The resource requirements of neural networks can be significantly reduced through pruning - the removal of seemingly less important parameters. However, for LLMs, full retraining to recover pruning-induced performance degradation is often prohibitive and classical approaches such as magnitude pruning are suboptimal on Transformers. State-of-the-art methods hence solve a layer-wise mask selection problem: finding a pruning mask that minimizes per-layer pruning error on a small set of calibration data. Exactly solving this problem is computationally infeasible due to its combinatorial nature and the size of the search space, and existing approaches rely on approximations or heuristics. We demonstrate that the mask selection problem can be made drastically more tractable at LLM scale. To that end, we decouple the rows by enforcing equal sparsity levels per row. This allows us to derive optimal 1-swaps (exchanging one kept and one pruned weight) computable efficiently via the Gram matrix. We propose a simple 1-swap algorithm that warmstarts from any pruning mask, runs efficiently on GPUs at LLM scale, and is essentially hyperparameter-free. Our approach reduces per-layer pruning error by up to 60% over Wanda (Sun et al., 2024) and consistently improves perplexity and zero-shot accuracy across state-of-the-art GPT architectures.

LLMs excel at reasoning, but validating their steps remains challenging. Formal verification offers a solution through mechanically checkable proofs. Interactive theorem provers (ITPs) dominate mathematical reasoning but require detailed low-level proof steps, while auto-active verifiers offer automation but focus on software verification. Recent work has begun bridging this divide by evaluating LLMs for software verification in ITPs, but the complementary direction--LLMs for mathematical theorem proving in auto-active verifiers--remains unexplored. We present MINIF2F-DAFNY, the first translation of the widely-used mathematical benchmark miniF2F to an auto-active verifier: Dafny. We find that Dafny's automation alone solves 39-44% of problems with empty proofs, whereas many require substantial proof guidance in ITPs. For remaining problems, we evaluate 7 off-the-shelf LLMs, achieving 55.7% success with the best model (Claude Sonnet 4.5) using modest resources. These results demonstrate effective division of labor: LLMs provide high-level guidance while automation handles low-level details. Our benchmark can be found on GitHub at http://github.com/dafny-lang/miniF2F .

We propose an audio aesthetics score (AES) prediction system by CyberAgent (AESCA) for AudioMOS Challenge 2025 (AMC25) Track 2. The AESCA comprises a Kolmogorov--Arnold Network (KAN)-based audiobox aesthetics and a predictor from the metric scores using the VERSA toolkit. In the KAN-based predictor, we replaced each multi-layer perceptron layer in the baseline model with a group-rational KAN and trained the model with labeled and pseudo-labeled audio samples. The VERSA-based predictor was designed as a regression model using extreme gradient boosting, incorporating outputs from existing metrics. Both the KAN- and VERSA-based models predicted the AES, including the four evaluation axes. The final AES values were calculated using an ensemble model that combined four KAN-based models and a VERSA-based model. Our proposed T12 system yielded the best correlations among the submitted systems, in three axes at the utterance level, two axes at the system level, and the overall average. We also released the inference model of the proposed KAN-based predictor (KAN #1-#4).

Recent image denoising methods have leveraged generative modeling for real noise synthesis to address the costly acquisition of real-world noisy data. However, these generative models typically require camera metadata and extensive target-specific noisy-clean image pairs, often showing limited generalization between settings. In this paper, to mitigate the prerequisites, we propose a Single-Pair Guided Diffusion for generalized noise synthesis GuidNoise, which uses a single noisy/clean pair as the guidance, often easily obtained by itself within a training set. To train GuidNoise, which generates synthetic noisy images from the guidance, we introduce a guidance-aware affine feature modification (GAFM) and a noise-aware refine loss to leverage the inherent potential of diffusion models. This loss function refines the diffusion model's backward process, making the model more adept at generating realistic noise distributions. The GuidNoise synthesizes high-quality noisy images under diverse noise environments without additional metadata during both training and inference. Additionally, GuidNoise enables the efficient generation of noisy-clean image pairs at inference time, making synthetic noise readily applicable for augmenting training data. This self-augmentation significantly improves denoising performance, especially in practical scenarios with lightweight models and limited training data. The code is available at https://github.com/chjinny/GuidNoise.

