The ability to accurately interpret implied meanings plays a crucial role in human communication and language use, and language models are also expected to possess this capability. This study demonstrates that providing language models with pragmatic theories as prompts is an effective in-context learning approach for tasks to understand implied meanings. Specifically, we propose an approach in which an overview of pragmatic theories, such as Gricean pragmatics and Relevance Theory, is presented as a prompt to the language model, guiding it through a step-by-step reasoning process to derive a final interpretation. Experimental results showed that, compared to the baseline, which prompts intermediate reasoning without presenting pragmatic theories (0-shot Chain-of-Thought), our methods enabled language models to achieve up to 9.6\% higher scores on pragmatic reasoning tasks. Furthermore, we show that even without explaining the details of pragmatic theories, merely mentioning their names in the prompt leads to a certain performance improvement (around 1-3%) in larger models compared to the baseline.

Understanding region-wise correspondences between manga line art images is fundamental for high-level manga processing, supporting downstream tasks such as line art colorization and in-between frame generation. Unlike natural images that contain rich visual cues, manga line art consists only of sparse black-and-white strokes, making it challenging to determine which regions correspond across images. In this work, we introduce a new task: predicting region-wise correspondence between raw manga line art images without any annotations. To address this problem, we propose a Transformer-based framework trained on large-scale, automatically generated region correspondences. The model learns to suppress noisy matches and strengthen consistent structural relationships, resulting in robust patch-level feature alignment within and across images. During inference, our method segments each line art and establishes coherent region-level correspondences through edge-aware clustering and region matching. We construct manually annotated benchmarks for evaluation, and experiments across multiple datasets demonstrate both high patch-level accuracy and strong region-level correspondence performance, achieving 78.4-84.4% region-level accuracy. These results highlight the potential of our method for real-world manga and animation applications.

Video Moment Retrieval is a task in video understanding that aims to localize a specific temporal segment in an untrimmed video based on a natural language query. Despite recent progress in moment retrieval from videos using both traditional techniques and Multimodal Large Language Models (MLLM), most existing methods still rely on coarse temporal understanding and a single visual modality, limiting performance on complex videos. To address this, we introduce \textit{S}hot-aware \textit{M}ultimodal \textit{A}udio-enhanced \textit{R}etrieval of \textit{T}emporal \textit{S}egments (SMART), an MLLM-based framework that integrates audio cues and leverages shot-level temporal structure. SMART enriches multimodal representations by combining audio and visual features while applying \textbf{Shot-aware Token Compression}, which selectively retains high-information tokens within each shot to reduce redundancy and preserve fine-grained temporal details. We also refine prompt design to better utilize audio-visual cues. Evaluations on Charades-STA and QVHighlights show that SMART achieves significant improvements over state-of-the-art methods, including a 1.61\% increase in R1@0.5 and 2.59\% gain in R1@0.7 on Charades-STA.

Deploying continuous-control reinforcement learning policies on embedded hardware requires meeting tight latency and power budgets. Small FPGAs can deliver these, but only if costly floating-point pipelines are avoided. We study quantization-aware training (QAT) of policies for integer inference and we present a learning-to-hardware pipeline that automatically selects low-bit policies and synthesizes them to an Artix-7 FPGA. Across five MuJoCo tasks, we obtain policy networks that are competitive with full precision (FP32) policies but require as few as 3 or even only 2 bits per weight, and per internal activation value, as long as input precision is chosen carefully. On the target hardware, the selected policies achieve inference latencies on the order of microseconds and consume microjoules per action, favorably comparing to a quantized reference. Last, we observe that the quantized policies exhibit increased input noise robustness compared to the floating-point baseline.

We find that current sentence-embedding models produce outputs with a consistent bias: every embedding $e$ decomposes as $\tilde e + μ$, where the mean $μ$ is near-identical across all sentences. We study two training-free corrections -- subtracting $μ$ directly (R1), or projecting each embedding off the mean direction (R2) -- and show, via a first-order error-propagation argument, that R2 cancels the parallel component of mean-estimation error that R1 retains. Across 38 models on the Massive Multilingual Text Embedding Benchmark (MMTEB)~\citep{MMTEB}, R2 yields consistent classification gains (paired $\bar t = 3.31$, 29 of 38 models with $t>2$, zero losses), and the per-model mean norm $\Vertμ\Vert$ correlates with which models benefit most. A nine-method dose-response ablation on five models further reveals that mild single-direction removal helps, but full principal component analysis (PCA) whitening hurts every model we test, and that R2 and All-but-the-Top with depth one agree within $0.18$ pp downstream despite weak geometric alignment between $\hatμ$ and the centered top principal component.

