We thank all reviewers for their valuable comments. The reviewers are generally satisfied with our writing and experiments. Other concerns are minor - we carefully respond below and will revise the paper accordingly. Can one-bit supervision reduce annotation costs? According to the authors of ILSVRC2012 [Russakovsky et al., IJCV'15], the average time for a full-bit annotation is A1: Please refer to the common question.

We study the following distribution clustering problem: Given a hidden partition of $k$ distributions into two groups, such that the distributions within each group are the same, and the two distributions associated with the two clusters are $\varepsilon$-far in total variation, the goal is to recover the partition. We establish upper and lower bounds on the sample complexity for two fundamental cases: (1) when one of the cluster's distributions is known, and (2) when both are unknown. Our upper and lower bounds characterize the sample complexity's dependence on the domain size $n$, number of distributions $k$, size $r$ of one of the clusters, and distance $\varepsilon$. In particular, we achieve tightness with respect to $(n,k,r,\varepsilon)$ (up to an $O(\log k)$ factor) for all regimes.

We introduce a two-stage self-supervised framework that combines the Joint-Embedding Predictive Architecture (JEPA) with a Density Adaptive Attention Mechanism (DAAM) for learning robust speech representations. Stage 1 uses JEPA with DAAM to learn semantic audio features via masked prediction in latent space, fully decoupled from waveform reconstruction. Stage 2 leverages these representations for efficient tokenization using Finite Scalar Quantization (FSQ) and a mixed-radix packing scheme, followed by high-fidelity waveform reconstruction with a HiFi-GAN decoder. By integrating Gaussian mixture-based density-adaptive gating into the JEPA encoder, the model performs adaptive temporal feature selection and discovers hierarchical speech structure at a low frame rate of 2.5 Hz. The resulting tokens (47.5 tokens/sec) provide a reversible, highly compressed, and language-model-friendly representation that is competitive with, and often more efficient than, existing neural audio codecs.

"Thinking with images" has emerged as an effective paradigm for advancing visual reasoning, extending beyond text-only chains of thought by injecting visual evidence into intermediate reasoning steps. However, existing methods fall short of human-like abstract visual thinking, as their flexibility is fundamentally limited by external tools. In this work, we introduce Monet, a training framework that enables multimodal large language models (MLLMs) to reason directly within the latent visual space by generating continuous embeddings that function as intermediate visual thoughts. We identify two core challenges in training MLLMs for latent visual reasoning: high computational cost in latent-vision alignment and insufficient supervision over latent embeddings, and address them with a three-stage distillation-based supervised fine-tuning (SFT) pipeline. We further reveal a limitation of applying GRPO to latent reasoning: it primarily enhances text-based reasoning rather than latent reasoning. To overcome this, we propose VLPO (Visual-latent Policy Optimization), a reinforcement learning method that explicitly incorporates latent embeddings into policy gradient updates. To support SFT, we construct Monet-SFT-125K, a high-quality text-image interleaved CoT dataset containing 125K real-world, chart, OCR, and geometry CoTs. Our model, Monet-7B, shows consistent gains across real-world perception and reasoning benchmarks and exhibits strong out-of-distribution generalization on challenging abstract visual reasoning tasks. We also empirically analyze the role of each training component and discuss our early unsuccessful attempts, providing insights for future developments in visual latent reasoning. Our model, data, and code are available at https://github.com/NOVAglow646/Monet.

Educational illustrations play a central role in communicating abstract concepts, yet current multimodal large language models (MLLMs) remain limited in producing pedagogically coherent and semantically consistent educational visuals. We introduce MAGMA-Edu, a self-reflective multi-agent framework that unifies textual reasoning and diagrammatic synthesis for structured educational problem generation. Unlike existing methods that treat text and image generation independently, MAGMA-Edu employs a two-stage co-evolutionary pipeline: (1) a generation-verification-reflection loop that iteratively refines question statements and solutions for mathematical accuracy, and (2) a code-based intermediate representation that enforces geometric fidelity and semantic alignment during image rendering. Both stages are guided by internal self-reflection modules that evaluate and revise outputs until domain-specific pedagogical constraints are met. Extensive experiments on multimodal educational benchmarks demonstrate the superiority of MAGMA-Edu over state-of-the-art MLLMs. Compared to GPT-4o, MAGMA-Edu improves the average textual metric from 57.01 to 92.31 (+35.3 pp) and boosts image-text consistency (ITC) from 13.20 to 85.24 (+72 pp). Across all model backbones, MAGMA-Edu achieves the highest scores (Avg-Text 96.20, ITC 99.12), establishing a new state of the art for multimodal educational content generation and demonstrating the effectiveness of self-reflective multi-agent collaboration in pedagogically aligned vision-language reasoning.

Vision-Language Model (VLM) is an important component to enable robust robot manipulation. Y et, using it to translate human instructions into an action-resolvable intermediate representation often needs a tradeoff between VLM-comprehensibility and generalizability. Inspired by context-free grammar, we design the Semantic Assembly representation named SEAM, by decomposing the intermediate representation into vocabulary and grammar . Doing so leads us to a concise vocabulary of semantically-rich operations and a VLM-friendly grammar for handling diverse unseen tasks. In addition, we design a new open-vocabulary segmentation paradigm with a retrieval-augmented few-shot learning strategy to localize fine-grained object parts for manipulation, effectively with the shortest inference time over all state-of-the-art parallel works. Also, we formulate new metrics for action-generalizability and VLM-comprehensibility, demonstrating the compelling performance of SEAM over mainstream representations on both aspects.

The Segment Anything Model 3 (SAM3) advances visual understanding with Promptable Concept Segmentation (PCS) across images and videos, but its unified architecture (shared vision backbone, DETR-style detector, dense-memory tracker) remains prohibitive for on-device use. W e present EfficientSAM3, a family of efficient models built on Progressive Hierarchical Distillation (PHD) that transfers capability from SAM3 to lightweight students in three stages: (1) Encoder Distillation aligns image features via prompt-in-the-loop training on SA-1B; (2) Temporal Memory Distillation replaces dense memory with a compact Perceiver-based module trained on SA-V to compress and retrieve spatiotemporal features efficiently; and (3) End-to-End Fine-Tuning refines the full pipeline on the official SAM3 PCS data to preserve concept-level performance. PHD yields a spectrum of student variants using RepViT, TinyViT, and EfficientViT backbones, enabling on-device concept segmentation and tracking while maintaining high fidelity to teacher behavior . W e benchmark on popular VOS datasets, and compare with varies of releated work, achieing strong performance-efficiency trade-offs.

In fact, Krause et al. (2008) showed that when the number of objectives