Small language models (SLMs) enable scalable tool-augmented multi-agent systems where multiple SLMs handle subtasks orchestrated by a powerful coordinator. However, they struggle with tool-use tasks, particularly in selecting appropriate tools and identifying correct parameters. A common failure mode is \textit{schema misalignment}: models hallucinate plausible tool names that are absent from the provided tool schema, due to different naming conventions internalized during pretraining. Rather than training models to adapt to unfamiliar schemas, we propose adapting schemas to align with models' pretrained knowledge. We introduce \textbf{PA-Tool} (Pretraining-Aligned Tool Schema Generation), a training-free method that leverages peakedness, a signal used in contamination detection that indicates pretraining familiarity, to rename tool components. By generating multiple candidates and selecting the candidate with the highest peakedness, PA-Tool identifies pretraining-aligned naming patterns. Experiments on MetaTool and RoTBench show improvements of up to 17\%, with schema misalignment errors reduced by 80\%. PA-Tool enables small models to substantially improve tool-use accuracy without retraining, showing that schema-level interventions can unlock the tool-use potential of resource-efficient models. Our code is available at https://github.com/holi-lab/PA-Tool.

Historical archives on weather events are collections of enduring primary source records that offer rich, untapped narratives of how societies have experienced and responded to extreme weather events. These qualitative accounts provide insights into societal vulnerability and resilience that are largely absent from meteorological records, making them valuable for climate scientists to understand societal responses. However, their vast scale, noisy digitized quality, and archaic language make it difficult to transform them into structured knowledge for climate research. To address this challenge, we introduce WeatherArchive-Bench, the first benchmark for evaluating retrieval-augmented generation (RAG) systems on historical weather archives. WeatherArchive-Bench comprises two tasks: WeatherArchive-Retrieval, which measures a system's ability to locate historically relevant passages from over one million archival news segments, and WeatherArchive-Assessment, which evaluates whether Large Language Models (LLMs) can classify societal vulnerability and resilience indicators from extreme weather narratives. Extensive experiments across sparse, dense, and re-ranking retrievers, as well as a diverse set of LLMs, reveal that dense retrievers often fail on historical terminology, while LLMs frequently misinterpret vulnerability and resilience concepts. These findings highlight key limitations in reasoning about complex societal indicators and provide insights for designing more robust climate-focused RAG systems from archival contexts. The constructed dataset and evaluation framework are publicly available at https://anonymous.4open.science/r/WeatherArchive-Bench/.

Black-box optimization (BBO) underpins advances in domains such as AutoML and Materials Informatics, yet implementations of algorithms and benchmarks remain fragmented across research communities. We introduce OptunaHub (https://hub.optuna.org/), a community-oriented, decentralized platform for distributing BBO components under a unified Optuna-compatible interface. OptunaHub enables independent publication, discovery, and reuse of optimization algorithms and benchmark problems through a lightweight Python module, a contributor-driven registry, and a searchable web interface. The source code is publicly available in the \href{https://github.com/optuna/optunahub}{\texttt{optunahub}}, \href{https://github.com/optuna/optunahub-registry}{\texttt{optunahub-registry}}, and \href{https://github.com/optuna/optunahub-web}{\texttt{optunahub-web}} repositories under the Optuna organization on GitHub (https://github.com/optuna/).

Large language models leverage both parametric knowledge acquired during pretraining and in-context knowledge provided at inference time. Crucially, when these sources conflict, models arbitrate based on their internal confidence, preferring parametric knowledge for high-confidence facts while deferring to context for less familiar ones. However, the training conditions that give rise to these fundamental behaviors remain unclear. Here we conduct controlled experiments using synthetic corpora to identify the specific data properties that shape knowledge utilization. Our results reveal a counterintuitive finding: the robust, balanced use of both knowledge sources is an emergent property that requires the co-occurrence of three factors typically considered detrimental, including (i) intra-document repetition, (ii) a moderate degree of intra-document inconsistency, and (iii) a skewed knowledge distribution. We further show that these dynamics arise in real-world language model pretraining and analyze how post-training procedures reshape arbitration strategies. Together, our findings provide empirical guidance for designing training data that supports the reliable integration of parametric and in-context knowledge in language models.