Representing the temporal evolution of legal norms is a critical challenge for automated processing. While foundational frameworks exist, they lack a formal pattern for granular, component-level versioning, hindering the deterministic point-in-time reconstruction of legal texts required by reliable AI applications. This paper proposes a structured, temporal modeling pattern grounded in the LRMoo ontology. Our approach models a norm's evolution as a diachronic chain of versioned F1 Works, distinguishing between language-agnostic Temporal Versions (TV), each being a distinct Work, and their monolingual Language Versions (LV), modeled as F2 Expressions. The legislative amendment process is formalized through event-centric modeling, allowing changes to be traced precisely. Using the Brazilian Constitution as a case study, we demonstrate that our architecture enables the exact reconstruction of any part of a legal text as it existed on a specific date. This provides a verifiable semantic backbone for legal knowledge graphs, offering a deterministic foundation for trustworthy legal AI.

Power-logistics scheduling in modern seaports typically follows a predict-then-optimize pipeline. To enhance the decision quality of predictions, decision-focused learning has been proposed, which aligns the training of forecasting models with downstream decision outcomes. However, this end-to-end design inherently restricts the value of forecasting models to a specific task structure and therefore generalizes poorly to evolving tasks induced by varying vessel arrivals. We address this gap with a decision-focused continual learning framework that adapts online to a stream of scheduling tasks. Specifically, we introduce Fisher-information-based regularization to enhance cross-task generalization by preserving parameters critical to prior tasks. A differentiable convex surrogate is also developed to stabilize gradient backpropagation. The proposed approach enables learning a decision-aligned forecasting model across a varying task stream with sustainable long-term computational and memory requirements. Experiments calibrated to Jurong Port show improved decision performance and cross-task generalization over existing methods, together with reduced computational cost and a bounded memory footprint.

We introduce Reinforcement Learning (RL) with Adaptive Verifiable Environments (RLVE), an approach using verifiable environments that procedurally generate problems and provide algorithmically verifiable rewards, to scale up RL for language models (LMs). RLVE enables each verifiable environment to dynamically adapt its problem difficulty distribution to the policy model's capabilities as training progresses. In contrast, static data distributions often lead to vanishing learning signals when problems are either too easy or too hard for the policy. To implement RLVE, we create RLVE-Gym, a large-scale suite of 400 verifiable environments carefully developed through manual environment engineering. Using RLVE-Gym, we show that environment scaling, i.e., expanding the collection of training environments, consistently improves generalizable reasoning capabilities. RLVE with joint training across all 400 environments in RLVE-Gym yields a 3.37% absolute average improvement across six reasoning benchmarks, starting from one of the strongest 1.5B reasoning LMs. By comparison, continuing this LM's original RL training yields only a 0.49% average absolute gain despite using over 3x more compute. We release our code publicly.

Even though we know the precise computations that lead from a large language model (LLM) input to its output this computation remains very hard to interpret. One way to make it easier to understand this process is by creating a sparse computational graph that captures most of the model behavior with smallest number of computational nodes. Cross-layer transcoders (CLT) decompose the dense computations of the MLP but the resulting circuits still contain thousands of nodes even for short prompts. Existing automated interpretation methods label individual features from corpus activations, and it often happens that these labels are not validated by causal intervention. We introduce probe prompting, a transparent rule-based pipeline that groups the features of an attribution graph into concept-aligned supernodes from their responses on a small set of concept-targeted probe prompts, summarized as Cross-Prompt Activation Signatures (CPAS). Across four factual domains, on Gemma-2-2B with a public CLT dictionary and 45,596 entity-swap interventions, we find that the labeled supernodes have the predicted steering behavior in every one of them. Code, datasets, and an interactive demo are released anonymously as a reusable harness for calibrating supernode labels against causal interventions.

