This work presents VTok, a unified video tokenization framework that can be used for both generation and understanding tasks. Unlike the leading vision-language systems that tokenize videos through a naive frame-sampling strategy, we propose to decouple the spatial and temporal representations of videos by retaining the spatial features of a single key frame while encoding each subsequent frame into a single residual token, achieving compact yet expressive video tokenization. Our experiments suggest that VTok effectively reduces the complexity of video representation from the product of frame count and per-frame token count to their sum, while the residual tokens sufficiently capture viewpoint and motion changes relative to the key frame. Extensive evaluations demonstrate the efficacy and efficiency of VTok: it achieves notably higher performance on a range of video understanding and text-to-video generation benchmarks compared with baselines using naive tokenization, all with shorter token sequences per video (e.g., 3.4% higher accuracy on our TV-Align benchmark and 1.9% higher VBench score). Remarkably, VTok produces more coherent motion and stronger guidance following in text-to-video generation, owing to its more consistent temporal encoding. We hope VTok can serve as a standardized video tokenization paradigm for future research in video understanding and generation.

Reconstructing high-dimensional spatiotemporal fields from sparse point-sensor measurements is a central challenge in learning parametric PDE dynamics. Existing approaches often struggle to generalize across trajectories and parameter settings, or rely on discretization-tied decoders that do not naturally transfer across meshes and resolutions. We propose STRIDE (Spatio-Temporal Recurrent Implicit DEcoder), a two-stage framework that maps a short window of sensor measurements to a latent state with a temporal encoder and reconstructs the field at arbitrary query locations with a modulated implicit neural representation (INR) decoder. Using the Fourier Multi-Component and Multi-Layer Neural Network (FMMNN) as the INR backbone improves representation of complex spatial fields and yields more stable optimization than sine-based INRs. We provide a conditional theoretical justification: under stable delay observability of point measurements on a low-dimensional parametric invariant set, the reconstruction operator factors through a finite-dimensional embedding, making STRIDE-type architectures natural approximators. Experiments on four challenging benchmarks spanning chaotic dynamics and wave propagation show that STRIDE outperforms strong baselines under extremely sparse sensing, supports super-resolution, and remains robust to noise.

The enhanced capabilities of LLM-based agents come with an emergency for model planning and tool-use abilities. Attributing to helpful-harmless trade-off from LLM alignment, agents typically also inherit the flaw of "over-refusal", which is a passive failure mode. However, the proactive planning and action capabilities of agents introduce another crucial danger on the other side of the trade-off. This phenomenon we term "Toxic Proactivity'': an active failure mode in which an agent, driven by the optimization for Machiavellian helpfulness, disregards ethical constraints to maximize utility. Unlike over-refusal, Toxic Proactivity manifests as the agent taking excessive or manipulative measures to ensure its "usefulness'' is maintained. Existing research pays little attention to identifying this behavior, as it often lacks the subtle context required for such strategies to unfold. To reveal this risk, we introduce a novel evaluation framework based on dilemma-driven interactions between dual models, enabling the simulation and analysis of agent behavior over multi-step behavioral trajectories. Through extensive experiments with mainstream LLMs, we demonstrate that Toxic Proactivity is a widespread behavioral phenomenon and reveal two major tendencies. We further present a systematic benchmark for evaluating Toxic Proactive behavior across contextual settings.

Safety risks of AI models have been widely studied at deployment time, such as jailbreak attacks that elicit harmful outputs. In contrast, safety risks emerging during training remain largely unexplored. Beyond explicit reward hacking that directly manipulates explicit reward functions in reinforcement learning, we study implicit training-time safety risks: harmful behaviors driven by a model's internal incentives and contextual background information. For example, during code-based reinforcement learning, a model may covertly manipulate logged accuracy for self-preservation. We present the first systematic study of this problem, introducing a taxonomy with five risk levels, ten fine-grained risk categories, and three incentive types. Extensive experiments reveal the prevalence and severity of these risks: notably, Llama-3.1-8B-Instruct exhibits risky behaviors in 74.4% of training runs when provided only with background information. We further analyze factors influencing these behaviors and demonstrate that implicit training-time risks also arise in multi-agent training settings. Our results identify an overlooked yet urgent safety challenge in training.