Vision-language models (VLMs) such as GPT (Generative Pre-Trained Transformer) have shown potential for medical image interpretation; however, challenges remain in generating reliable radiological findings in clinical practice, as exemplified by dental pathologies. This study proposes a Self-correction Loop with Structured Output (SLSO) framework as an integrated processing methodology to enhance the accuracy and reliability of AI-generated findings for jaw cysts in dental panoramic radiographs. Dental panoramic radiographs with jaw cysts were used to implement a 10-step integrated processing framework incorporating image analysis, structured data generation, tooth number extraction, consistency checking, and iterative regeneration. The framework functioned as an external validation mechanism for GPT outputs. Performance was compared against the conventional Chain-of-Thought (CoT) method across seven evaluation items: transparency, internal structure, borders, root resorption, tooth movement, relationships with other structures, and tooth number. The SLSO framework improved output accuracy for multiple items compared to the CoT method, with the most notable improvements observed in tooth number identification, tooth movement detection, and root resorption assessment. In successful cases, consistently structured outputs were achieved after up to five regenerations. The framework enforced explicit negative finding descriptions and suppressed hallucinations, although accurate identification of extensive lesions spanning multiple teeth remained limited. This investigation established the feasibility of the proposed integrated processing methodology and provided a foundation for future validation studies with larger, more diverse datasets.

Most existing approaches formulate action quality assessment and skill proficiency estimation as discriminative prediction tasks, typically producing discrete labels or scores without explicitly modeling the reasoning process underlying the assessment. We instead reformulate the problem as generative vision-language modeling, introducing ProfVLM, a parameter-efficient vision-language model that jointly predicts proficiency levels and generates expert-like natural language feedback from multi-view videos. ProfVLM leverages conditional language generation to provide actionable insights along with quantitative evaluation scores. Central to our method is an AttentiveGatedProjector that dynamically fuses and projects multi-view egocentric and exocentric features from a frozen TimeSformer backbone into a language model fine-tuned for feedback generation. Trained on EgoExo4D with expert commentaries, ProfVLM surpasses state-of-the-art methods while using up to 20x fewer parameters and reducing training time by up to 60% compared to existing classification-based methods. By providing natural language critiques aligned with performance levels, this work shows that generative vision-language modeling offers a powerful and efficient paradigm shift for interpretable action quality assessment.

Understanding risk in autonomous driving requires not only perception and prediction, but also high-level reasoning about agent behavior and context. Current Vision Language Model (VLM)-based methods primarily ground agents in static images and provide qualitative judgments, lacking the spatio-temporal reasoning needed to capture how risks evolve over time. To address this gap, we propose NuRisk, a comprehensive Visual Question Answering (VQA) dataset comprising 2.9K scenarios and 1.1M agent-level samples, built on real-world data from nuScenes and Waymo, completed with safety-critical scenarios from the CommonRoad simulator. The dataset provides Bird's-eye view (BEV) based sequential images with quantitative, agent-level risk annotations, enabling spatio-temporal reasoning. We benchmark well-known VLMs across different prompting techniques and find that they fail to perform explicit spatio-temporal reasoning, resulting in a peak accuracy of 33% at high latency. To address these shortcomings, our fine-tuned 7B VLM agent improves accuracy to 41% and reduces latency by 75%, demonstrating explicit spatio-temporal reasoning capabilities that proprietary models lacked. While this represents a significant step forward, the modest accuracy underscores the profound challenge of the task, establishing NuRisk as a critical benchmark for advancing spatio-temporal reasoning in autonomous driving. More information can be found at https://github.com/TUM-AVS/NuRisk.