Social and collaborative platforms emit multivariate time-series traces in which early interactions -- such as views, likes, or downloads -- are followed, sometimes months or years later, by higher impact like citations, sales, or reviews. We formalize this setting as Lead-Lag Forecasting (LLF): given an early usage channel (the lead), predict a correlated but temporally shifted outcome channel (the lag). Despite the ubiquity of such patterns, LLF has not been treated as a unified forecasting problem within the time-series community, largely due to the absence of standardised datasets. To anchor research in LLF, here we present two high-volume benchmark datasets: arXiv (accesses -> citations of 2.3M papers) and GitHub (pushes/stars -> forks of 3M repositories). Our datasets provide ideal testbeds for lead-lag forecasting, by capturing long-horizon dynamics across years, spanning the full spectrum of outcomes, and avoiding survivorship bias in sampling. We documented all technical details of data curation and cleaning, verified the presence of lead-lag dynamics through statistical and classification tests, and benchmarked parametric and non-parametric baselines for regression. Our study establishes LLF as a novel forecasting paradigm and lays an empirical foundation for its systematic exploration in social and usage data.

We present the bulk--boundary decomposition as a new framework for understanding the training dynamics of deep neural networks. Starting from the stochastic gradient descent formulation, we show that the Lagrangian can be reorganized into a data-independent bulk term and a data-dependent boundary term. The bulk captures the intrinsic dynamics set by network architecture and activation functions, while the boundary reflects stochastic interactions from training samples at the input and output layers. This decomposition exposes the local and homogeneous structure underlying deep networks. As a physical consequence of locality and homogeneity, we derive the energy continuity equation within a deep neural network.

Game environments provide rich, controllable settings that stimulate many aspects of real-world complexity. As such, game agents offer a valuable testbed for exploring capabilities relevant to Artificial General Intelligence. Recently, the emergence of Large Language Models (LLMs) provides new opportunities to endow these agents with generalizable reasoning, memory, and adaptability in complex game environments. This survey offers an up-to-date review of LLM-based game agents (LLMGAs) through a unified reference architecture. At the single-agent level, we synthesize existing studies around three core components: memory, reasoning, and perception-action interfaces, which jointly characterize how language enables agents to perceive, think, and act. At the multi-agent level, we outline how communication protocols and organizational models support coordination, role differentiation, and large-scale social behaviors. To contextualize these designs, we introduce a challenge-centered taxonomy linking six major game genres to their dominant agent requirements, from low-latency control in action games to open-ended goal formation in sandbox worlds. A curated list of related papers is available at https://github.com/git-disl/awesome-LLM-game-agent-papers

Temporal reasoning involves understanding how systems evolve over time through input-driven state transitions. A key aspect is temporal causal reasoning, causally reasoning about what prior inputs were necessary in causing an observed outcome. While large language models (LLMs) perform well at forward simulation, predicting outputs from inputs, they struggle to identify the minimal causal inputs of outcomes. To study this distinction, we define two tasks: \textit{trace simulation} (SIM), which requires models to simulate system execution, and \textit{minimal causal attribution} (MIN), which identifies the minimal set of inputs necessary for a given outcome. We introduce \textsc{TempoBench}, the first formally verified benchmark for temporal causal reasoning, built from synthesized Mealy machines with controllable complexity and provably correct causal labels. Across frontier models, we observe that despite achieving up to 96\% accuracy on the SIM task, performance on the causal attribution MIN task drops below 25\%; models fail to reason about causal necessity. Over 94\% of causal errors involve overspecification, where models perform retrieval and list all possible inputs rather than reasoning about the minimal causal subset. Fine-tuning on \textsc{TempoBench} training corpus improves causal reasoning and generalizes better than math, code, or instruction training, with gains across standard reasoning benchmarks.

The choice of activation function plays a critical role in neural networks, yet most architectures still rely on fixed, uniform activation functions across all neurons. We introduce SmartMixed, a novel two-phase training strategy that allows networks to learn optimal per-neuron activation functions while preserving computational efficiency at inference. In the first phase, neurons adaptively select from a pool of candidate activation functions (ReLU, Sigmoid, Tanh, Leaky\_ReLU, ELU, SELU) using a differentiable hard mixture mechanism. In the second phase, each neuron's activation function is fixed according to the learned selection, resulting in a computationally efficient network that supports continued training with optimized vectorized operations. We evaluate SmartMixed on the MNIST dataset using feedforward neural networks of different architectures. Our analysis reveals that neurons in different layers exhibit distinct preferences for activation functions, providing insights into the functional diversity within neural architectures. We also demonstrated that SmartMixed effectively trains the network by allowing neurons to select their preferred activation functions, competing against models using a single fixed state-of-the-art activation function.

