviors. Consequently, relearning attacks exploit this vulnerability by using just a few fine-tuning samples to navigate the steep gradients surrounding these unstable regions, thereby rapidly recovering knowledge that was supposedly erased. This exposes a critical robustness gap between apparent unlearning and actual knowledge removal. To address this issue, we propose StableUN, a bi-level feedback-guided optimization framework that explicitly seeks more stable parameter regions via neighborhood-aware optimization. It integrates forgetting feedback, which uses adversarial perturbations to probe parameter neighborhoods, with remembering feedback to preserve model utility, aligning the two objectives through gradient projection. Experiments on WMDP and MUSE benchmarks demonstrate that our method is significantly more robust against both relearning and jailbreaking attacks while maintaining competitive utility performance.

A foundational humanoid motion tracker is expected to be able to track diverse, highly dynamic, and contact-rich motions. More importantly, it needs to operate stably in real-world scenarios against various dynamics disturbances, including terrains, external forces, and physical property changes for general practical use. To achieve this goal, we propose Any2Track (Track Any motions under Any disturbances), a two-stage RL framework to track various motions under multiple disturbances in the real world. Any2Track reformulates dynamics adaptability as an additional capability on top of basic action execution and consists of two key components: AnyTracker and AnyAdapter. AnyTracker is a general motion tracker with a series of careful designs to track various motions within a single policy. AnyAdapter is a history-informed adaptation module that endows the tracker with online dynamics adaptability to overcome the sim2real gap and multiple real-world disturbances. We deploy Any2Track on Unitree G1 hardware and achieve a successful sim2real transfer in a zero-shot manner. Any2Track performs exceptionally well in tracking various motions under multiple real-world disturbances.

It is crucial to efficiently execute instructions such as "Find an apple and a banana" or "Get ready for a field trip," which require searching for multiple objects or understanding context-dependent commands. This study addresses the challenging problem of determining which robot should be assigned to which part of a task when each robot possesses different situational on-site knowledge-specifically, spatial concepts learned from the area designated to it by the user. We propose a task planning framework that leverages large language models (LLMs) and spatial concepts to decompose natural language instructions into subtasks and allocate them to multiple robots. We designed a novel few-shot prompting strategy that enables LLMs to infer required objects from ambiguous commands and decompose them into appropriate subtasks. In our experiments, the proposed method achieved 47/50 successful assignments, outperforming random (28/50) and commonsense-based assignment (26/50). Furthermore, we conducted qualitative evaluations using two actual mobile manipulators. The results demonstrated that our framework could handle instructions, including those involving ad hoc categories such as "Get ready for a field trip," by successfully performing task decomposition, assignment, sequential planning, and execution.

Text-based person search (TBPS) enables the retrieval of person images from large-scale databases using natural language descriptions, offering critical value in surveillance applications. However, a major challenge lies in the labor-intensive process of obtaining high-quality textual annotations, which limits scalability and practical deployment. To address this, we introduce two complementary modules: Multi-Turn Text Generation (MTG) and Multi-Turn Text Interaction (MTI). MTG generates rich pseudo-labels through simulated dialogues with MLLMs, producing fine-grained and diverse visual descriptions without manual supervision. MTI refines user queries at inference time through dynamic, dialogue-based reasoning, enabling the system to interpret and resolve vague, incomplete, or ambiguous descriptions - characteristics often seen in real-world search scenarios. Together, MTG and MTI form a unified and annotation-free framework that significantly improves retrieval accuracy, robustness, and usability. Extensive evaluations demonstrate that our method achieves competitive or superior results while eliminating the need for manual captions, paving the way for scalable and practical deployment of TBPS systems.

We propose MindVL, a multimodal large language model (MLLMs) trained on Ascend NPUs. The training of state-of-the-art MLLMs is often confined to a limited set of hardware platforms and relies heavily on massive, undisclosed data recipes, which hinders reproducibility and open research. To change the common perception that Ascend hardware is unsuitable for efficient full-stage MLLM training, we introduce MindSpeed-MLLM, a highly efficient training framework that supports stable and high-performance training of large-scale Dense and Mixture-of-Experts (MoE) models on Ascend hardware. Based on this, we provide a systematic and open description of the data production methods and mixing strategies for all training stages. Furthermore, we present MindVL, a data-efficient multimodal large language model trained end-to-end on Ascend NPUs. In addition, we find that averaging weights from checkpoints trained with different sequence lengths is particularly effective and yields further gains when combined with test-time resolution search. Our experiments demonstrate superior data efficiency: MindVL-8B matches the performance of Qwen2.5VL-7B using only 10\% of its training data, while our MoE model, MindVL-671B-A37B, matches Qwen2.5VL-72B using only 3\% of the Qwen2.5VL training data, and achieves comparable performance with other leading multimodal MoE models. Our work provides the community with a valuable hardware alternative, open data recipes, and effective performance-enhancing techniques.