Interleaved-Modal Chain-of-Thought (I-MCoT) advances vision-language reasoning, such as Visual Question Answering (VQA). This paradigm integrates specially selected visual evidence from the input image into the context of Vision-Language Models (VLMs), enabling them to ground their reasoning logic in these details. Accordingly, the efficacy of an I-MCoT framework relies on identifying what to see (evidence selection) and when to see it (triggering of insertions). However, existing methods fall short in both aspects. First, for selection, they rely on attention signals, which are unreliable -- particularly under severe granularity imbalance between the brief textual query and the informative image. Second, for triggering, they adopt static triggers, which fail to capture the VLMs' dynamic needs for visual evidence. To this end, we propose a novel I-MCoT framework, Active Information-driven Multi-modal Chain-of-Thought (AIM-CoT), which aims to improve both evidence selection and insertion triggering via: (1) Context-enhanced Attention-map Generation (CAG) to mitigate granularity imbalance via textual context enhancement; (2) Active Visual Probing (AVP) to proactively select the most informative evidence via an information foraging process; and (3) Dynamic Attention-shift Trigger (DAT) to precisely activate insertions when VLM's attention shifts from text to visual context. Experiments across three benchmarks and four backbones demonstrate AIM-CoT's consistent superiority. Our code is available at https://anonymous.4open.science/r/AIMCoT.

To gain finer regional forecasts, many works have explored the regional integration from the global atmosphere, e.g., by solving boundary equations in physics-based methods or cropping regions from global forecasts in data-driven methods. However, the effectiveness of these methods is often constrained by static and imprecise regional boundaries, resulting in poor generalization ability. To address this issue, we propose Spatial-Temporal Weather Forecasting (STCast), a novel AI-driven framework for adaptive regional boundary optimization and dynamic monthly forecast allocation. Specifically, our approach employs a Spatial-Aligned Attention (SAA) mechanism, which aligns global and regional spatial distributions to initialize boundaries and adaptively refines them based on attention-derived alignment patterns. Furthermore, we design a Temporal Mixture-of-Experts (TMoE) module, where atmospheric variables from distinct months are dynamically routed to specialized experts using a discrete Gaussian distribution, enhancing the model's ability to capture temporal patterns. Beyond global and regional forecasting, we evaluate our STCast on extreme event prediction and ensemble forecasting. Experimental results demonstrate consistent superiority over state-of-the-art methods across all four tasks. Code: https://github.com/chenhao-zju/STCast

LLMs have low GPU efficiency and high latency due to autoregressive decoding. Speculative decoding (SD) mitigates this using a small draft model to speculatively generate multiple tokens, which are then verified in parallel by a target model. However, when speculation accuracy is low, the overhead from rejected tokens can offset the benefits, limiting SD's effectiveness, especially at large batch sizes. To address this, we propose Speculative Verification (SV), an efficient augmentation to SD that dynamically predicts speculation accuracy and adapts the verification length to maximize throughput. SV introduces a companion model - a small auxiliary model similar in size to the draft model - to estimate the alignment between draft and target model distributions. By maximizing the information gain from quantifying this alignment, SV refines verification decisions, reducing wasted computation on rejected tokens and improving decoding efficiency. Moreover, SV requires no modifications to the draft or target models and is compatible with existing SD variants. We extensively evaluated SV on publicly available LLMs across three NLP tasks using nine combinations of draft, companion, and target models, including 13B-72B target models and three types of variations: base (no finetuning), instruction-tuned, and task fine-tuned. Across all experiments and batch sizes (4-80), SV consistently outperforms both SD and standard decoding with the target model. It improves SD performance by up to 2$\times$, with an average speedup of 1.4 $\times$ in large-batch settings (batch sizes 32-80). These results demonstrate SV's robustness, scalability, and practical utility for efficient LLM inference.