Pedestrian simulation traditionally relies on expert-tuned, hand-crafted models that limit scalability and generalization. Meanwhile, large-scale video generation models have achieved high visual realism across diverse settings, motivating exploration of their potential as general-purpose world simulators. Existing benchmarks primarily assess single-subject realism rather than scenes with multiple interacting people, leaving the plausibility of multi-agent dynamics in generated videos untested. We propose a rigorous evaluation protocol to benchmark text-to-video (T2V) and image-to-video (I2V) models as implicit simulators of pedestrian dynamics. For I2V, we leverage start frames from established datasets to enable direct comparison with ground truth videos, while for T2V we design a prompt suite covering varied crowd densities and interaction types. A key component is a method to reconstruct 2D bird's-eye view trajectories from pixel-space without known camera parameters. Our analysis shows that leading models exhibit effective priors for plausible multi-agent behavior, though issues such as merging and disappearing pedestrians reveal limits to their physical consistency.

Evaluating conversational AI systems that use external tools is challenging, as errors can arise from complex interactions among user, agent, and tools. While existing evaluation methods assess either user satisfaction or agents' tool-calling capabilities, they fail to capture critical errors in multi-turn tool-augmented dialogues-such as when agents misinterpret tool results yet appear satisfactory to users. We introduce TRACE, a benchmark of systematically synthesized tool-augmented conversations covering diverse error cases. Evaluation with state-of-the-art conversation evaluation frameworks reveals that all approaches remain far from ideal performance, demonstrating the fundamental difficulty of this benchmark.

Stochastic differential equations (SDEs) describe dynamical systems where deterministic flows, governed by a drift function, are superimposed with random fluctuations, dictated by a diffusion function. The accurate estimation (or discovery) of these functions from data is a central problem in machine learning, with wide application across the natural and social sciences. Yet current solutions either rely heavily on prior knowledge of the dynamics or involve intricate training procedures. We introduce FIM-SDE (Foundation Inference Model for SDEs), a pretrained recognition model that delivers accurate in-context (or zero-shot) estimation of the drift and diffusion functions of low-dimensional SDEs, from noisy time series data, and allows rapid finetuning to target datasets. Leveraging concepts from amortized inference and neural operators, we (pre)train FIM-SDE in a supervised fashion to map a large set of noisy, discretely observed SDE paths onto the space of drift and diffusion functions. We demonstrate that FIM-SDE achieves robust in-context function estimation across a wide range of synthetic and real-world processes -- from canonical SDE systems (e.g., double-well dynamics or weakly perturbed Lorenz attractors) to stock price recordings and oil-price and wind-speed fluctuations -- while matching the performance of symbolic, Gaussian process and Neural SDE baselines trained on the target datasets. When finetuned to the target processes, we show that FIM-SDE consistently outperforms all these baselines.

The reasoning pattern of Large language models (LLMs) remains opaque, and reinforcement learning (RL) typically applies uniform credit across an entire generation, blurring the distinction between pivotal and routine steps. This work positions attention as a privileged substrate that renders the internal logic of LLMs legible, not merely as a byproduct of computation, but as a mechanistic blueprint of reasoning itself. We first distinguish attention heads between locally and globally focused information processing and reveal that locally focused heads produce a sawtooth pattern near the diagonal indicating phrasal chunks, while globally focused heads expose tokens that exert broad downstream influence over future tokens. We formalize these with two metrics: 1) Windowed Average Attention Distance, which measures the extent of backward attention within a clipped window; 2) Future Attention Influence, which quantifies a token's global importance as the average attention it receives from subsequent tokens. Taken together, these signals reveal a recurring preplan-and-anchor mechanism, where the model first performs a long-range contextual reference to generate an introductory token, which is immediately followed by or coincides with a semantic anchor token that organizes subsequent reasoning. Leveraging these insights, we introduce three novel RL strategies that dynamically perform targeted credit assignment to critical nodes (preplan tokens, anchor tokens, and their temporal coupling) and show consistent performance gains across various reasoning tasks. By aligning optimization with the model's intrinsic reasoning rhythm, we aim to transform opaque optimization into an actionable structure-aware process, hoping to offer a potential step toward more transparent and effective optimization of LLM reasoning.

