Token pruning is essential for enhancing the computational efficiency of vision-language models (VLMs), particularly for video-based tasks where temporal redundancy is prevalent. Prior approaches typically prune tokens either (1) within the vision transformer (ViT) exclusively for unimodal perception tasks such as action recognition and object segmentation, without adapting to downstream vision-language tasks; or (2) only within the LLM while leaving the ViT output intact, often requiring complex text-conditioned token selection mechanisms. In this paper, we introduce Spatio-Temporal Token Scoring (STTS), a simple and lightweight module that prunes vision tokens across both the ViT and the LLM without text conditioning or token merging, and is fully compatible with end-to-end training. By learning how to score temporally via an auxiliary loss and spatially via LLM downstream gradients, aided by our efficient packing algorithm, STTS prunes 50% of vision tokens throughout the entire architecture, resulting in a 62% improvement in efficiency during both training and inference with only a 0.7% drop in average performance across 13 short and long video QA tasks. Efficiency gains increase with more sampled frames per video. Applying test-time scaling for long-video QA further yields performance gains of 0.5-1% compared to the baseline. Overall, STTS represents a novel, simple yet effective technique for unified, architecture-wide vision token pruning.

Multimodal Large Language Models (MLLMs) have made impressive progress in connecting vision and language, but they still struggle with spatial understanding and viewpoint-aware reasoning. Recent efforts aim to augment the input representations with geometric cues rather than explicitly teaching models to reason in 3D space. We introduce Loc3R-VLM, a framework that equips 2D Vision-Language Models with advanced 3D understanding capabilities from monocular video input. Inspired by human spatial cognition, Loc3R-VLM relies on two joint objectives: global layout reconstruction to build a holistic representation of the scene structure, and explicit situation modeling to anchor egocentric perspective. These objectives provide direct spatial supervision that grounds both perception and language in a 3D context. To ensure geometric consistency and metric-scale alignment, we leverage lightweight camera pose priors extracted from a pre-trained 3D foundation model. Loc3R-VLM achieves state-of-the-art performance in language-based localization and outperforms existing 2D- and video-based approaches on situated and general 3D question-answering benchmarks, demonstrating that our spatial supervision framework enables strong 3D understanding. Project page: https://kevinqu7.github.io/loc3r-vlm

In this work, we present EchoGen, a unified framework for layout-to-image generation and image grounding, capable of generating images with accurate layouts and high fidelity to text descriptions (e.g., spatial relationships), while grounding the image robustly at the same time. We believe that image grounding possesses strong text and layout understanding abilities, which can compensate for the corresponding limitations in layout-to-image generation. At the same time, images generated from layouts exhibit high diversity in content, thereby enhancing the robustness of image grounding. Jointly training both tasks within a unified model can promote performance improvements for each. However, we identify that this joint training paradigm encounters several optimization challenges and results in restricted performance. To address these issues, we propose progressive training strategies. First, the Parallel Multi-Task Pre-training (PMTP) stage equips the model with basic abilities for both tasks, leveraging shared tokens to accelerate training. Next, the Dual Joint Optimization (DJO) stage exploits task duality to sequentially integrate the two tasks, enabling unified optimization. Finally, the Cycle RL stage eliminates reliance on visual supervision by using consistency constraints as rewards, significantly enhancing the model's unified capabilities via the GRPO strategy. Extensive experiments demonstrate state-of-the-art results on both layout-to-image generation and image grounding benchmarks, and reveal clear synergistic gains from optimizing the two tasks together.

Building LLM-based agents has become increasingly important. Recent works on LLM-based agent self-evolution primarily record successful experiences as textual prompts or reflections, which cannot reliably guarantee efficient task re-execution in complex scenarios. We propose AgentFactory, a new self-evolution paradigm that preserves successful task solutions as executable subagent code rather than textual experience. Crucially, these subagents are continuously refined based on execution feedback, becoming increasingly robust and efficient as more tasks are encountered. Saved subagents are pure Python code with standardized documentation, enabling portability across any Python-capable system. We demonstrate that AgentFactory enables continuous capability accumulation: its library of executable subagents grows and improves over time, progressively reducing the effort required for similar tasks without manual intervention. Our implementation is open-sourced at https://github.com/zzatpku/AgentFactory, and our demonstration video is available at https://youtu.be/iKSsuAXJHW0.