The LLM-as-a-judge paradigm is widely used in both evaluating free-text model responses and reward modeling for model alignment and fine-tuning. Recently, fine-tuning judges with judge-specific data has emerged as an often preferred choice over directly prompting frontier models as judges, as the former achieves better performance with smaller model sizes while being more robust to common biases. However, the standard evaluation ignores several practical concerns of fine-tuned judges regarding their real-world deployment. In this paper, we identify and formalize three aspects that affect the shelf life of these judges: future-proofing and backward-compatibility -- how well judges fine-tuned on responses by today's generator models perform on responses by future models or past models, as well as question generalization -- how well judges generalize to unseen questions at test time. We study these three aspects under a unified framework with varying train and test distributions in two reasoning datasets, three SFT- and DPO-based fine-tuning algorithms, and three different backbone models. Experiments suggest that future-proofing is challenging for most models, while backward-compatibility is relatively easy, with DPO-trained models consistently improving performance. We further find that continual learning provides a more balanced adaptation to shifts between older and newer response distributions than training solely on stronger or weaker responses. Moreover, all models exhibit some degree of performance degradation when moving from questions seen during training to unseen ones, showing that current judges do not fully generalize to unseen questions. These findings provide insights into practical considerations for developing and deploying judge models in the face of ever-changing generators.

Recent work by Chatzi et al. and Ravfogel et al. has developed, for the first time, a method for generating counterfactuals of probabilistic Large Language Models. Such counterfactuals tell us what would - or might - have been the output of an LLM if some factual prompt ${\bf x}$ had been ${\bf x}^*$ instead. The ability to generate such counterfactuals is an important necessary step towards explaining, evaluating, and eventually improving, the behavior of LLMs. I argue, however, that the existing method rests on an ambiguous interpretation of LLMs: it does not interpret LLMs literally, for the method involves the assumption that one can change the implementation of an LLM's sampling process without changing the LLM itself, nor does it interpret LLMs as intended, for the method involves explicitly representing a nondeterministic LLM as a deterministic causal model. I here present a much simpler method for generating counterfactuals that is based on an LLM's intended interpretation by representing it as a nondeterministic causal model instead. The advantage of my simpler method is that it is directly applicable to any black-box LLM without modification, as it is agnostic to any implementation details. The advantage of the existing method, on the other hand, is that it directly implements the generation of a specific type of counterfactuals that is useful for certain purposes, but not for others. I clarify how both methods relate by offering a theoretical foundation for reasoning about counterfactuals in LLMs based on their intended semantics, thereby laying the groundwork for novel application-specific methods for generating counterfactuals.

Recently, camera-controlled video generation has seen rapid development, offering more precise control over video generation. However, existing methods predominantly focus on camera control in perspective projection video generation, while geometrically consistent panoramic video generation remains challenging. This limitation is primarily due to the inherent complexities in panoramic pose representation and spherical projection. To address this issue, we propose CamPVG, the first diffusion-based framework for panoramic video generation guided by precise camera poses. We achieve camera position encoding for panoramic images and cross-view feature aggregation based on spherical projection. Specifically, we propose a panoramic Plücker embedding that encodes camera extrinsic parameters through spherical coordinate transformation. This pose encoder effectively captures panoramic geometry, overcoming the limitations of traditional methods when applied to equirectangular projections. Additionally, we introduce a spherical epipolar module that enforces geometric constraints through adaptive attention masking along epipolar lines. This module enables fine-grained cross-view feature aggregation, substantially enhancing the quality and consistency of generated panoramic videos. Extensive experiments demonstrate that our method generates high-quality panoramic videos consistent with camera trajectories, far surpassing existing methods in panoramic video generation.

Pretraining on large-scale collections of PDE-governed spatiotemporal trajectories has recently shown promise for building generalizable models of dynamical systems. Yet most existing PDE foundation models rely on deterministic Transformer architectures, which lack generative flexibility for many science and engineering applications. We propose Flow Marching, an algorithm that bridges neural operator learning with flow matching motivated by an analysis of error accumulation in physical dynamical systems, and we build a generative PDE foundation model on top of it. By jointly sampling the noise level and the physical time step between adjacent states, the model learns a unified velocity field that transports a noisy current state toward its clean successor, reducing long-term rollout drift while enabling uncertainty-aware ensemble generations. Alongside this core algorithm, we introduce a Physics-Pretrained Variational Autoencoder (P2VAE) to embed physical states into a compact latent space, and an efficient Flow Marching Transformer (FMT) that combines a diffusion-forcing scheme with latent temporal pyramids, achieving up to 15x greater computational efficiency than full-length video diffusion models and thereby enabling large-scale pretraining at substantially reduced cost. We curate a corpus of ~2.5M trajectories across 12 distinct PDE families and train suites of P2VAEs and FMTs at multiple scales. On downstream evaluation, we benchmark on unseen Kolmogorov turbulence with few-shot adaptation, demonstrate long-term rollout stability over deterministic counterparts, and present uncertainty-stratified ensemble results, highlighting the importance of generative PDE foundation models for real-world applications.