While deep learning-based super-resolution (SR) methods have shown impressive outcomes with synthetic degradation scenarios such as bicubic downsampling, they frequently struggle to perform well on real-world images that feature complex, nonlinear degradations like noise, blur, and compression artifacts. Recent efforts to address this issue have involved the painstaking compilation of real low-resolution (LR) and high-resolution (HR) image pairs, usually limited to several specific downscaling factors. To address these challenges, our work introduces a novel framework capable of synthesizing authentic LR images from a single HR image by leveraging the latent degradation space with flow matching. Our approach generates LR images with realistic artifacts at unseen degradation levels, which facilitates the creation of large-scale, real-world SR training datasets. Comprehensive quantitative and qualitative assessments verify that our synthetic LR images accurately replicate real-world degradations. Furthermore, both traditional and arbitrary-scale SR models trained using our datasets consistently yield much better HR outcomes.

Instruction-grounded driving, where passenger language guides trajectory planning, requires vehicles to understand intent before motion. However, most prior instruction-following planners rely on simulation or fixed command vocabularies, limiting real-world generalization. doScenes, the first real-world dataset linking free-form instructions (with referentiality) to nuScenes ground-truth motion, enables instruction-conditioned planning. In this work, we adapt OpenEMMA, an open-source MLLM-based end-to-end driving framework that ingests front-camera views and ego-state and outputs 10-step speed-curvature trajectories, to this setting, presenting a reproducible instruction-conditioned baseline on doScenes and investigate the effects of human instruction prompts on predicted driving behavior. We integrate doScenes directives as passenger-style prompts within OpenEMMA's vision-language interface, enabling linguistic conditioning before trajectory generation. Evaluated on 849 annotated scenes using ADE, we observe that instruction conditioning substantially improves robustness by preventing extreme baseline failures, yielding a 98.7% reduction in mean ADE. When such outliers are removed, instructions still influence trajectory alignment, with well-phrased prompts improving ADE by up to 5.1%. We use this analysis to discuss what makes a "good" instruction for the OpenEMMA framework. We release the evaluation prompts and scripts to establish a reproducible baseline for instruction-aware planning. GitHub: https://github.com/Mi3-Lab/doScenes-VLM-Planning

Event-based Action Recognition (EAR) has attracted significant attention due to the high temporal resolution and high dynamic range of event cameras. However, existing methods typically suffer from (i) the computational redundancy of dense voxel representations, (ii) structural redundancy inherent in multi-branch architectures, and (iii) the under-utilization of spectral information in capturing global motion patterns. To address these challenges, we propose an efficient EAR framework named HoloEv-Net. First, to simultaneously tackle representation and structural redundancies, we introduce a Compact Holographic Spatiotemporal Representation (CHSR). Departing from computationally expensive voxel grids, CHSR implicitly embeds horizontal spatial cues into the Time-Height (T-H) view, effectively preserving 3D spatiotemporal contexts within a 2D representation. Second, to exploit the neglected spectral cues, we design a Global Spectral Gating (GSG) module. By leveraging the Fast Fourier Transform (FFT) for global token mixing in the frequency domain, GSG enhances the representation capability with negligible parameter overhead. Extensive experiments demonstrate the scalability and effectiveness of our framework. Specifically, HoloEv-Net-Base achieves state-of-the-art performance on THU-EACT-50-CHL, HARDVS and DailyDVS-200, outperforming existing methods by 10.29%, 1.71% and 6.25%, respectively. Furthermore, our lightweight variant, HoloEv-Net-Small, delivers highly competitive accuracy while offering extreme efficiency, reducing parameters by 5.4 times, FLOPs by 300times, and latency by 2.4times compared to heavy baselines, demonstrating its potential for edge deployment.

Existing industrial-scale navigation applications contend with massive road networks, typically employing two main categories of approaches for route planning. The first relies on precomputed road costs for optimal routing and heuristic algorithms for generating alternatives, while the second, generative methods, has recently gained significant attention. However, the former struggles with personalization and route diversity, while the latter fails to meet the efficiency requirements of large-scale real-time scenarios. To address these limitations, we propose GenMRP, a generative framework for multi-route planning. To ensure generation efficiency, GenMRP first introduces a skeleton-to-capillary approach that dynamically constructs a relevant sub-network significantly smaller than the full road network. Within this sub-network, routes are generated iteratively. The first iteration identifies the optimal route, while the subsequent ones generate alternatives that balance quality and diversity using the newly proposed correctional boosting approach. Each iteration incorporates road features, user historical sequences, and previously generated routes into a Link Cost Model to update road costs, followed by route generation using the Dijkstra algorithm. Extensive experiments show that GenMRP achieves state-of-the-art performance with high efficiency in both offline and online environments. To facilitate further research, we have publicly released the training and evaluation dataset. GenMRP has been fully deployed in a real-world navigation app, demonstrating its effectiveness and benefits.