Large Language Models have shown strong scientific reasoning ability, but their performance on materials science problems remains less studied. To fill this gap, we introduce MatSciBench, a comprehensive college-level benchmark comprising 1340 problems that span the essential subdisciplines of materials science. MatSciBench features a structured and fine-grained taxonomy that categorizes materials science questions into 6 primary fields and 31 subfields, together with a three-tier difficulty classification based on the reasoning length needed to solve each problem. MatSciBench includes detailed reference solutions for 946 questions, supports process-level error analysis, and contains 315 questions with images for evaluating multimodal reasoning. We evaluate leading thinking and non-thinking LLMs on MatSciBench, and further test three reasoning methods for non-thinking models: basic chain-of-thought prompting, tool augmentation, and self-correction. The results show that current models still face clear limits in college-level materials science reasoning. DeepSeek-R1 achieves the highest score on text-only questions at 75.22% accuracy, and GPT-5 performs the best on questions with images at 53.02%. Our analysis shows that tool augmentation improves many non-thinking models in a token-efficient way, while self-correction often fails to provide reliable gains and can revise correct answers into incorrect ones. We further analyze performance across difficulty levels, reasoning efficiency, multimodal reasoning, and failure patterns, and find that current models are mainly limited by domain knowledge gaps, calculation errors, problem comprehension failures, and difficulty in extracting precise information from scientific figures. Overall, MatSciBench provides a clear testbed for measuring current LLM limitations and guiding future work on scientific reasoning in materials science.

Search agents trained with reinforcement learning (RL) interleave reasoning with tool calls in a multi-turn, tool-integrated reasoning (TIR) loop, where each tool invocation returns an environment observation that is appended to the agent's context. As the rollout proceeds, these raw observations accumulate, inflating token cost and diluting the signal available for downstream reasoning. Unlike single-pass retrieve-then-read pipelines, where context compression is a one-time postprocessing step, the multi-turn RL setting requires compression that runs at every observation step while remaining decoupled from policy optimization. We introduce RECON (REasoning with CONdensation), a framework that addresses this challenge by inserting a dedicated observation compressor into the reasoning loop. The compressor is trained via a two-stage curriculum: relevance pretraining on QA datasets followed by multi-aspect distillation from proprietary LLMs, and remains frozen during RL training to preserve policy stability. Integrated into the Search-R1 search-agent pipeline, RECON reduces total context length by 35%, improves training speed by 5.4% and inference latency by 30.9%, while boosting average exact-match by 14.5% on the 3B agent and 3.0% on the 7B agent, with particular strength in multi-hop QA. These results establish learned observation compression as a key component for building practical, scalable RL-trained search agents.

The rapid advancement of Low-Altitude Economy Networks (LAENets) has enabled a variety of applications, including aerial surveillance, environmental sensing, and semantic data collection. To support these scenarios, unmanned aerial vehicles (UAVs) equipped with onboard vision-language models (VLMs) offer a promising solution for real-time multimodal inference. However, ensuring both inference accuracy and communication efficiency remains a significant challenge due to limited onboard resources and dynamic network conditions. In this paper, we first propose a UAV-enabled LAENet system model that jointly captures UAV mobility, user-UAV communication, and the onboard visual question answering (VQA) pipeline. Based on this model, we formulate a mixed-integer non-convex optimization problem to minimize task latency and power consumption under user-specific accuracy constraints. To solve the problem, we design a hierarchical optimization framework composed of two parts: (i) an Alternating Resolution and Power Optimization (ARPO) algorithm for resource allocation under accuracy constraints, and (ii) a Large Language Model-augmented Reinforcement Learning Approach (LLaRA) for adaptive UAV trajectory optimization. The large language model (LLM) serves as an expert in refining reward design of reinforcement learning in an offline fashion, introducing no additional latency in real-time decision-making. Numerical results demonstrate the efficacy of our proposed framework in improving inference performance and communication efficiency under dynamic LAENet conditions.