Real-time video motion transfer applications such as immersive gaming and vision-based anomaly detection require accurate yet diverse future predictions to support realistic synthesis and robust downstream decision making under uncertainty. To improve the diversity of such sequential forecasts we propose a novel inference-time refinement technique that combines Gated Recurrent Unit-Normalizing Flows (GRU-NF) with stochastic sampling methods. While GRU-NF can capture multimodal distributions through its integration of normalizing flows within a temporal forecasting framework, its deterministic transformation structure can limit expressivity. To address this, inspired by Stochastic Normalizing Flows (SNF), we introduce Markov Chain Monte Carlo (MCMC) steps during GRU-NF inference, enabling the model to explore a richer output space and better approximate the true data distribution without retraining. We validate our approach in a keypoint-based video motion transfer pipeline, where capturing temporally coherent and perceptually diverse future trajectories is essential for realistic samples and low bandwidth communication. Experiments show that our inference framework, Gated Recurrent Unit- Stochastic Normalizing Flows (GRU-SNF) outperforms GRU-NF in generating diverse outputs without sacrificing accuracy, even under longer prediction horizons. By injecting stochasticity during inference, our approach captures multimodal behavior more effectively. These results highlight the potential of integrating stochastic dynamics with flow-based sequence models for generative time series forecasting. The code is available at: https://github.com/Tasmiah1408028/Inference-Time-Stochastic-Refinement-Of-GRU-NF-For-Real-Time-Video-Motion-Transfer

Large language models are increasingly embedded into academic writing workflows, yet existing assistants remain external to the editor, preventing deep interaction with document state, structure, and revision history. This separation makes it impossible to support agentic, context-aware operations directly within LaTeX editors such as Overleaf. We present PaperDebugger, an in-editor, multi-agent, and plugin-based academic writing assistant that brings LLM-driven reasoning directly into the writing environment. Enabling such in-editor interaction is technically non-trivial: it requires reliable bidirectional synchronization with the editor, fine-grained version control and patching, secure state management, multi-agent scheduling, and extensible communication with external tools. PaperDebugger addresses these challenges through a Chrome-approved extension, a Kubernetes-native orchestration layer, and a Model Context Protocol (MCP) toolchain that integrates literature search, reference lookup, document scoring, and revision pipelines. Our demo showcases a fully integrated workflow, including localized edits, structured reviews, parallel agent execution, and diff-based updates, encapsulated within a minimal-intrusion user interface (UI). Early aggregated analytics demonstrate active user engagement and validate the practicality of an editor-native, agentic writing assistant. More details about this demo and video could be found at https://github.com/PaperDebugger/PaperDebugger.

Multimodal foundation models are increasingly trained on sensitive data across domains such as finance, biomedicine, and personal identifiers. However, this distributed setup raises serious privacy concerns due to the need for cross-partition data sharing. Split learning addresses these concerns by enabling collaborative model training without raw data exchange between partitions, yet it introduces a significant challenge: transmitting high-dimensional intermediate feature representations between partitions leads to substantial communication costs. To address this challenge, we propose Quantized-TinyLLaVA, a multimodal foundation model with an integrated communication-efficient split learning framework. Our approach adopts a compression module that quantizes intermediate feature into discrete representations before transmission, substantially reducing communication overhead. Besides, we derive a principled quantization strategy grounded in entropy coding theory to determine the optimal number of discrete representation levels. We deploy our framework in a two-partition setting, with one partition operating as the client and the other as the server, to realistically simulate distributed training. Under this setup, Quantized-TinyLLaVA achieves an approximate \textbf{87.5\%} reduction in communication overhead with 2-bit quantization, while maintaining performance of the original 16-bit model across five benchmark datasets. Furthermore, our compressed representations exhibit enhanced resilience against feature inversion attacks, validating the privacy of transmission. The code is available at https://github.com/anonymous-1742/Quantized-TinyLLaVA.