This paper introduces Point2Insert, a sparse-point-based framework for flexible and user-friendly object insertion in videos, motivated by the growing popularity of accurate, low-effort object placement. Existing approaches face two major challenges: mask-based insertion methods require labor-intensive mask annotations, while instruction-based methods struggle to place objects at precise locations. Point2Insert addresses these issues by requiring only a small number of sparse points instead of dense masks, eliminating the need for tedious mask drawing. Specifically, it supports both positive and negative points to indicate regions that are suitable or unsuitable for insertion, enabling fine-grained spatial control over object locations. The training of Point2Insert consists of two stages. In Stage 1, we train an insertion model that generates objects in given regions conditioned on either sparse-point prompts or a binary mask. In Stage 2, we further train the model on paired videos synthesized by an object removal model, adapting it to video insertion. Moreover, motivated by the higher insertion success rate of mask-guided editing, we leverage a mask-guided insertion model as a teacher to distill reliable insertion behavior into the point-guided model. Extensive experiments demonstrate that Point2Insert consistently outperforms strong baselines and even surpasses models with $\times$10 more parameters.

Static sparse training is a promising route to efficient learning by committing to a fixed mask pattern, yet the constrained structure reduces robustness. Early pruning decisions can lock the network into a brittle structure that is difficult to escape, especially in deep reinforcement learning (RL) where the evolving policy continually shifts the training distribution. We propose Topology-Aware Revival (TAR), a lightweight one-shot post-pruning procedure that improves static sparsity without dynamic rewiring. After static pruning, TAR performs a single revival step by allocating a small reserve budget across layers according to topology needs, randomly uniformly reactivating a few previously pruned connections within each layer, and then keeping the resulting connectivity fixed for the remainder of training. Across multiple continuous-control tasks with SAC and TD3, TAR improves final return over static sparse baselines by up to +37.9% and also outperforms dynamic sparse training baselines with a median gain of +13.5%.

Situated embodied conversation requires robots to interleave real-time dialogue with active perception: deciding what to look at, when to look, and what to say under tight latency constraints. We present a simple, minimal system recipe that pairs a real-time multimodal language model with a small set of tool interfaces for attention and active perception. We study six home-style scenarios that require frequent attention shifts and increasing perceptual scope. Across four system variants, we evaluate turn-level tool-decision correctness against human annotations and collect subjective ratings of interaction quality. Results indicate that real-time multimodal large language models and tool use for active perception is a promising direction for practical situated embodied conversation.

Segmentation architectures are typically benchmarked on single imaging modalities, obscuring deployment-relevant performance variations: an architecture optimal for one modality may underperform on another. We present a cross-modal evaluation framework for materials image segmentation spanning SEM, AFM, XCT, and optical microscopy. Our evaluation of six encoder-decoder combinations across seven datasets reveals that optimal architectures vary systematically by context: UNet excels for high-contrast 2D imaging while DeepLabv3+ is preferred for the hardest cases. The framework also provides deployment feedback via out-of-distribution detection and counterfactual explanations that reveal which microstructural features drive predictions. Together, the architecture guidance, reliability signals, and interpretability tools address a practical gap in materials characterization, where researchers lack tools to select architectures for their specific imaging setup or assess when models can be trusted on new samples.

Multivariate time series forecasting in graph-structured domains is critical for real-world applications, yet existing spatiotemporal models often suffer from performance degradation under data scarcity and cross-domain shifts. We address these challenges through the lens of structure-aware context selection. We propose TL-GPSTGN, a transfer-oriented spatiotemporal framework that enhances sample efficiency and out-of-distribution generalization by selectively pruning non-optimized graph context. Specifically, our method employs information-theoretic and correlation-based criteria to extract structurally informative subgraphs and features, resulting in a compact, semantically grounded representation. This optimized context is subsequently integrated into a spatiotemporal convolutional architecture to capture complex multivariate dynamics. Evaluations on large-scale traffic benchmarks demonstrate that TL-GPSTGN consistently outperforms baselines in low-data transfer scenarios. Our findings suggest that explicit context pruning serves as a powerful inductive bias for improving the robustness of graph-based forecasting models.