We present a training-free framework for continuous and controllable image editing at test time for text-conditioned generative models. In contrast to prior approaches that rely on additional training or manual user intervention, we find that a simple steering in the text-embedding space is sufficient to produce smooth edit control. Given a target concept (e.g., enhancing photorealism or changing facial expression), we use a large language model to automatically construct a small set of debiased contrastive prompt pairs, from which we compute a steering vector in the generator's text-encoder space. We then add this vector directly to the input prompt representation to control generation along the desired semantic axis. To obtain a continuous control, we propose an elastic range search procedure that automatically identifies an effective interval of steering magnitudes, avoiding both under-steering (no-edit) and over-steering (changing other attributes). Adding the scaled versions of the same vector within this interval yields smooth and continuous edits. Since our method modifies only textual representations, it naturally generalizes across text-conditioned modalities, including image and video generation. To quantify the steering continuity, we introduce a new evaluation metric that measures the uniformity of semantic change across edit strengths. We compare the continuous editing behavior across methods and find that, despite its simplicity and lightweight design, our approach is comparable to training-based alternatives, outperforming other training-free methods.

Tokenization is a fundamental technique in the generative modeling of various modalities. In particular, it plays a critical role in autoregressive (AR) models, which have recently emerged as a compelling option for 3D generation. However, optimal tokenization of 3D shapes remains an open question. State-of-the-art (SOTA) methods primarily rely on geometric level-of-detail (LoD) hierarchies, originally designed for rendering and compression. These spatial hierarchies are often token-inefficient and lack semantic coherence for AR modeling. We propose Level-of-Semantics Tokenization (LoST), which orders tokens by semantic salience, such that early prefixes decode into complete, plausible shapes that possess principal semantics, while subsequent tokens refine instance-specific geometric and semantic details. To train LoST, we introduce Relational Inter-Distance Alignment (RIDA), a novel 3D semantic alignment loss that aligns the relational structure of the 3D shape latent space with that of the semantic DINO feature space. Experiments show that LoST achieves SOTA reconstruction, surpassing previous LoD-based 3D shape tokenizers by large margins on both geometric and semantic reconstruction metrics. Moreover, LoST achieves efficient, high-quality AR 3D generation and enables downstream tasks like semantic retrieval, while using only 0.1%-10% of the tokens needed by prior AR models.

Synthesizing controllable 6-DOF object manipulation trajectories in 3D environments is essential for enabling robots to interact with complex scenes, yet remains challenging due to the need for accurate spatial reasoning, physical feasibility, and multimodal scene understanding. Existing approaches often rely on 2D or partial 3D representations, limiting their ability to capture full scene geometry and constraining trajectory precision. We present GMT, a multimodal transformer framework that generates realistic and goal-directed object trajectories by jointly leveraging 3D bounding box geometry, point cloud context, semantic object categories, and target end poses. The model represents trajectories as continuous 6-DOF pose sequences and employs a tailored conditioning strategy that fuses geometric, semantic, contextual, and goaloriented information. Extensive experiments on synthetic and real-world benchmarks demonstrate that GMT outperforms state-of-the-art human motion and human-object interaction baselines, such as CHOIS and GIMO, achieving substantial gains in spatial accuracy and orientation control. Our method establishes a new benchmark for learningbased manipulation planning and shows strong generalization to diverse objects and cluttered 3D environments. Project page: https://huajian- zeng.github. io/projects/gmt/.

This paper introduces a novel shape-sensing approach for Concentric Tube Steerable Drilling Robots (CT-SDRs) based on Optical Frequency Domain Reflectometry (OFDR). Unlike traditional FBG-based methods, OFDR enables continuous strain measurement along the entire fiber length with enhanced spatial resolution. In the proposed method, a Shape Sensing Assembly (SSA) is first fabricated by integrating a single OFDR fiber with a flat NiTi wire. The calibrated SSA is then routed through and housed within the internal channel of a flexible drilling instrument, which is guided by the pre-shaped NiTi tube of the CT-SDR. In this configuration, the drilling instrument serves as a protective sheath for the SSA during drilling, eliminating the need for integration or adhesion to the instrument surface that is typical of conventional optical sensor approaches. The performance of the proposed SSA, integrated within the cannulated CT-SDR, was thoroughly evaluated under free-bending conditions and during drilling along multiple J-shaped trajectories in synthetic Sawbones phantoms. Results demonstrate accurate and reliable shape-sensing capability, confirming the feasibility and robustness of this integration strategy.