Multimodal Large Language Models are primarily trained and evaluated on aligned image-text pairs, which leaves their ability to detect and resolve real-world inconsistencies largely unexplored. In open-domain applications visual and textual cues often conflict, requiring models to perform structured reasoning beyond surface-level alignment. We introduce CrossCheck-Bench, a diagnostic benchmark for evaluating contradiction detection in multimodal inputs. The benchmark adopts a hierarchical task framework covering three levels of reasoning complexity and defines seven atomic capabilities essential for resolving cross-modal inconsistencies. CrossCheck-Bench includes 15k question-answer pairs sourced from real-world artifacts with synthetically injected contradictions. The dataset is constructed through a multi-stage annotation pipeline involving more than 450 expert hours to ensure semantic validity and calibrated difficulty across perception, integration, and reasoning. We evaluate 13 state-of-the-art vision-language models and observe a consistent performance drop as tasks shift from perceptual matching to logical contradiction detection. Most models perform well on isolated entity recognition but fail when multiple clues must be synthesized for conflict reasoning. Capability-level analysis further reveals uneven skill acquisition, especially in tasks requiring multi-step inference or rule-based validation. Additional probing shows that conventional prompting strategies such as Chain-of-Thought and Set-of-Mark yield only marginal gains. By contrast, methods that interleave symbolic reasoning with grounded visual processing achieve more stable improvements. These results highlight a persistent bottleneck in multimodal reasoning and suggest new directions for building models capable of robust cross-modal verification.

Learning safe and stable robot motions from demonstrations remains a challenge, especially in complex, nonlinear tasks involving dynamic, obstacle-rich environments. In this paper, we propose Safe and Stable Neural Network Dynamical Systems S$^2$-NNDS, a learning-from-demonstration framework that simultaneously learns expressive neural dynamical systems alongside neural Lyapunov stability and barrier safety certificates. Unlike traditional approaches with restrictive polynomial parameterizations, S$^2$-NNDS leverages neural networks to capture complex robot motions, providing probabilistic guarantees through split conformal prediction in learned certificates. Experimental results in various 2D and 3D datasets -- including LASA handwriting and demonstrations recorded kinesthetically from the Franka Emika Panda robot -- validate the effectiveness of S$^2$-NNDS in learning robust, safe, and stable motions from potentially unsafe demonstrations. The source code, supplementary material and experiment videos can be accessed via https://github.com/allemmbinn/S2NNDS

Large language models (LLMs) can internalize private or harmful content, motivating unlearning that removes a forget set while preserving retaining knowledge. However, forgetting updates often cause collateral degradation on retaining knowledge, creating a persistent trade-off. Existing LLM unlearning methods are often heuristic, and other theoretical approaches rely on offline feature constructions that do not capture update-time forget-retain interaction in LLMs. To address this limitation, we aim to develop an LLM unlearning method that reduces the forget-retain trade-off with theoretical guarantees. We take a first-principles view by formalizing "no side effects" as local retain invariance under small parameter updates, and prove an equivalence under optimizer-induced geometry: the retain loss is locally invariant if and only if the update direction is orthogonal to the subspace spanned by retain gradients. Based on the insight, we propose Geometric-disentanglement Unlearning (GU), a lightweight and theoretically grounded projection that can be plug-and-play to existing gradient-based unlearning methods to mitigate forget-retain side effects. Experiments on TOFU, MUSE, and WMDP-cyber show that GU strengthens forgetting while reducing retain drift. When added to SimNPO, it achieves up to 62\% improved forgetting Extraction Strength (ES) and 31\% higher retain ES. We open-sourced our code in https://github.com/Lemutisme/Geometric-Unlearning.