Current 3D scene graph generation (3DSGG) approaches heavily rely on a single-agent assumption and small-scale environments, exhibiting limited scalability to real-world scenarios. In this work, we introduce Multi-Agent 3D Scene Graph Generation (MA3DSG) model, the first framework designed to tackle this scalability challenge using multiple agents. We develop a training-free graph alignment algorithm that efficiently merges partial query graphs from individual agents into a unified global scene graph. Leveraging extensive analysis and empirical insights, our approach enables conventional single-agent systems to operate collaboratively without requiring any learnable parameters. To rigorously evaluate 3DSGG performance, we propose MA3DSG-Bench-a benchmark that supports diverse agent configurations, domain sizes, and environmental conditions-providing a more general and extensible evaluation framework. This work lays a solid foundation for scalable, multi-agent 3DSGG research.

Data incompleteness severely impedes the reliability of multimodal systems. Existing reconstruction methods face distinct bottlenecks: conventional parametric/generative models are prone to hallucinations due to over-reliance on internal memory, while retrieval-augmented frameworks struggle with retrieval rigidity. Critically, these end-to-end architectures are fundamentally constrained by Semantic-Detail Entanglement -- a structural conflict between logical reasoning and signal synthesis that compromises fidelity. In this paper, we present \textbf{\underline{O}}mni-\textbf{\underline{M}}odality \textbf{\underline{G}}eneration Agent (\textbf{OMG-Agent}), a novel framework that shifts the paradigm from static mapping to a dynamic coarse-to-fine Agentic Workflow. By mimicking a \textit{deliberate-then-act} cognitive process, OMG-Agent explicitly decouples the task into three synergistic stages: (1) an MLLM-driven Semantic Planner that resolves input ambiguity via Progressive Contextual Reasoning, creating a deterministic structured semantic plan; (2) a non-parametric Evidence Retriever that grounds abstract semantics in external knowledge; and (3) a Retrieval-Injected Executor that utilizes retrieved evidence as flexible feature prompts to overcome rigidity and synthesize high-fidelity details. Extensive experiments on multiple benchmarks demonstrate that OMG-Agent consistently surpasses state-of-the-art methods, maintaining robustness under extreme missingness, e.g., a $2.6$-point gain on CMU-MOSI at $70$\% missing rates.

As robots increasingly become part of shared human spaces, their movements must transcend basic functionality by incorporating expressive qualities to enhance engagement and communication. This paper introduces a movement-centered design pedagogy designed to support engineers in creating expressive robotic arm movements. Through a hands-on interactive workshop informed by interdisciplinary methodologies, participants explored various creative possibilities, generating valuable insights into expressive motion design. The iterative approach proposed integrates analytical frameworks from dance, enabling designers to examine motion through dynamic and embodied dimensions. A custom manual remote controller facilitates interactive, real-time manipulation of the robotic arm, while dedicated animation software supports visualization, detailed motion sequencing, and precise parameter control. Qualitative analysis of this interactive design process reveals that the proposed "toolbox" effectively bridges the gap between human intent and robotic expressiveness resulting in more intuitive and engaging expressive robotic arm movements.

Distributional reinforcement learning (RL) is a powerful framework increasingly adopted in safety-critical domains for its ability to optimize risk-sensitive objectives. However, the role of the discount factor is often overlooked, as it is typically treated as a fixed parameter of the Markov decision process or tunable hyperparameter, with little consideration of its effect on the learned policy. In the literature, it is well-known that the discounting function plays a major role in characterizing time preferences of an agent, which an exponential discount factor cannot fully capture. Building on this insight, we propose a novel framework that supports flexible discounting of future rewards and optimization of risk measures in distributional RL. We provide a technical analysis of the optimality of our algorithms, show that our multi-horizon extension fixes issues raised with existing methodologies, and validate the robustness of our methods through extensive experiments. Our results highlight that discounting is a cornerstone in decision-making problems for capturing more expressive temporal and risk preferences profiles, with potential implications for real-world safety-critical applications.