Controlled video generation has seen drastic improvements in recent years. However, editing actions and dynamic events, or inserting contents that should affect the behaviors of other objects in real-world videos, remains a major challenge. Existing trained models struggle with complex edits, likely due to the difficulty of collecting relevant training data. Similarly, existing training-free methods are inherently restricted to structure- and motion-preserving edits and do not support modification of motion or interactions. Here, we introduce DynaEdit, a training-free editing method that unlocks versatile video editing capabilities with pretrained text-to-video flow models. Our method relies on the recently introduced inversion-free approach, which does not intervene in the model internals, and is thus model-agnostic. We show that naively attempting to adapt this approach to general unconstrained editing results in severe low-frequency misalignment and high-frequency jitter. We explain the sources for these phenomena and introduce novel mechanisms for overcoming them. Through extensive experiments, we show that DynaEdit achieves state-of-the-art results on complex text-based video editing tasks, including modifying actions, inserting objects that interact with the scene, and introducing global effects.

We present AHOY, a method for reconstructing complete, animatable 3D Gaussian avatars from in-the-wild monocular video despite heavy occlusion. Existing methods assume unoccluded input-a fully visible subject, often in a canonical pose-excluding the vast majority of real-world footage where people are routinely occluded by furniture, objects, or other people. Reconstructing from such footage poses fundamental challenges: large body regions may never be observed, and multi-view supervision per pose is unavailable. We address these challenges with four contributions: (i) a hallucination-as-supervision pipeline that uses identity-finetuned diffusion models to generate dense supervision for previously unobserved body regions; (ii) a two-stage canonical-to-pose-dependent architecture that bootstraps from sparse observations to full pose-dependent Gaussian maps; (iii) a map-pose/LBS-pose decoupling that absorbs multi-view inconsistencies from the generated data; (iv) a head/body split supervision strategy that preserves facial identity. We evaluate on YouTube videos and on multi-view capture data with significant occlusion and demonstrate state-of-the-art reconstruction quality. We also demonstrate that the resulting avatars are robust enough to be animated with novel poses and composited into 3DGS scenes captured using cell-phone video. Our project page is available at https://miraymen.github.io/ahoy/

Orthogonalized-momentum optimizers such as Muon improve transformer training by approximately whitening/orthogonalizing matrix-valued momentum updates via a short polar-decomposition iteration. However, polar-factor approximations typically require multiple large matrix multiplications, and the resulting overhead can be substantial and hardware-dependent. We introduce MUD (MomentUm Decorrelation), a complementary whitening approach that replaces Muon's polar update with a triangular (Cholesky-like) whitening surrogate inspired by classical Gram--Schmidt and Gauss-Seidel ideas. We show that row-orthonormal matrices are fixed points of the MUD map, relate the inner step to symmetric Gauss-Seidel preconditioning of the Gram matrix, and prove quadratic local convergence near the fixed point. In terms of time-to-perplexity, MUD yields consistent 10-50\% wall-clock improvements over tuned AdamW and Muon in time-to-perplexity, typically converging slightly slower per step than Muon but with substantially lower optimizer overhead -- relative to Muon, MUD improves peak tokens/s by roughly $1.3-2.6\times$ across most settings and up to nearly $3\times$ on GPT-2 large on an A100. We also demonstrate training a ESM-2 150M protein language model, where MUD matches Muon-level validation perplexity in significantly less wall-clock time.

Robotics foundation models have demonstrated strong capabilities in executing natural language instructions across diverse tasks and environments. However, they remain largely data-driven and lack formal guarantees on safety and satisfaction of time-dependent specifications during deployment. In practice, robots often need to comply with operational constraints involving rich spatio-temporal requirements such as time-bounded goal visits, sequential objectives, and persistent safety conditions. In this work, we propose a specification-aware action distribution optimization framework that enforces a broad class of Signal Temporal Logic (STL) constraints during execution of a pretrained robotics foundation model without modifying its parameters. At each decision step, the method computes a minimally modified action distribution that satisfies a hard STL feasibility constraint by reasoning over the remaining horizon using forward dynamics propagation. We validate the proposed framework in simulation using a state-of-the-art robotics foundation model across multiple environments and complex specifications.