Multimodal large language models (MLLMs) show strong potential as judges. However, existing approaches face a fundamental trade-off: adapting MLLMs to output a single score misaligns with the generative nature of MLLMs and limits fine-grained requirement understanding, whereas autoregressively generating judging analyses is prohibitively slow in high-throughput settings. Observing that judgment reduces to verifying whether inputs satisfy a set of structured requirements, we propose YOFO, a template-conditioned method that judges all requirements in a single forward pass. Built on an autoregressive model, YOFO accepts a structured requirement template and, in one inference step, produces a binary yes/no decision for each requirement by reading the logits of the final token associated with that requirement. This design yields orders-of-magnitude speedups while preserving interpretability. Extensive experiments show that YOFO not only achieves state-of-the-art results on standard recommendation datasets, but also supports dependency-aware analysis -- where subsequent judgments are conditioned on previous ones -- and further benefits from post-hoc CoT.

Geospatial Foundation Models (GeoFMs) are transforming Earth Observation (EO), but evaluation lacks standardized protocols. GEO-Bench-2 addresses this with a comprehensive framework spanning classification, segmentation, regression, object detection, and instance segmentation across 19 permissively-licensed datasets. We introduce ''capability'' groups to rank models on datasets that share common characteristics (e.g., resolution, bands, temporality). This enables users to identify which models excel in each capability and determine which areas need improvement in future work. To support both fair comparison and methodological innovation, we define a prescriptive yet flexible evaluation protocol. This not only ensures consistency in benchmarking but also facilitates research into model adaptation strategies, a key and open challenge in advancing GeoFMs for downstream tasks. Our experiments show that no single model dominates across all tasks, confirming the specificity of the choices made during architecture design and pretraining. While models pretrained on natural images (ConvNext ImageNet, DINO V3) excel on high-resolution tasks, EO-specific models (TerraMind, Prithvi, and Clay) outperform them on multispectral applications such as agriculture and disaster response. These findings demonstrate that optimal model choice depends on task requirements, data modalities, and constraints. This shows that the goal of a single GeoFM model that performs well across all tasks remains open for future research. GEO-Bench-2 enables informed, reproducible GeoFM evaluation tailored to specific use cases. Code, data, and leaderboard for GEO-Bench-2 are publicly released under a permissive license.

Federated Learning (FL) enables collaborative training across decentralized data, but faces key challenges of bidirectional communication overhead and client-side data heterogeneity. To address communication costs while embracing data heterogeneity, we propose pFed1BS, a novel personalized federated learning framework that achieves extreme communication compression through one-bit random sketching. In personalized FL, the goal shifts from training a single global model to creating tailored models for each client. In our framework, clients transmit highly compressed one-bit sketches, and the server aggregates and broadcasts a global one-bit consensus. To enable effective personalization, we introduce a sign-based regularizer that guides local models to align with the global consensus while preserving local data characteristics. To mitigate the computational burden of random sketching, we employ the Fast Hadamard Transform for efficient projection. Theoretical analysis guarantees that our algorithm converges to a stationary neighborhood of the global potential function. Numerical simulations demonstrate that pFed1BS substantially reduces communication costs while achieving competitive performance compared to advanced communication-efficient FL algorithms.

Successful autonomous robot navigation in off-road domains requires the ability to generate high-quality terrain costmaps that are able to both generalize well over a wide variety of terrains and rapidly adapt relative costs at test time to meet mission-specific needs. Existing approaches for costmap generation allow for either rapid test-time adaptation of relative costs (e.g., semantic segmentation methods) or generalization to new terrain types (e.g., representation learning methods), but not both. In this work, we present scaled preference conditioned all-terrain costmap generation (SPACER), a novel approach for generating terrain costmaps that leverages synthetic data during training in order to generalize well to new terrains, and allows for rapid test-time adaptation of relative costs by conditioning on a user-specified scaled preference context. Using large-scale aerial maps, we provide empirical evidence that SPACER outperforms other approaches at generating costmaps for terrain navigation, with the lowest measured regret across varied preferences in five of seven environments for global path planning.