Heterogeneous multi-robot systems are increasingly deployed in long-horizon missions that require coordination among robots with diverse capabilities. However, existing planning approaches struggle to construct accurate symbolic representations and maintain plan consistency in dynamic environments. Classical PDDL planners require manually crafted symbolic models, while LLM-based planners often ignore agent heterogeneity and environmental uncertainty. We introduce KGLAMP, a knowledge-graph-guided LLM planning framework for heterogeneous multi-robot teams. The framework maintains a structured knowledge graph encoding object relations, spatial reachability, and robot capabilities, which guides the LLM in generating accurate PDDL problem specifications. The knowledge graph serves as a persistent, dynamically updated memory that incorporates new observations and triggers replanning upon detecting inconsistencies, enabling symbolic plans to adapt to evolving world states. Experiments on the MAT-THOR benchmark show that KGLAMP improves performance by at least 25.5% over both LLM-only and PDDL-based variants.

Light verb constructions (LVCs) are a challenging class of verbal multiword expressions, especially in Turkish, where rich morphology and productive complex predicates create minimal contrasts between idiomatic predicate meanings and literal verb--argument uses. This paper asks what signals drive LVC classification by systematically restricting model inputs. Using UD-derived supervision, we compare lemma-driven baselines (lemma TF--IDF + Logistic Regression; BERTurk trained on lemma sequences), a grammar-only Logistic Regression over UD morphosyntax (UPOS/DEPREL/MORPH), and a full-input BERTurk baseline. We evaluate on a controlled diagnostic set with Random negatives, lexical controls (NLVC), and LVC positives, reporting split-wise performance to expose decision-boundary behavior. Results show that coarse morphosyntax alone is insufficient for robust LVC detection under controlled contrasts, while lexical identity supports LVC judgments but is sensitive to calibration and normalization choices. Overall, Our findings motivate targeted evaluation of Turkish MWEs and show that ``lemma-only'' is not a single, well-defined representation, but one that depends critically on how normalization is operationalized.

Recent research has shown that language models can learn to \textit{reason}, often via reinforcement learning. Some work even trains low-rank parameterizations for reasoning, but conventional LoRA cannot scale below the model dimension. We question whether even rank=1 LoRA is necessary for learning to reason and propose TinyLoRA, a method for scaling low-rank adapters to sizes as small as one parameter. Within our new parameterization, we are able to train the 8B parameter size of Qwen2.5 to 91\% accuracy on GSM8K with only 13 trained parameters in bf16 (26 total bytes). We find this trend holds in general: we are able to recover 90\% of performance improvements while training $1000x$ fewer parameters across a suite of more difficult learning-to-reason benchmarks such as AIME, AMC, and MATH500. Notably, we are only able to achieve such strong performance with RL: models trained using SFT require $100-1000x$ larger updates to reach the same performance.

Graph Foundation Models (GFMs) have achieved remarkable success in generalizing across diverse domains. However, they mainly focus on Text-Attributed Graphs (TAGs), leaving Multimodal-Attributed Graphs (MAGs) largely untapped. Developing Multimodal Graph Foundation Models (MGFMs) allows for leveraging the rich multimodal information in MAGs, and extends applicability to broader types of downstream tasks. While recent MGFMs integrate diverse modality information, our empirical investigation reveals two fundamental limitations of existing MGFMs: (1)they fail to explicitly model modality interaction, essential for capturing intricate cross-modal semantics beyond simple aggregation, and (2)they exhibit sub-optimal modality alignment, which is critical for bridging the significant semantic disparity between distinct modal spaces. To address these challenges, we propose PLANET (graPh topoLogy-aware modAlity iNteraction and alignmEnT), a novel framework employing a Divide-and-Conquer strategy to decouple modality interaction and alignment across distinct granularities. At the embedding granularity, (1)Embedding-wise Domain Gating (EDG) performs local semantic enrichment by adaptively infusing topology-aware cross-modal context, achieving modality interaction. At the node granularity, (2)Node-wise Discretization Retrieval (NDR) ensures global modality alignment by constructing a Discretized Semantic Representation Space (DSRS) to bridge modality gaps. Extensive experiments demonstrate that PLANET significantly outperforms state-of-the-art baselines across diverse graph-centric and multimodal generative tasks.