Harmonization methods such as ComBat and its variants are widely used to mitigate diffusion MRI (dMRI) site-specific biases. However, ComBat assumes that subject distributions exhibit a Gaussian profile. In practice, patients with neurological disorders often present diffusion metrics that deviate markedly from those of healthy controls, introducing pathological outliers that distort site-effect estimation. This problem is particularly challenging in clinical practice as most patients undergoing brain imaging have an underlying and yet undiagnosed condition, making it difficult to exclude them from harmonization cohorts, as their scans were precisely prescribed to establish a diagnosis. In this paper, we show that harmonizing data to a normative reference population with ComBat while including pathological cases induces significant distortions. Across 7 neurological conditions, we evaluated 10 outlier rejection methods with 4 ComBat variants over a wide range of scenarios, revealing that many filtering strategies fail in the presence of pathology. In contrast, a simple MLP provides robust outlier compensation enabling reliable harmonization while preserving disease-related signal. Experiments on both control and real multi-site cohorts, comprising up to 80% of subjects with neurological disorders, demonstrate that Robust-ComBat consistently outperforms conventional statistical baselines with lower harmonization error across all ComBat variants.

Media design layer generation enables the creation of fully editable, layered design documents such as posters, flyers, and logos using only natural language prompts. Existing methods either restrict outputs to a fixed number of layers or require each layer to contain only spatially continuous regions, causing the layer count to scale linearly with design complexity. We propose LaDe (Layered Media Design), a latent diffusion framework that generates a flexible number of semantically meaningful layers. LaDe combines three components: an LLM-based prompt expander that transforms a short user intent into structured per-layer descriptions that guide the generation, a Latent Diffusion Transformer with a 4D RoPE positional encoding mechanism that jointly generates the full media design and its constituent RGBA layers, and an RGBA VAE that decodes each layer with full alpha-channel support. By conditioning on layer samples during training, our unified framework supports three tasks: text-to-image generation, text-to-layers media design generation, and media design decomposition. We compare LaDe to Qwen-Image-Layered on text-to-layers and image-to-layers tasks on the Crello test set. LaDe outperforms Qwen-Image-Layered in text-to-layers generation by improving text-to-layer alignment, as validated by two VLM-as-a-judge evaluators (GPT-4o mini and Qwen3-VL).

Handling gender across languages remains a persistent challenge for Machine Translation (MT) and Large Language Models (LLMs), especially when translating from gender-neutral languages into morphologically gendered ones, such as English to Italian. English largely omits grammatical gender, while Italian requires explicit agreement across multiple grammatical categories. This asymmetry often leads MT systems to default to masculine forms, reinforcing bias and reducing translation accuracy. To address this issue, we present the Contextual Gender Annotation (ConGA) framework, a linguistically grounded set of guidelines for word-level gender annotation. The scheme distinguishes between semantic gender in English through three tags, Masculine (M), Feminine (F), and Ambiguous (A), and grammatical gender realisation in Italian (Masculine (M), Feminine (F)), combined with entity-level identifiers for cross-sentence tracking. We apply ConGA to the gENder-IT dataset, creating a gold-standard resource for evaluating gender bias in translation. Our results reveal systematic masculine overuse and inconsistent feminine realisation, highlighting persistent limitations of current MT systems. By combining fine-grained linguistic annotation with quantitative evaluation, this work offers both a methodology and a benchmark for building more gender-aware and multilingual NLP systems.

While Large Language Models achieve state-of-the-art results across a wide range of NLP tasks, they remain prone to systematic biases. Among these, gender bias is particularly salient in MT, due to systematic differences across languages in whether and how gender is marked. As a result, translation often requires disambiguating implicit source signals into explicit gender-marked forms. In this context, standard benchmarks may capture broad disparities but fail to reflect the full complexity of gender bias in modern MT. In this paper, we extend recent frameworks on bias evaluation by: (i) introducing a novel measure coined "Prior Bias", capturing a model's default gender assumptions, and (ii) applying the framework to decoder-only MT models. Our results show that, despite their scale and state-of-the-art status, decoder-only models do not generally outperform encoder-decoder architectures on gender-specific metrics; however, post-training (e.g., instruction tuning) not only improves contextual awareness but also reduces the masculine Prior Bias.