Scientific idea generation is central to discovery, requiring the joint satisfaction of novelty and scientific soundness. Unlike standard reasoning or general creative generation, scientific ideation is inherently open-ended and multi-objective, making its automation particularly challenging. Recent advances in large language models (LLMs) have enabled the generation of coherent and plausible scientific ideas, yet the nature and limits of their creative capabilities remain poorly understood. This survey provides a structured synthesis of methods for LLM-driven scientific ideation, focusing on how different approaches trade off novelty and scientific validity. We organize existing methods into five complementary families: External knowledge augmentation, Prompt-based distributional steering, Inference-time scaling, Multi-agent collaboration, and Parameter-level adaptation. To interpret their contributions, we adopt two complementary creativity frameworks: Boden taxonomy to characterize the expected level of creative novelty, and Rhodes 4Ps framework to analyze the aspects or sources of creativity emphasized by each method. By aligning methodological developments with cognitive creativity frameworks, this survey clarifies the evaluation landscape and identifies key challenges and directions for reliable and systematic LLM-based scientific discovery.

Evaluating large language models (LLMs) typically requires thousands of benchmark items, making the process expensive, slow, and increasingly impractical at scale. Existing evaluation protocols rely on average accuracy over fixed item sets, treating all items as equally informative despite substantial variation in difficulty and discrimination. We introduce ATLAS, an adaptive testing framework based on Item Response Theory (IRT) that estimates model ability using Fisher information-guided item selection. ATLAS reduces the number of required items by up to 90% while maintaining measurement precision. For instance, it matches whole-bank ability estimates using only 41 items (0.157 MAE) on HellaSwag (5,600 items). We further reconstruct accuracy from ATLAS's ability estimates and find that reconstructed accuracies closely match raw accuracies across all five benchmarks, indicating that ability $θ$ preserves the global performance structure. At the same time, $θ$ provides finer discrimination within accuracy-equivalent models: among more than 3,000 evaluated models, 23-31% shift by more than 10 rank positions, and models with identical accuracies receive meaningfully different ability estimates. Code and calibrated item banks are available at https://github.com/Peiyu-Georgia-Li/ATLAS.git.

Recently, the ChannelComp framework has proposed digital over-the-air computation by designing digital modulations that enable the computation of arbitrary functions. Unlike traditional analog over-the-air computation, which is restricted to nomographic functions, ChannelComp enables a broader range of computational tasks while maintaining compatibility with digital communication systems. This framework is intended for applications that favor local information processing over the mere acquisition of data. However, ChannelComp is currently designed for scalar function computation, while numerous data-centric applications necessitate vector-based computations, and it is susceptible to channel fading. In this work, we introduce a generalization of the ChannelComp framework, called VecComp, by integrating ChannelComp with multiple-antenna technology. This generalization not only enables vector function computation but also ensures scalability in the computational complexity, which increases only linearly with the vector dimension. As such, VecComp remains computationally efficient and robust against channel impairments, making it suitable for high-dimensional, data-centric applications. We establish a non-asymptotic upper bound on the mean squared error of VecComp, affirming its computation efficiency under fading channel conditions. Numerical experiments show the effectiveness of VecComp in improving the computation of vector functions and fading compensation over noisy and fading multiple-access channels.

The ability to connect visual patterns with the processes that form them represents one of the deepest forms of visual understanding. Textures of clouds and waves, the growth of cities and forests, or the formation of materials and landscapes are all examples of patterns emerging from underlying mechanisms. We present the SciTextures dataset, a large-scale collection of textures and visual patterns from all domains of science, tech, and art, along with the models and code that generate these images. Covering over 1,270 different models and 100,000 images of patterns and textures from physics, chemistry, biology, sociology, technology, mathematics, and art, this dataset offers a way to explore the deep connection between the visual patterns that shape our world and the mechanisms that produce them. Built through an agentic AI pipeline that autonomously collects, implements, and standardizes scientific and generative models. This AI pipeline is also used to autonomously invent and implement novel methods for generating visual patterns and textures. SciTextures enables systematic evaluation of vision language models (VLM's) ability to link visual patterns to the models and code that generate them, and to identify different patterns that emerge from the same underlying process. We also test VLMs ability to infer and recreate the mechanisms behind visual patterns by providing a natural image of a real-world phenomenon and asking the AI to identify and code a model of the process that formed it, then run this code to generate a simulated image that is compared to the reference image. These benchmarks reveal that VLM's can understand and simulate physical systems beyond visual patterns at multiple levels of abstraction. The dataset and code are available at: https://zenodo.org/records/17485502