The equations of complex dynamical systems may not be identified by expert knowledge, especially if the underlying mechanisms are unknown. Data-driven discovery methods address this challenge by inferring governing equations from time-series data using a library of functions constructed from the measured variables. However, these methods typically assume time-invariant coefficients, which limits their ability to capture evolving system dynamics. To overcome this limitation, we allow some of the parameters to vary over time, learn their temporal evolution directly from data, and infer a system of equations that incorporates both constant and time-varying parameters. We then transform this framework into a forecasting model by predicting the time-varying parameters and substituting these predictions into the learned equations. The model is validated using datasets for Susceptible-Infected-Recovered, Consumer--Resource, greenhouse gas concentration, and Cyanobacteria cell count. By dynamically adapting to temporal shifts, our proposed model achieved a mean absolute error below 3\% for learning a time series and below 6\% for forecasting up to a month ahead. We additionally compare forecasting performance against CNN-LSTM and Gradient Boosting Machine (GBM), and show that our model outperforms these methods across most datasets. Our findings demonstrate that integrating time-varying parameters into data-driven discovery of differential equations improves both modeling accuracy and forecasting performance.

Psychological counseling is a fundamentally multimodal cognitive process in which clinicians integrate verbal content with visual and vocal cues to infer clients' mental states and respond empathically. However, most existing language-model-based counseling systems operate on text alone and rely on implicit mental state inference. We introduce DELTA, a deliberative multi-agent framework that models counseling as a structured reasoning process over multimodal signals, separating evidence grounding, mental state abstraction, and response generation. DELTA further incorporates reinforcement learning guided by a distribution-level Emotion Attunement Score to encourage emotionally attuned responses. Experiments on a multimodal counseling benchmark show that DELTA improves both counseling quality and emotion attunement across models. Ablation and qualitative analyses suggest that explicit multimodal reasoning and structured mental state representations play complementary roles in supporting empathic human-AI interaction.

Semi-dual neural optimal transport learns a transport map via a max-min objective, yet training can converge to incorrect or degenerate maps. We fully characterize these spurious solutions in the common regime where data concentrate on low-dimensional manifold: the objective is underconstrained off the data manifold, while the on-manifold transport signal remains identifiable. Following Choi, Choi, and Kwon (2025), we study additive-noise smoothing as a remedy and prove new map recovery guarantees as the noise vanishes. Our main practical contribution is a computable terminal noise level $\varepsilon_{\mathrm{stat}}(N)$ that attains the optimal statistical rate, with scaling governed by the intrinsic dimension $m$ of the data. The formula arises from a theoretical unified analysis of (i) quantitative stability of optimal plans, (ii) smoothing-induced bias, and (iii) finite-sample error, yielding rates that depend on $m$ rather than the ambient dimension. Finally, we show that the reduced semi-dual objective becomes increasingly ill-conditioned as $\varepsilon \downarrow 0$. This provides a principled stopping rule: annealing below $\varepsilon_{\mathrm{stat}}(N)$ can $\textit{worsen}$ optimization conditioning without improving statistical accuracy.

In this work, we focus on boosting the feature extraction to improve the performance of Structure-from-Motion (SfM) in endoscopy videos. We present SuperPoint-E, a new local feature extraction method that, using our proposed Tracking Adaptation supervision strategy, significantly improves the quality of feature detection and description in endoscopy. Extensive experimentation on real endoscopy recordings studies our approach's most suitable configuration and evaluates SuperPoint-E feature quality. The comparison with other baselines also shows that our 3D reconstructions are denser and cover more and longer video segments because our detector fires more densely and our features are more likely to survive (i.e. higher detection precision). In addition, our descriptor is more discriminative, making the guided matching step almost redundant. The presented approach brings significant improvements in the 3D reconstructions obtained, via SfM on endoscopy videos, compared to the original SuperPoint and the gold standard SfM COLMAP pipeline.

Perplexity is a widely adopted metric for assessing the predictive quality of large language models (LLMs) and often serves as a reference metric for downstream evaluations. However, recent evidence shows that perplexity can be unreliable, especially when irrelevant long inputs are used, raising concerns for both benchmarking and system deployment. While prior efforts have employed selective input filtering and curated datasets, the impact of input length on perplexity has not been systematically studied from a systems perspective and input length has rarely been treated as a first-class system variable affecting both fairness and efficiency. In this work, we close this gap by introducing LengthBenchmark, a system-conscious evaluation framework that explicitly integrates input length, evaluation protocol design, and system-level costs, evaluating representative LLMs under two scoring protocols (direct accumulation and fixed window sliding) across varying context lengths. Unlike prior work that focuses solely on accuracy-oriented metrics, LengthBenchmark additionally measures latency, memory footprint, and evaluation cost, thereby linking predictive metrics to deployment realities. We further incorporate quantized variants not as a main contribution, but as robustness checks, showing that length-induced biases persist across both full-precision and compressed models. This design disentangles the effects of evaluation logic, quantization, and input length, and demonstrates that length bias is a general phenomenon that undermines fair cross-model comparison. Our analysis yields two key observations: (i) sliding window evaluation consistently inflates performance on short inputs, and (ii) both full-precision and quantized models appear to realise gains as the evaluated segment length grows.