Extending language models to video introduces two challenges: representation, where existing methods rely on lossy approximations, and long-context, where caption- or agent-based pipelines collapse video into text and lose visual fidelity. To overcome this, we introduce \textbf{VideoAtlas}, a task-agnostic environment to represent video as a hierarchical grid that is simultaneously lossless, navigable, scalable, caption- and preprocessing-free. An overview of the video is available at a glance, and any region can be recursively zoomed into, with the same visual representation used uniformly for the video, intermediate investigations, and the agent's memory, eliminating lossy text conversion end-to-end. This hierarchical structure ensures access depth grows only logarithmically with video length. For long-context, Recursive Language Models (RLMs) recently offered a powerful solution for long text, but extending them to visual domain requires a structured environment to recurse into, which \textbf{VideoAtlas} provides. \textbf{VideoAtlas} as a Markov Decision Process unlocks Video-RLM: a parallel Master-Worker architecture where a Master coordinates global exploration while Workers concurrently drill into assigned regions to accumulate lossless visual evidence. We demonstrate three key findings: (1)~logarithmic compute growth with video duration, further amplified by a 30-60\% multimodal cache hit rate arising from the grid's structural reuse. (2)~environment budgeting, where bounding the maximum exploration depth provides a principled compute-accuracy hyperparameter. (3)~emergent adaptive compute allocation that scales with question granularity. When scaling from 1-hour to 10-hour benchmarks, Video-RLM remains the most duration-robust method with minimal accuracy degradation, demonstrating that structured environment navigation is a viable and scalable paradigm for video understanding.

Rapid adaptation in complex control systems remains a central challenge in reinforcement learning. We introduce a framework in which policy and value functions share a low-dimensional coefficient vector - a goal embedding - that captures task identity and enables immediate adaptation to novel tasks without retraining representations. During pretraining, we jointly learn structured value bases and compatible policy bases through a bilinear actor-critic decomposition. The critic factorizes as Q = sum_k G_k(g) y_k(s,a), where G_k(g) is a goal-conditioned coefficient vector and y_k(s,a) are learned value basis functions. This multiplicative gating - where a context signal scales a set of state-dependent bases - is reminiscent of gain modulation observed in Layer 5 pyramidal neurons, where top-down inputs modulate the gain of sensory-driven responses without altering their tuning. Building on Successor Features, we extend the decomposition to the actor, which composes a set of primitive policies weighted by the same coefficients G_k(g). At test time the bases are frozen and G_k(g) is estimated zero-shot via a single forward pass, enabling immediate adaptation to novel tasks without any gradient update. We train a Soft Actor-Critic agent on the MuJoCo Ant environment under a multi-directional locomotion objective, requiring the agent to walk in eight directions specified as continuous goal vectors. The bilinear structure allows each policy head to specialize to a subset of directions, while the shared coefficient layer generalizes across them, accommodating novel directions by interpolating in goal embedding space. Our results suggest that shared low-dimensional goal embeddings offer a general mechanism for rapid, structured adaptation in high-dimensional control, and highlight a potentially biologically plausible principle for efficient transfer in complex reinforcement learning systems.

Converting pretrained attention modules such as grouped-query attention (GQA) into multi-head latent attention (MLA) can improve expressivity without increasing KV-cache cost, making it attractive for efficient inference. However, many practical conversion baselines rely on weight-only low-rank approximations (e.g., SVD-style initializations) and uniform rank allocation. They focus on minimizing the difference between weight matrices rather than on how those weights affect input activations, ignore the covariance structure of activations, and enforce uniform rank across layers, causing activation drift and degraded attention fidelity. To address these issues, we propose CARE, a Covariance-Aware, Rank-Enhanced MLA conversion pipeline under a fixed KV width. CARE introduces three key steps: (i) activation-preserving factorization, which aligns the approximation with the actual input activations rather than just the weights; (ii) adjusted-rank allocation, which spreads a fixed KV budget across layers by giving more capacity to layers that need it most; and (iii) KV-parity mapping, which reparameterizes the converted K and V to fit the MLA format while keeping the KV-cache size unchanged. Our method outperforms a uniform-rank SVD baseline on Qwen3-4B/30B-A3B-Instruct-2507 and Llama-3.1-8B/70B-Instruct, reducing one-shot perplexity by up to 215x and improving mean accuracy by up to 1.70x at matched KV budgets. With a brief post-SVD healing fine-tune, we fully recover the original model's accuracy.