We present a simple perturbation mechanism for the release of $d$-dimensional covariance matrices $Σ$ under pure differential privacy. For large datasets with at least $n\geq d^2/\varepsilon$ elements, our mechanism recovers the provably optimal Frobenius norm error guarantees of \cite{nikolov2023private}, while simultaneously achieving best known error for all other $p$-Schatten norms, with $p\in [1,\infty]$. Our error is information-theoretically optimal for all $p\ge 2$, in particular, our mechanism is the first purely private covariance estimator that achieves optimal error in spectral norm. For small datasets $n< d^2/\varepsilon$, we further show that by projecting the output onto the nuclear norm ball of appropriate radius, our algorithm achieves the optimal Frobenius norm error $O(\sqrt{d\;\text{Tr}(Σ) /n})$, improving over the known bounds of $O(\sqrt{d/n})$ of \cite{nikolov2023private} and ${O}\big(d^{3/4}\sqrt{\text{Tr}(Σ)/n}\big)$ of \cite{dong2022differentially}.

We introduce Bounded Numerical Differentiation (BOND), a perturbative method for estimating the gradients of black-box functions. BOND is distinguished by its formulation, which adaptively bounds perturbations to ensure accurate sign estimation, and by its implementation, which operates at black-box interfaces. This enables BOND to be more accurate and scalable compared to existing methods, facilitating end-to-end training of architectures that incorporate non-autodifferentiable modules. We observe that these modules, implemented in our experiments as frozen networks, can enhance model performance without increasing the number of trainable parameters. Our findings highlight the potential of leveraging fixed transformations to expand model capacity, pointing to hybrid analogue - digital devices as a path to scaling networks, and provides insights into the dynamics of adaptive optimizers.

Vision-Language-Action models (VLAs) represent a significant frontier in embodied intelligence, aiming to bridge digital knowledge with physical-world interaction. Despite their remarkable performance, foundational VLAs are hindered by the prohibitive computational and data demands inherent to their large-scale architectures. While a surge of recent research has focused on enhancing VLA efficiency, the field lacks a unified framework to consolidate these disparate advancements. To bridge this gap, this survey presents the first comprehensive review of Efficient Vision-Language-Action models (Efficient VLAs) across the entire model-training-data pipeline. Specifically, we introduce a unified taxonomy to systematically organize the disparate efforts in this domain, categorizing current techniques into three core pillars: (1) Efficient Model Design, focusing on efficient architectures and model compression; (2) Efficient Training, which reduces computational burdens during model learning; and (3) Efficient Data Collection, which addresses the bottlenecks in acquiring and utilizing robotic data. Through a critical review of state-of-the-art methods within this framework, this survey not only establishes a foundational reference for the community but also summarizes representative applications, delineates key challenges, and charts a roadmap for future research. We maintain a continuously updated project page to track our latest developments: https://evla-survey.github.io/.

Recent cooperative perception datasets have played a crucial role in advancing smart mobility applications by enabling information exchange between intelligent agents, helping to overcome challenges such as occlusions and improving overall scene understanding. While some existing real-world datasets incorporate both vehicle-to-vehicle and vehicle-to-infrastructure interactions, they are typically limited to a single intersection or a single vehicle. A comprehensive perception dataset featuring multiple connected vehicles and infrastructure sensors across several intersections remains unavailable, limiting the benchmarking of algorithms in diverse traffic environments. Consequently, overfitting can occur, and models may demonstrate misleadingly high performance due to similar intersection layouts and traffic participant behavior. To address this gap, we introduce UrbanIng-V2X, the first large-scale, multi-modal dataset supporting cooperative perception involving vehicles and infrastructure sensors deployed across three urban intersections in Ingolstadt, Germany. UrbanIng-V2X consists of 34 temporally aligned and spatially calibrated sensor sequences, each lasting 20 seconds. All sequences contain recordings from one of three intersections, involving two vehicles and up to three infrastructure-mounted sensor poles operating in coordinated scenarios. In total, UrbanIng-V2X provides data from 12 vehicle-mounted RGB cameras, 2 vehicle LiDARs, 17 infrastructure thermal cameras, and 12 infrastructure LiDARs. All sequences are annotated at a frequency of 10 Hz with 3D bounding boxes spanning 13 object classes, resulting in approximately 712k annotated instances across the dataset. We provide comprehensive evaluations using state-of-the-art cooperative perception methods and publicly release the codebase, dataset, HD map, and a digital twin of the complete data collection environment.