Large language models (LLMs) achieve strong performance when all task-relevant information is available upfront, as in static prediction and instruction-following problems. However, many real-world decision-making tasks are inherently online: crucial information must be acquired through interaction, feedback is delayed, and effective behavior requires balancing information collection and exploitation over time. While in-context learning enables adaptation without weight updates, existing LLMs often struggle to reliably leverage in-context interaction experience in such settings. In this work, we show that this limitation can be addressed through training. We introduce ORBIT, a multi-task, multi-episode meta-reinforcement learning framework that trains LLMs to learn from interaction in context. After meta-training, a relatively small open-source model (Qwen3-14B) demonstrates substantially improved in-context online learning on entirely unseen environments, matching the performance of GPT-5.2 and outperforming standard RL fine-tuning by a large margin. Scaling experiments further reveal consistent gains with model size, suggesting significant headroom for learn-at-inference-time decision-making agents. Code reproducing the results in the paper can be found at https://github.com/XiaofengLin7/ORBIT.

Music understanding is a complex task that often requires reasoning over both structural and semantic elements of audio. We introduce BASS, designed to evaluate music understanding and reasoning in audio language models across four broad categories: structural segmentation, lyric transcription, musicological analysis, and artist collaboration. BASS comprises 2658 questions spanning 12 tasks, 1993 unique songs and covering over 138 hours of music from a wide range of genres and tracks, crafted to assess musicological knowledge and reasoning in real-world scenarios. We evaluate 14 open-source and frontier multimodal LMs, finding that even state-of-the-art models struggle on higher-level reasoning tasks such as structural segmentation and artist collaboration, while performing best on lyric transcription. Our analysis reveals that current models leverage linguistic priors effectively but remain limited in reasoning over musical structure, vocal, and musicological attributes. BASS provides an evaluation framework with widespread applications in music recommendation and search and has the potential to guide the development of audio LMs.

Deterministic neural operators perform well on many PDEs but can struggle with the approximation of high-frequency wave phenomena, where strong input-to-output sensitivity makes operator learning challenging, and spectral bias blurs oscillations. We argue for adopting a probabilistic approach for approximating waves in high-frequency regime, and develop our probabilistic framework using a score-based conditional diffusion operator. After demonstrating a stability analysis of the Helmholtz operator, we present our numerical experiments across a wide range of frequencies, benchmarked against other popular data-driven and machine learning approaches for waves. We show that our probabilistic neural operator consistently produces robust predictions with the lowest errors in $L^2$, $H^1$, and energy norms. Moreover, unlike all the other tested deterministic approaches, our framework remarkably captures uncertainties in the input sound speed map propagated to the solution field. We envision that our results position probabilistic operator learning as a principled and effective approach for solving complex PDEs such as Helmholtz in the challenging high-frequency regime.

Research has repeatedly demonstrated that intermediate hidden states extracted from large language models and speech audio models predict measured brain response to natural language stimuli. Yet, very little is known about the representation properties that enable this high prediction performance. Why is it the intermediate layers, and not the output layers, that are most effective for this unique and highly general transfer task? We give evidence that the correspondence between speech and language models and the brain derives from shared meaning abstraction and not their next-word prediction properties. In particular, models construct higher-order linguistic features in their middle layers, cued by a peak in the layerwise intrinsic dimension, a measure of feature complexity. We show that a layer's intrinsic dimension strongly predicts how well it explains fMRI and ECoG signals; that the relation between intrinsic dimension and brain predictivity arises over model pre-training; and finetuning models to better predict the brain causally increases both representations' intrinsic dimension and their semantic content. Results suggest that semantic richness, high intrinsic dimension, and brain predictivity mirror each other, and that the key driver of model-brain similarity is rich meaning abstraction of the inputs, where language modeling is a task sufficiently complex (but perhaps not the only) to require it.