The widespread adoption of dashcams has made video evidence in traffic accidents increasingly abundant, yet transforming "what happened in the video" into "who is responsible under which legal provisions" still relies heavily on human experts. Existing ego-view traffic accident studies mainly focus on perception and semantic understanding, while LLM-based legal methods are mostly built on textual case descriptions and rarely incorporate video evidence, leaving a clear gap between the two. We first propose C-TRAIL, a multimodal legal dataset that, under the Chinese traffic regulation system, explicitly aligns dashcam videos and textual descriptions with a closed set of responsibility modes and their corresponding Chinese traffic statutes. On this basis, we introduce a two-stage framework: (1) a traffic accident understanding module that generates textual video descriptions; and (2) a legal multi-agent framework that outputs responsibility modes, statute sets, and complete judgment reports. Experimental results on C-TRAIL and MM-AU show that our method outperforms general and legal LLMs, as well as existing agent-based approaches, while providing a transparent and interpretable legal reasoning process.

Legal age estimation plays a critical role in forensic and medico-legal contexts, where decisions must be supported by accurate, robust, and reproducible methods with explicit uncertainty quantification. While prior artificial intelligence (AI)-based approaches have primarily focused on hand radiographs or dental imaging, clavicle computed tomography (CT) scans remain underexplored despite their documented effectiveness for legal age estimation. In this work, we present an interpretable, multi-stage pipeline for legal age estimation from clavicle CT scans. The proposed framework combines (i) a feature-based connected-component method for automatic clavicle detection that requires minimal manual annotation, (ii) an Integrated Gradients-guided slice selection strategy used to construct the input data for a multi-slice convolutional neural network that estimates legal age, and (iii) conformal prediction intervals to support uncertainty-aware decisions in accordance with established international protocols. The pipeline is evaluated on 1,158 full-body post-mortem CT scans from a public forensic dataset (the New Mexico Decedent Image Database). The final model achieves state-of-the-art performance with a mean absolute error (MAE) of 1.55 $\pm$ 0.16 years on a held-out test set, outperforming both human experts (MAE of approximately 1.90 years) and previous methods (MAEs above 1.75 years in our same dataset). Furthermore, conformal prediction enables configurable coverage levels aligned with forensic requirements. Attribution maps indicate that the model focuses on anatomically relevant regions of the medial clavicular epiphysis. The proposed method, which is currently being added as part of the Skeleton-ID software (https://skeleton-id.com/skeleton-id/), is intended as a decision-support component within multi-factorial forensic workflows.

Semantic segmentation for uncrewed aerial vehicles (UAVs) is fundamental for aerial scene understanding, yet existing RGB and RGB-T datasets remain limited in scale, diversity, and annotation efficiency due to the high cost of manual labeling and the difficulties of accurate RGB-T alignment on off-the-shelf UAVs. To address these challenges, we propose a scalable geometry-driven 2D-3D-2D paradigm that leverages multi-view redundancy in high-overlap aerial imagery to automatically propagate labels from a small subset of manually annotated RGB images to both RGB and thermal modalities within a unified framework. By lifting less than 3% of RGB images into a semantic 3D point cloud and reprojecting it into all views, our approach enables dense pseudo ground-truth generation across large image collections, automatically producing 97% of RGB labels and 100% of thermal labels while achieving 91% and 88% annotation accuracy without any 2D manual refinement. We further extend this 2D-3D-2D paradigm to cross-modal image registration, using 3D geometry as an intermediate alignment space to obtain fully automatic, strong pixel-level RGB-T alignment with 87% registration accuracy and no hardware-level synchronization. Applying our framework to existing geo-referenced aerial imagery, we construct SegFly, a large-scale benchmark with over 20,000 high-resolution RGB images and more than 15,000 geometrically aligned RGB-T pairs spanning diverse urban, industrial, and rural environments across multiple altitudes and seasons. On SegFly, we establish the Firefly baseline for RGB and thermal semantic segmentation and show that both conventional architectures and vision foundation models benefit substantially from SegFly supervision, highlighting the potential of geometry-driven 2D-3D-2D pipelines for scalable multi-modal scene understanding. Data and Code available at https://github.com/markus-42/SegFly.