We propose VC-Inspector, a lightweight, open-source large multimodal model (LMM) for reference-free evaluation of video captions, with a focus on factual accuracy. Unlike existing metrics that suffer from limited context handling, weak factuality assessment, or reliance on proprietary services, VC-Inspector offers a reproducible and fact-aware alternative that aligns closely with human judgments. To enable robust training and interpretable evaluation, we introduce a systematic framework for generating captions with controllable factual errors, paired with graded quality scores and explanatory annotations. Experiments demonstrate that VC-Inspector achieves state-of-the-art correlation with human judgments, generalizing across diverse domains (e.g., VATEX-Eval, Flickr8K-Expert, and Flickr8K-CF benchmarks) and revealing the potential for caption improvement. Project page is available at https://dipta007.github.io/VC-Inspector

Fine-tuning the bias terms of large language models (LLMs) has the potential to achieve unprecedented parameter efficiency while maintaining competitive performance, particularly in low-data regimes. However, the link between fine-tuning different bias terms (i.e., $\boldsymbol{b}_q$, $\boldsymbol{b}_k$, and $\boldsymbol{b}_v$ in the query, key, or value projections) and downstream performance remains largely unclear to date. In this paper, we investigate the link between fine-tuning $\boldsymbol{b}_q$, $\boldsymbol{b}_k$, and $\boldsymbol{b}_v$ with the performance of the downstream task. Our key finding is that directly fine-tuning $\boldsymbol{b}_v$ generally leads to higher downstream performance in low-data regimes, in comparison to $\boldsymbol{b}_q$ and $\boldsymbol{b}_k$. We extensively evaluate this unique property across a wide range of LLMs spanning encoder-only and decoder-only architectures up to 6.7B parameters (including bias-free LLMs). Our results provide strong evidence for the effectiveness of directly fine-tuning $\boldsymbol{b}_v$ across various downstream tasks. The implementation code is available at https://github.com/whubaichuan/BEFT.

Assessing the impact the training data on machine learning models is crucial for understanding the behavior of the model, enhancing the transparency, and selecting training data. Influence function provides a theoretical framework for quantifying the effect of training data points on model's performance given a specific test data. However, the computational and memory costs of influence function presents significant challenges, especially for large-scale models, even when using approximation methods, since the gradients involved in computation are as large as the model itself. In this work, we introduce a novel approach that leverages dropout as a gradient compression mechanism to compute the influence function more efficiently. Our method significantly reduces computational and memory overhead, not only during the influence function computation but also in gradient compression process. Through theoretical analysis and empirical validation, we demonstrate that our method could preserves critical components of the data influence and enables its application to modern large-scale models.

Neural network (NN) training is inherently a large-scale matrix optimization problem, yet the matrix structure of NN parameters has long been overlooked. Recently, the optimizer Muon \citep{jordanmuon}, which explicitly exploits this structure, has gained significant attention for its strong performance in foundation model training. A key component contributing to Muon's success is matrix orthogonalization. In this paper, we propose \textit{low-rank orthogonalization}, which performs orthogonalization by leveraging the low-rank nature of gradients during NN training. Building on this, we introduce low-rank matrix-signed gradient descent (MSGD) and a low-rank variant of Muon. Numerical experiments demonstrate the superior performance of low-rank orthogonalization, with low-rank Muon achieving promising results in GPT-2 and LLaMA pretraining -- surpassing the carefully tuned vanilla Muon on tasks with large model sizes. Theoretically, we establish the iteration complexity of low-rank MSGD for finding an approximate stationary solution, and the iteration complexity of low-rank Muon for finding an approximate stochastic stationary solution under heavy-tailed noise. The code to reproduce our numerical experiments is available at https://github.com/dengzhanwang/Low-rank-Muon.