Multiplicity, the existence of equally good yet competing models, has received growing attention in recent years. While prior work has emphasized modelling choices, the critical role of data in shaping multiplicity has been largely overlooked. In this work, we first introduce a neighbouring datasets framework, arguing that much of data processing can be reframed as choosing between neighbouring datasets. Under this framework, we find a counterintuitive theoretical relationship: neighbouring datasets with greater inter-class distribution overlap exhibit lower multiplicity. Building on this insight, we apply our framework to two domains: active learning and data imputation. For each, we establish natural extensions of the neighbouring datasets perspective, conduct the first systematic study of multiplicity in existing algorithms, and finally, propose novel multiplicity-aware methods, namely, multiplicity-aware data acquisition strategies for active learning and multiplicity-aware data imputation.

As Large Language Models (LLMs) scale to million-token contexts, traditional Mechanistic Interpretability techniques for analyzing attention scale quadratically with context length, demanding terabytes of memory beyond 100,000 tokens. We introduce Sparse Tracing, a novel technique that leverages dynamic sparse attention to efficiently analyze long context attention patterns. We present Stream, a compilable hierarchical pruning algorithm that estimates per-head sparse attention masks in near-linear time $O(T \log T)$ and linear space $O(T)$, enabling one-pass interpretability at scale. Stream performs a binary-search-style refinement to retain only the top-$k$ key blocks per query while preserving the model's next-token behavior. We apply Stream to long chain-of-thought reasoning traces and identify thought anchors while pruning 97-99\% of token interactions. On the RULER benchmark, Stream preserves critical retrieval paths while discarding 90-96\% of interactions and exposes layer-wise routes from the needle to output. Our method offers a practical drop-in tool for analyzing attention patterns and tracing information flow without terabytes of caches. By making long context interpretability feasible on consumer GPUs, Sparse Tracing helps democratize chain-of-thought monitoring. Code is available at https://anonymous.4open.science/r/stream-03B8/.

Human communication is motivated: people speak, write, and create content with a particular communicative intent in mind. As a result, information that large language models (LLMs) and AI agents process is inherently framed by humans' intentions and incentives. People are adept at navigating such nuanced information: we routinely identify benevolent or self-serving motives in order to decide what statements to trust. For LLMs to be effective in the real world, they too must critically evaluate content by factoring in the motivations of the source -- for instance, weighing the credibility of claims made in a sales pitch. In this paper, we undertake a comprehensive study of whether LLMs have this capacity for motivational vigilance. We first employ controlled experiments from cognitive science to verify that LLMs' behavior is consistent with rational models of learning from motivated testimony, and find they successfully discount information from biased sources in a human-like manner. We then extend our evaluation to sponsored online adverts, a more naturalistic reflection of LLM agents' information ecosystems. In these settings, we find that LLMs' inferences do not track the rational models' predictions nearly as closely -- partly due to additional information that distracts them from vigilance-relevant considerations. However, a simple steering intervention that boosts the salience of intentions and incentives substantially increases the correspondence between LLMs and the rational model. These results suggest that LLMs possess a basic sensitivity to the motivations of others, but generalizing to novel real-world settings will require further improvements to these models.

Reformulating nonlinear optimization problems into solver-ready linear optimization problems is often necessary for practical applications, but the process is often manual and requires domain expertise. We propose LinearizeLLM, an agent-based LLM framework that produces solver-ready linear reformulations of nonlinear optimization problems. Agents first detect the nonlinearity pattern (e.g., bilinear products) and apply nonlinearity pattern-aware reformulation techniques, selecting the most suitable linearization technique. We benchmark on 40 instances: 27 derived from ComplexOR by injecting exactly-linearizable operators, and 13 automatically generated instances with deeply nested nonlinearities. LinearizeLLM achieves 73\% mean end-to-end overall success (OSR) across nonlinearity depths (8.3x higher than a one-shot LLM baseline; 4.3x higher than Pyomo). The results suggest that a set of pattern-specialized agents can automate linearization, supporting natural-language-based modeling of nonlinear optimization.