Large language models (LLMs) contain billions of parameters, yet many exact values are not essential. We show that what matters most is the relative rank of weights-whether one connection is stronger or weaker than another-rather than precise magnitudes. To reduce the number of unique weight values, we apply weight clustering to pretrained models, replacing every weight matrix with K shared values from K-means. For Llama 3.1-8B-Instruct and SmolLM2-135M, reducing each matrix to only 16-64 distinct values preserves strong accuracy without retraining, providing a simple, training-free method to compress LLMs on disk. Optionally fine-tuning only the cluster means (centroids) recovers 30-40 percent of the remaining accuracy gap at minimal cost. We then systematically randomize cluster means while keeping assignments fixed. Scrambling the relative ranks of the clusters degrades quality sharply-perplexity can increase by orders of magnitude-even when global statistics such as mean and variance are preserved. In contrast, rank-preserving randomizations cause almost no loss at mid and late layers. On the other hand, when many layers are perturbed simultaneously, progressive layer-by-layer replacement reveals that scale drift-not rank distortion-is the dominant collapse mechanism; however, an affine correction w' = aw + b with a > 0 (which preserves both rank order and overall weight distribution) can substantially delay this drift. This rank-based perspective offers a new lens on model compression and robustness.

Collaborative inference of object classification Deep neural Networks (DNNs) where resource-constrained end-devices offload partially processed data to remote edge servers to complete end-to-end processing, is becoming a key enabler of edge-AI. However, such edge-offloading is vulnerable to malicious data injections leading to stealthy misclassifications that are tricky to detect, especially in the presence of environmental noise. In this paper, we propose a semi-gray-box and noise- aware anomaly detection framework fueled by a variational autoencoder (VAE) to capture deviations caused by adversarial manipulation. The proposed framework incorporates a robust noise-aware feature that captures the characteristic behavior of environmental noise to improve detection accuracy while reducing false alarm rates. Our evaluation with popular object classification DNNs demonstrate the robustness of the proposed detection (up to 90% AUROC across DNN configurations) under realistic noisy conditions while revealing limitations caused by feature similarity and elevated noise levels.

Understanding the distance between human languages is central to linguistics, anthropology, and tracing human evolutionary history. Yet, while linguistics has long provided rich qualitative accounts of cross-linguistic variation, a unified and scalable quantitative approach to measuring language distance remains lacking. In this paper, we introduce a method that leverages pretrained multilingual language models as systematic instruments for linguistic measurement. Specifically, we show that the spontaneously emerged attention mechanisms of these models provide a robust, tokenization-agnostic measure of cross-linguistic distance, termed Attention Transport Distance (ATD). By treating attention matrices as probability distributions and measuring their geometric divergence via optimal transport, we quantify the representational distance between languages during translation. Applying ATD to a large and diverse set of languages, we demonstrate that the resulting distances recover established linguistic groupings with high fidelity and reveal patterns aligned with geographic and contact-induced relationships. Furthermore, incorporating ATD as a regularizer improves transfer performance in low-resource machine translation. Our results establish a principled foundation for testing linguistic hypotheses using artificial neural networks. This framework transforms multilingual models into powerful tools for quantitative linguistic discovery, facilitating more equitable multilingual AI.

We introduce SpiderCam, an FPGA-based snapshot depth-from-defocus camera which produces 480x400 sparse depth maps in real-time at 32.5 FPS over a working range of 52 cm while consuming 624 mW of power in total. SpiderCam comprises a custom camera that simultaneously captures two differently focused images of the same scene, processed with a SystemVerilog implementation of depth from differential defocus (DfDD) on a low-power FPGA. To achieve state-of-the-art power consumption, we present algorithmic improvements to DfDD that overcome challenges caused by low-power sensors, and design a memory-local implementation for streaming depth computation on a device that is too small to store even a single image pair. We report the first sub-Watt total power measurement for passive FPGA-based 3D cameras in the literature.