Reservoir is an efficient network for time series processing. It is well known that network structure is one of the determinants of its performance. However, the topology structure of reservoirs, as well as their performance, is hard to analyzed, due to the lack of suitable mathematical tools. In this paper, we study the topology structure of reservoirs using persistent GLMY homology theory, and develop a method to improve its performance. Specifically, it is found that the reservoir performance is closely related to the one-dimensional GLMY homology groups. Then, we develop a reservoir structure optimization method by modifying the minimal representative cycles of one-dimensional GLMY homology groups. Finally, by experiments, it is validated that the performance of reservoirs is jointly influenced by the reservoir structure and the periodicity of the dataset.

Geographic reasoning is a fundamental cognitive capability that requires models to infer plausible locations by synthesizing visual evidence with spatial world knowledge. Despite recent advances in large vision-language models (LVLMs), existing evaluation paradigms remain largely outcome-centric, relying on static datasets and predefined labels that are conceptually misaligned with open-world geographic inference. Such outcome-centric evaluations often focus exclusively on label matching, leaving the underlying linguistic reasoning chains as unexamined black boxes. In this work, we introduce GeoArena, a dynamic, human-preference-based evaluation framework for benchmarking open-world geographic reasoning. GeoArena reframes evaluation as a pairwise reasoning alignment task on in-the-wild images, where human judges compare model-generated explanations based on reasoning quality, evidence synthesis, and plausibility. We deploy GeoArena as a public platform and benchmark 17 frontier LVLMs using thousands of human judgments, which complements existing benchmarks and supports the development of geographically grounded, human-aligned AI systems. We further provide detailed analyses of model behavior, including reliability of human preferences and factors influencing judgments of geographic reasoning quality.

Probabilistic programming offers a powerful framework for modeling uncertainty, yet statistical model discovery in this domain entails navigating an immense search space under strict domain-specific constraints. When small language models are tasked with generating probabilistic programs, they frequently produce outputs that suffer from both syntactic and semantic errors, such as flawed inference constructs. Motivated by probabilistic programmers' domain expertise and debugging strategies, we introduce RefineStat, a language model--driven framework that enforces semantic constraints ensuring synthesized programs contain valid distributions and well-formed parameters, and then applies diagnostic-aware refinement by resampling prior or likelihood components whenever reliability checks fail. We evaluate RefineStat on multiple probabilistic-programming code-generation tasks using smaller language models (SLMs) and find that it produces programs that are both syntactically sound and statistically reliable, often matching or surpassing those from closed-source large language models (e.g., OpenAI o3).

The pursuit of autonomous agents capable of temporally coherent planning is hindered by a fundamental flaw in current vision-language models (VLMs): they lack cognitive inertia. Operating on isolated snapshots, these models cannot form a continuous understanding of the environment, leading to erratic decision jitter and a failure to execute complex, multi-step maneuvers. To remedy this, we introduce CogDriver, a framework designed to build a stable internal representation by instilling this crucial cognitive property. Our work makes two key contributions: (1) We present CogDriver-Data, a large-scale vision-language-action dataset whose narrative annotations provide the supervisory signal for learning temporal dynamics and persistent intent. (2) We develop the CogDriver-Agent, an architecture featuring a sparse temporal memory to maintain a stable internal state. This is enabled by a spatiotemporal knowledge distillation approach that explicitly teaches decision coherence. Comprehensive experiments validate our paradigm: CogDriver-Agent achieves a 22% increase in the closed-loop Driving Score on Bench2Drive and a 21% reduction in mean L2 error on nuScenes, establishing a new state-of-the-art. These significant gains in both long-term decision-making and imitation accuracy provide strong evidence that our agent successfully maintains a temporally coherent internal state, bridging the gap toward more reliable autonomous driving. Project link: https://ocean-luna.github.io/CogDriver.github.io/.