Midtraining, the practice of mixing specialized data with more general pretraining data in an intermediate training phase, has become widespread in language model development, yet there is little understanding of what makes it effective. We propose that midtraining functions as distributional bridging by providing better initialization for posttraining. We conduct controlled pretraining experiments, and find that midtraining benefits are largest for domains distant from general pretraining data, such as code and math, and scale with the proximity advantage the midtraining data provides toward the target distribution. In these domains, midtraining consistently outperforms continued pretraining on specialized data alone both in-domain and in terms of mitigating forgetting. We further conduct an investigation on the starting time and mixture weight of midtraining data, using code as a case study, and find that time of introduction and mixture weight interact strongly such that early introduction of specialized data is amenable to high mixture weights, while late introduction requires lower ones. This suggests that late introduction of specialized data outside a plasticity window cannot be compensated for by increasing data mixtures later in training. Beyond midtraining itself, this suggests that distributional transitions between any training phases may benefit from similar bridging strategies.

Transformer-based models have achieved remarkable performance in NLP tasks. However, their structural characteristics-multiple layers and attention heads-introduce efficiency challenges in inference and deployment. To address these challenges, various pruning methods have recently been proposed. Notably, gradient-based methods using Head Importance Scores (HIS) have gained traction for interpretability, efficiency, and ability to identify redundant heads. However, HIS alone has limitations as it captures only the gradient-driven contribution, overlooking the diversity of attention patterns. To overcome these limitations, we introduce a novel pruning criterion, HIES (Head Importance-Entropy Score), which integrates head importance scores with attention entropy, providing complementary evidence on per-head contribution. Empirically, HIES-based pruning yields up to 15.2% improvement in model quality and 2.04x improvement in stability over HIS-only methods, enabling substantial model compression without sacrificing either accuracy or stability. Code will be released upon publication.

Computational replication of Chinese calligraphy remains challenging. Existing methods falter, either creating high-quality isolated characters while ignoring page-level aesthetics like ligatures and spacing, or attempting page synthesis at the expense of calligraphic correctness. We introduce \textbf{UniCalli}, a unified diffusion framework for column-level recognition and generation. Training both tasks jointly is deliberate: recognition constrains the generator to preserve character structure, while generation provides style and layout priors. This synergy fosters concept-level abstractions that improve both tasks, especially in limited-data regimes. We curated a dataset of over 8,000 digitized pieces, with ~4,000 densely annotated. UniCalli employs asymmetric noising and a rasterized box map for spatial priors, trained on a mix of synthetic, labeled, and unlabeled data. The model achieves state-of-the-art generative quality with superior ligature continuity and layout fidelity, alongside stronger recognition. The framework successfully extends to other ancient scripts, including Oracle bone inscriptions and Egyptian hieroglyphs. Code and data can be viewed in \href{https://github.com/EnVision-Research/UniCalli}{this URL}.

Reinforcement learning (RL) has garnered increasing attention in text-to-image (T2I) generation. However, most existing RL approaches are tailored to either diffusion models or autoregressive models, overlooking an important alternative: masked generative models. In this work, we propose Mask-GRPO, the first method to incorporate Group Relative Policy Optimization (GRPO)-based RL into this overlooked paradigm. Our core insight is to redefine the transition probability, which is different from current approaches, and formulate the unmasking process as a multi-step decision-making problem. To further enhance our method, we explore several useful strategies, including removing the KL constraint, applying the reduction strategy, and filtering out low-quality samples. Using Mask-GRPO, we improve a base model, Show-o, with substantial improvements on standard T2I benchmarks and preference alignment, outperforming existing state-of-the-art approaches. The code is available on https://github.com/xingzhejun/Mask-GRPO