Text-to-image (T2I) models have substantially improved image fidelity and prompt adherence, yet their creativity remains constrained by reliance on discrete natural language prompts. When presented with fuzzy prompts such as ``a creative vinyl record-inspired skyscraper'', these models often fail to infer the underlying creative intent, leaving creative ideation and prompt design largely to human users. Recent reasoning- or agent-driven approaches iteratively augment prompts but incur high computational and monetary costs, as their instance-specific generation makes ``creativity'' costly and non-reusable, requiring repeated queries or reasoning for subsequent generations. To address this, we introduce \textbf{CAT}, a framework for \textbf{C}reative \textbf{A}gent \textbf{T}okenization that encapsulates agents' intrinsic understanding of ``creativity'' through a \textit{Creative Tokenizer}. Given the embeddings of fuzzy prompts, the tokenizer generates a reusable token template that can be directly concatenated with them to inject creative semantics into T2I models without repeated reasoning or prompt augmentation. To enable this, the tokenizer is trained via creative semantic disentanglement, leveraging relations among partially overlapping concept pairs to capture the agent's latent creative representations. Extensive experiments on \textbf{\textit{Architecture Design}}, \textbf{\textit{Furniture Design}}, and \textbf{\textit{Nature Mixture}} tasks demonstrate that CAT provides a scalable and effective paradigm for enhancing creativity in T2I generation, achieving a $3.7\times$ speedup and a $4.8\times$ reduction in computational cost, while producing images with superior human preference and text-image alignment compared to state-of-the-art T2I models and creative generation methods.

Post training quantization is essential for deploying large language models (LLMs) on resource constrained hardware, yet state of the art methods enforce uniform bit widths across layers, yielding suboptimal accuracy efficiency trade offs. We present RAMP (Reinforcement Adaptive Mixed Precision), an off policy Soft Actor Critic framework that learns per layer bit width assignments to minimize perplexity under a global bit budget. The policy conditions on an 11 dimensional embedding of activation statistics, weight properties, and structural descriptors, enabling zero shot transfer across model families and scales. To enable stable sub 4 bit quantization, we introduce Scale Folding, a preconditioning technique that migrates activation outliers into weights via per channel scaling and normalization layer compensation. A quality prioritized reward with asymmetric penalties and budget cliffs drives rapid convergence. On Llama 2 7B, RAMP achieves 5.54 perplexity at 3.68GB (3.65 effective bits), outperforming uniform 4 bit AWQ (5.60 at 3.90 GB) and GPTQ by 6% in size and 1% to3% in quality. Critically, a policy trained only on Llama 2 7B generalizes zero shot to Llama 2 13B and Mistral 7B, often surpassing target specific training, supporting the hypothesis that quantization sensitivity is primarily architectural. The HALO pipeline exports allocations to GGUF format for kernel free inference on CPUs, GPUs, and edge devices, retaining 99.5% of FP16 commonsense reasoning performance.

Large language models (LLMs) are trained through multi-stage pipelines over heterogeneous data sources, yet developers lack a principled way to pinpoint the specific data responsible for an observed behavior. This lack of observability reduces debugging to reactive patching and makes failures prone to recur under distribution shift or subsequent model updates. To address this limitation, we propose DebugLM, a framework that equips LLMs with built-in data provenance, enabling them to explicitly trace the origins of their behaviors to specific training data sources. Specifically, the model learns to associate its responses with unique provenance tags that indicate the responsible dataset, empowering developers to precisely identify where undesirable behaviors are learned. Building on this capability, DebugLM further supports targeted test-time remediation, enabling developers to selectively trigger targeted refusal for specified data sources without retraining or modifying model parameters. Experiments demonstrate that DebugLM provides accurate behavior tracing in multi-stage training pipelines and effective test-time remediation while preserving the general utility of the model.

Instruction-following image editing models are expected to modify only the specified region while keeping the rest of the image unchanged. However, in practice, we observe a pervasive phenomenon -- edit spillover: models alter semantically related but unspecified content outside the edit region. This raises a fundamental question -- does spillover reflect genuine implicit world understanding, or is it merely attention leakage? We propose EditSpilloverProbe, a systematic framework that repurposes edit spillover as a natural probe for world knowledge in image editing models. We introduce a spillover taxonomy (spatial, semantic, mixed, random), an automated detection-and-classification pipeline, and a benchmark dataset constructed from real-world Chinese text editing tasks, EditSpilloverBench. Systematic evaluation of 5 representative editing models reveals three core findings: (1) spillover rates vary dramatically across architectures, from 3.49% to 11.46%, with a 3.3x ratio; (2) absolute semantic spillover quantity reveals models' world understanding capability -- nano_banana produces the most semantic spillover (27.8 per image), while qwen_2511 has the most precise editing control but lower semantic spillover (16.3 per image), revealing a trade-off between editing control and world understanding; (3) spatial decay analysis shows spillover area density decays exponentially with distance, but the proportion of semantically relevant spillover remains constant (40%-58%), providing direct evidence that semantic spillover reflects genuine world understanding rather than spatial diffusion.