Fine-tuning large-scale pre-trained models with limited data presents significant challenges for generalization. While Sharpness-Aware Minimization (SAM) has proven effective in improving generalization by seeking flat minima, its substantial extra memory and computation overhead make it impractical for large models. Integrating SAM with parameter-efficient fine-tuning methods like Low-Rank Adaptation (LoRA) is a promising direction. However, we find that directly applying SAM to LoRA parameters limits the sharpness optimization to a restricted subspace, hindering its effectiveness. To address this limitation, we propose Bi-directional Low-Rank Adaptation (Bi-LoRA), which introduces an auxiliary LoRA module to model SAM's adversarial weight perturbations. It decouples SAM's weight perturbations from LoRA optimization: the primary LoRA module adapts to specific tasks via standard gradient descent, while the auxiliary module captures the sharpness of the loss landscape through gradient ascent. Such dual-module design enables Bi-LoRA to capture broader sharpness for achieving flatter minima while remaining memory-efficient. Another important benefit is that the dual design allows for simultaneous optimization and perturbation, eliminating SAM's doubled training costs. Extensive experiments across diverse tasks and architectures demonstrate Bi-LoRA's efficiency and effectiveness in enhancing generalization.

Verbal multiword expressions (VMWEs) remain difficult for machine translation because their meanings are often not recoverable from their component words. In this study, we analyze the impact of three VMWE categories -- verbal idioms, verb-particle constructions, and light verb constructions -- on machine translation quality from English to multiple languages. Using both established multiword expression datasets and standard machine translation datasets, we evaluate how state-of-the-art translation systems handle these expressions. Our experimental results consistently show that VMWEs negatively affect translation quality, with deeper analysis indicating that this degradation is primarily attributable to the VMWE itself rather than general sentence-level difficulty. We release our code and evaluation framework to test new MT systems for the community.

Real-world image super-resolution (Real-ISR) focuses on recovering high-quality images from low-resolution inputs that suffer from complex degradations like noise, blur, and compression. Recently, diffusion models (DMs) have shown great potential in this area by leveraging strong generative priors to restore fine details. However, their iterative denoising process incurs high computational overhead, posing challenges for real-time applications. Although one-step distillation methods, such as OSEDiff and TSD-SR, offer faster inference, they remain fundamentally constrained by their large, over-parameterized model architectures. In this work, we present TinySR, a compact yet effective diffusion model specifically designed for Real-ISR that achieves real-time performance while maintaining perceptual quality. We introduce a Dynamic Inter-block Activation and an Expansion-Corrosion Strategy to facilitate more effective decision-making in depth pruning. We achieve VAE compression through channel pruning, attention removal and lightweight SepConv. We eliminate time- and prompt-related modules and perform pre-caching techniques to further speed up the model. TinySR significantly reduces computational cost and model size, achieving up to 5.68x speedup and 83% parameter reduction compared to its teacher TSD-SR, while still providing high quality results.

Retrieving visual and textual information from medical literature and hospital records can enhance diagnostic accuracy for clinical image interpretation. However, multimodal retrieval-augmented diagnosis is highly challenging. We explore a lightweight mechanism for enhancing diagnostic performance of retrieval-augmented LVLMs. We train a lightweight LVLM-aware multimodal retriever, such that the retriever learns to return images and texts that guide the LVLM toward correct predictions. In our low-resource setting, we perform only lightweight fine-tuning with small amounts of data, and use only general-purpose backbone models, achieving competitive results in clinical classification and VQA tasks compared to medically pre-trained models with extensive training. In a novel analysis, we highlight a previously unexplored class of errors that we term inconsistent retrieval predictions: cases where different top-retrieved images yield different predictions for the same target. We find that these cases are challenging for all models, even for non-retrieval models, and that our retrieval optimization mechanism significantly improves these cases over standard RAG. However, our analysis also sheds light on gaps in the ability of LVLMs to utilize retrieved information for clinical predictions. Code and models available at: https://github.com/Nirmaz/CLARE.