Oversampling is one of the most widely used approaches for addressing imbalanced classification. The core idea is to generate additional minority samples to rebalance the dataset. Most existing methods, such as SMOTE, require converting categorical variables into numerical vectors, which often leads to information loss. Recently, large language model (LLM)-based methods have been introduced to overcome this limitation. However, current LLM-based approaches typically generate minority samples with limited diversity, reducing robustness and generalizability in downstream classification tasks. To address this gap, we propose a novel LLM-based oversampling method designed to enhance diversity. First, we introduce a sampling strategy that conditions synthetic sample generation on both minority labels and features. Second, we develop a new permutation strategy for fine-tuning pre-trained LLMs. Third, we fine-tune the LLM not only on minority samples but also on interpolated samples to further enrich variability. Extensive experiments on 10 tabular datasets demonstrate that our method significantly outperforms eight SOTA baselines. The generated synthetic samples are both realistic and diverse. Moreover, we provide theoretical analysis through an entropy-based perspective, proving that our method encourages diversity in the generated samples.

Macro-aligned micro-records are crucial for credible simulations in social science and urban studies. For example, epidemic models are only reliable when individual-level mobility and contacts mirror real behavior, while aggregates match real-world statistics like case counts or travel flows. However, collecting such fine-grained data at scale is impractical, leaving researchers with only macro-level data. LLMSynthor addresses this by turning a pretrained LLM into a macro-aware simulator that generates realistic micro-records consistent with target macro-statistics. It iteratively builds synthetic datasets: in each step, the LLM generates batches of records to minimize discrepancies between synthetic and target aggregates. Treating the LLM as a nonparametric copula allows the model to capture realistic joint dependencies among variables. To improve efficiency, LLM Proposal Sampling guides the LLM to propose targeted record batches, specifying variable ranges and counts, to efficiently correct discrepancies while preserving realism grounded in the model's priors. Evaluations across domains (mobility, e-commerce, population) show that LLMSynthor achieves strong realism, statistical fidelity, and practical utility, making it broadly applicable to economics, social science, and urban studies.

Vision-and-Language Navigation (VLN) has been studied mainly in either discrete or continuous spaces, with little attention to dynamic, crowded environments. We present HA-VLN 2.0, a unified benchmark introducing explicit social-awareness constraints. Our contributions are: (i) a standardized task and metrics capturing both goal accuracy and personal-space adherence; (ii) HAPS 2.0 dataset and simulators modeling multi-human interactions, outdoor contexts, and finer language-motion alignment; (iii) benchmarks on 16,844 socially grounded instructions, revealing sharp performance drops of leading agents under human dynamics and partial observability; and (iv) real-world robot experiments validating sim-to-real transfer, with an open leaderboard enabling transparent comparison. Results show that explicit social modeling improves navigation robustness and reduces collisions, underscoring necessity of human-centric approaches. By releasing datasets, simulators, baselines, and protocols, HA-VLN 2.0 provides a strong foundation for safe, human-aware navigation research.

Latent world models are a promising approach for learning state representations and dynamics directly from high-dimensional observations, enabling robot control in hard-to-model settings. However, control performance ultimately depends on the latent representation encoding the required information for the task. In this work, we study latent-space safe control problems and show how partial observability can induce control failures when safety-relevant information is not preserved in the latent state. Specifically, we identify two world model failure modes: estimation gaps, where current observations do not reveal safety-critical quantities (e.g., temperature in a cooking task), and prediction gaps, where failures are observable once they occur but cannot be reliably anticipated from available observations. We introduce two diagnostics for these gaps: a mutual-information-based measure of safety observability and a rollout-based measure of future safety predictability. Finally, we present mitigation strategies for each failure mode: privileged multimodal supervision for estimation gaps and conformal risk calibration for prediction gaps. Across two hardware case studies -- using unimodal RGB world models and multimodal RGB+Tactile and RGB+Thermal variants -- we show that these mitigation strategies improve the safety of a Franka Research 3 manipulator on challenging cooking tasks under partial observability, albeit with increased conservativeness. More broadly, our work raises the question of when world model state representations are sufficient for reliable robot control

Reasoning models enhance performance by tackling problems in a step-by-step manner, decomposing them into sub-problems and exploring long chains of thought before producing an answer. However, applying extended reasoning to every step introduces substantial redundancy, as sub-problems vary widely in difficulty and complexity: a small number of pivotal steps are genuinely challenging and decisive for the final answer, while many others only involve straightforward revisions or simple computations. Therefore, a natural idea is to endow reasoning models with the ability to adaptively respond to this variation, rather than treating all steps with the same level of elaboration. To this end, we propose MixReasoning, a framework that dynamically adjusts the depth of reasoning within a single response. The resulting chain of thought then becomes a mixture of detailed reasoning on difficult steps and concise inference on simpler ones. Experiments on GSM8K, MATH-500, and AIME show that MixReasoning shortens reasoning length and substantially improves efficiency without compromising accuracy.