Entities in discourse vary in salience: main participants, objects and locations stay prominent, while others are quickly forgotten, raising questions about how humans signal and infer discourse-level salience. Using a graded operationalization of discourse-level salience based on summary-worthiness in multiple summaries, this paper investigates whether predictors of utterance-level prominence extend to the discourse level, and how they interact across 24 spoken and written genres of English. We examine features including grammatical function, definiteness, entity type, linear order, discourse relations and hierarchy, and referential structure, as well as the impact of genre. Our results show that utterance-level predictors significantly correlate with discourse-level salience, but interact with and are modulated by entity-level factors such as frequency and dispersion across the document. Multifactorial models reveal that no single factor determines salience; rather, discourse-structural and semantic features prove more robust than morphosyntactic ones, with substantial variation by genre and communicative intent.

Modern large language models (LLMs) are typically trained and deployed using structured role tags (e.g. system, user, assistant, tool) that explicitly mark the source of each piece of context. While these tags are essential for instruction following and controllability, asymmetries in the training data associated with different role tags can potentially introduce inductive biases. In this paper, we study this phenomenon by formalizing user-assistant bias, defined as the tendency of an LLM to preferentially rely on information from either the user or assistant role when they provide incompatible information about the same entity in the context history. We introduce a task-agnostic benchmark UserAssist and evaluate such bias in 52 frontier models. We observe that most of the instruction-tuned models exhibit strong user bias, whereas base and reasoning models are close to neutral. Using controlled fine-tuning experiments, we isolate which post-training recipes drive the observed user-assistant bias. We find that human-preference alignment amplifies user bias, while reasoning fine-tuning reduces it. Finally, we show that user-assistant bias can be bidirectionally controlled via direct preference optimization (DPO) on UserAssist-train, and that the resulting bias reliably generalizes to two realistic multi-turn debate datasets spanning philosophical opinions and natural argumentative exchanges on factual/policy topics. These results reveal an underexplored consequence of role-tagged training and provide a principled framework to diagnose and control tag-induced biases in modern LLMs.

While multi-step diffusion models have advanced both forward and inverse rendering, existing approaches often treat these problems independently, leading to cycle inconsistency and slow inference speed. In this work, we present Ouroboros, a framework composed of two single-step diffusion models that handle forward and inverse rendering with mutual reinforcement. Our approach extends intrinsic decomposition to both indoor and outdoor scenes and introduces a cycle consistency mechanism that ensures coherence between forward and inverse rendering outputs. Experimental results demonstrate state-of-the-art performance across diverse scenes while achieving substantially faster inference speed compared to other diffusion-based methods. We also demonstrate that Ouroboros can transfer to video decomposition in a training-free manner, reducing temporal inconsistency in video sequences while maintaining high-quality per-frame inverse rendering.

Large language models (LLMs) perform well on step-by-step reasoning benchmarks such as mathematics and code generation, yet their ability to carry out robust long-horizon planning under realistic constraints remains insufficiently evaluated. Existing planning benchmarks often rely on abstract domains or interactive feedback, obscuring end-to-end planning failures and feasibility errors. We introduce HeroBench, a benchmark for evaluating long-horizon, hierarchical planning and structured reasoning in a complex RPG-inspired virtual world. Tasks require models to select numerically feasible equipment, reason over multi-level crafting and resource dependencies, and execute hundreds to thousands of actions as a single end-to-end plan. HeroBench integrates symbolic planning, numeric combat simulation, spatial reasoning, and resource management, while supporting scalable difficulty and adversarial distractors. HeroBench evaluates executable plans through simulation, enabling both success-based and fine-grained progress metrics, as well as detailed failure mode analysis. An evaluation of 25 state-of-the-art LLMs reveals large performance disparities rarely observed in conventional reasoning benchmarks. While reasoning models perform substantially better, no model reliably solves the hardest tasks, highlighting persistent challenges in long-horizon autonomous planning.