Hallucinations in diffusion models are samples with structural inconsistencies that can emerge due to the excessive smoothing of the learned score function, which in turn leads to interpolations between modes of the data distribution. Since semantic interpolations are often desirable and contribute to sample diversity, we believe that a nuanced and targeted solution is required to address diffusion model hallucinations. In this work, we introduce Dynamic Guidance, which mitigates hallucinations by selectively sharpening the score function only along the pre-determined directions known to cause artifacts, while preserving valid semantic variations. This sharpening can be performed using either pre-determined classes or semantically coherent clusters that form pseudo-classes over the data distribution. The latter allows for a principled extension of Dynamic Guidance to text-to-image generation, where we select modes to correspond to fine-grained contextual differences in textual descriptions. To our knowledge, this is the first approach that addresses hallucinations at generation time rather than through post-hoc filtering. Dynamic Guidance substantially reduces hallucinations on both controlled and natural image datasets, significantly outperforming baselines.

Creating ultra-high-resolution spatially varying bidirectional reflectance functions (SVBRDFs) is critical for photorealistic 3D content creation, to faithfully represent fine-scale surface details required for close-up rendering. However, achieving 4K generation faces two key challenges: (1) the need to synthesize multiple reflectance maps at full resolution, which multiplies the pixel budget and imposes prohibitive memory and computational cost, and (2) the requirement to maintain strong pixel-level alignment across maps at 4K, which is particularly difficult when adapting pretrained models designed for the RGB image domain. We introduce HiMat, a diffusion-based framework tailored for efficient and diverse 4K SVBRDF generation. To address the first challenge, HiMat performs generation in a high-compression latent space via DC-AE, and employs a pretrained diffusion transformer with linear attention to improve per-map efficiency. To address the second challenge, we propose CrossStitch, a lightweight convolutional module that enforces cross-map consistency without incurring the cost of global attention. Our experiments show that HiMat achieves high-fidelity 4K SVBRDF generation with superior efficiency, structural consistency, and diversity compared to prior methods. Beyond materials, our framework also generalizes to related applications such as intrinsic decomposition.

Large time series foundation models often adopt channel-independent architectures to handle varying data dimensions, but this design ignores crucial cross-channel dependencies. Meanwhile, existing cross-modal methods predominantly rely on textual modalities, leaving the spatial pattern recognition capabilities of vision models underexplored for time series analysis. To address these limitations, we propose VFEM, a cross-modal forecasting model that leverages pre-trained large vision models (LVMs) to capture complex cross-variable patterns. VFEM transforms multivariate time series into visual representations, enabling LVMs to perceive spatial relationships that are not explicitly modeled by channel-independent models. Through a dual-branch architecture, visual and temporal features are independently extracted and then fused via cross-modal attention, allowing complementary information from both modalities to enhance forecasting. By freezing the LVM and training only 7.45% of the total parameters, VFEM achieves competitive performance on multiple benchmarks, offering a new perspective on multivariate time series forecasting.

Multilingual Retrieval-Augmented Generation (mRAG) systems enable language models to answer knowledge-intensive queries with citation-supported responses across languages. Despite their growing use, an open questions is whether the mixture of different document languages impacts generation and citation behavior in unintended ways. To investigate this, we introduce a controlled methodology using model internals to measure language preference while holding other factors such as document relevance constant. Across eight languages and six open-weight models, we find that models preferentially cite English sources when queries are in English, with this bias amplified for lower-resource languages and for documents positioned mid-context. More crucially, we find that models sometimes trade-off document relevance for language preference, indicating that citation choices are not always driven by informativeness alone. Our findings shed light on how language models leverage multilingual context and influence citation behavior.