Instruction following is critical for large language models, yet real-world instructions often involve multiple constraints with logical structures, such as parallel composition, sequential dependencies, and conditional branching. Existing methods typically construct data by simply combining constraints and aggregate rewards by averaging individual constraint scores during training, overlooking logical dependencies and introducing noisy signals. We propose LsrIF, a training framework for logic-structured instruction following. LsrIF constructs data by organizing atomic constraints into parallel, sequential, conditional, and nested structures, and applies structure-aware reward aggregation aligned with their execution semantics: averaging rewards for parallel constraints, decaying later rewards after early failures in sequential structures, and rewarding only active branches in conditional structures. Experiments show that LsrIF improves instruction following in both in-domain and out-of-domain settings while also benefiting logic reasoning. Further analysis indicates that logic-structured training increases attention to constraint-related tokens and logical connectors, suggesting improved modeling of instruction logic. We will release our data and code for future research.

Machine-Generated Text (MGT) is becoming increasingly difficult to distinguish from Human-Written Text (HWT). This trend has exacerbated malicious activities such as fake news and online fraud. The generalization ability of fine-tuned detectors relies heavily on dataset quality, and simply expanding the sources of MGT may become increasingly insufficient. Further augmentation of the generation process is required. Based on HC-Var's theory, enhancing the human-like alignment of MGT not only facilitates robustness testing of existing detectors but also boosts the generalization ability of detectors fine-tuned on such aligned MGT datasets. Therefore, we propose the \textbf{M}achine-\textbf{A}ugment-\textbf{G}enerated Text via \textbf{A}lignment (MAGA) Detection Benchmark. MAGA integrates several alignment methods, ranging from prompt construction to \textbf{G}enerator-\textbf{D}etector \textbf{A}dversarial \textbf{R}einforcement \textbf{L}earning (GDARL) and the reasoning process. In our experiments, the RoBERTa detector fine-tuned on MAGA achieves an average improvement of 4.60\% in generalization AUC. Conversely, the aligned MGTs in MAGA also lead to an average decrease of 8.13\% in the AUC of selected detectors. We hope the MAGA Benchmark will provide valuable insights for future research on the generalization ability of MGT detectors.

As LLMs gain persuasive capabilities through extended dialogues, they create new opportunities for studying adversarial conversational behavior in extended interaction settings that traditional single-turn safety evaluations fail to capture. We systematically study these interactional dynamics using a controlled LLM-to-LLM simulation framework for automated red-teaming across bilingual social engineering scenarios. Evaluating eight state-of-the-art models in English and Chinese, we analyze dialogue-level outcomes, annotate attacker and defender strategy families, and model interaction dynamics between them. Results show that multi-turn adversarial dialogues follow recurrent escalation patterns, while defensive responses frequently rely on verification, delay, and channel control. We further find statistically significant cross-model and cross-lingual differences in outcome distributions, and transition analysis reveals systematic structural variation in how defender strategies respond to attacker tactics across languages. These findings highlight the importance of studying interactional structure in multi-turn adversarial dialogue settings and demonstrate how controlled LLM-to-LLM simulations can support mechanistic analysis of adversarial conversational dynamics.

Neural networks for time series forecasting have relied on error metrics and architecture-specific interpretability approaches for model selection that don't apply across models of different families. To interpret forecasting models agnostic to the types of layers across state-of-the-art model families, we introduce Horizon Activation Mapping (HAM), a visual interpretability technique inspired by grad-CAM that uses gradient norm averages to study the horizon's subseries where grad-CAM studies attention maps over image data. We introduce causal and anti-causal modes to calculate gradient update norm averages across subseries at every timestep and lines of proportionality signifying uniform distributions of the norm averages. Optimization landscape studies with respect to changes in batch sizes, early stopping, train-val-test splits, architectural choices, univariate forecasting and dropouts are studied with respect to performances and subseries in HAM. Interestingly, batch size based differences in activities seem to indicate potential for existence of an exponential approximation across them per epoch relative to each other. Multivariate forecasting models including MLP-based CycleNet, N-Linear, N-HITS, self attention-based FEDformer, Pyraformer, SSM-based SpaceTime and diffusion-based Multi-Resolution DDPM over different horizon sizes trained over the ETTm2 dataset are used for HAM plots in this study. NHITS' neural approximation theorem and SpaceTime's exponential autoregressive activities have been attributed to trends in HAM plots over their training, validation and test sets. In general, HAM can be used for granular model selection, validation set choices and comparisons across different neural network model families.

Qualitative data are widespread in domains such as healthcare, marketing, and bioinformatics, where clustering offers a fundamental tool for pattern discovery. A core difficulty of qualitative-data clustering lies in measuring similarity among attribute values that carry no inherent ordering or distance. To recover such relationships, existing studies typically rely on within-dataset co-occurrence statistics. This statistical route, however, becomes unreliable once the sample size is small, and the semantic context of each value is therefore left underexploited. Motivated by this limitation, this paper proposes BREVE (Balanced Representation via External Value Enrichment), a clustering framework that enriches each qualitative value with extra semantic dimensions drawn from an external knowledge base. That is, every unique value is expanded by a dense embedding that encodes its semantic content. To prevent the original value identity from being diluted by the added dimensions, a lightweight one-hot component is further appended. An adaptive weight, guided by cluster compactness, then determines how strongly the enrichment dimensions enter the final representation. With this design, experiments on eight benchmark datasets yield an average ARI rank of 1.3 against seven representative competitors.

Large language models (LLMs) achieve strong question answering (QA) performance but can produce fluent answers unsupported by available evidence. Existing hallucination detectors often rely on external verification, repeated sampling, or test-time judge calls, which can be costly for real-time QA. We propose \textbf{HaluNet}, a lightweight hallucination risk estimator that uses internal signals from one model generation. HaluNet jointly models token likelihood, predictive entropy, and hidden-state information, allowing probabilistic, distributional, and semantic evidence to inform an answer-level risk score. It is trained with LLM-as-a-Judge labels as scalable weak supervision and evaluated with independent human and multi-judge assessments. Experiments on SQuAD, TriviaQA, and Natural Questions show that HaluNet improves answer-level risk ranking across in-domain and out-of-domain settings. On a 300-example human evaluation, HaluNet achieves 0.874 AUROC and 0.869 AUPRC; its top 20\% highest-risk answers contain 96.5\% errors, yielding a 2.06$\times$ lift over the base error rate.

Modern surveillance systems increasingly rely on multi-wavelength sensors and deep neural networks to recognize faces in infrared images captured at night. However, most facial recognition models are trained on visible light datasets, leading to substantial performance degradation on infrared inputs due to significant domain shifts. Early feature-based methods for infrared face recognition proved ineffective, prompting researchers to adopt generative approaches that convert infrared images into visible light images for improved recognition. This paradigm, known as Heterogeneous Face Recognition (HFR), faces challenges such as model and modality discrepancies, leading to distortion and feature loss in generated images. To address these limitations, this paper introduces a novel latent diffusion-based model designed to generate high-quality visible face images from thermal inputs while preserving critical identity features. A multi-attribute classifier is incorporated to extract key facial attributes from visible images, mitigating feature loss during infrared-to-visible image restoration. Additionally, we propose the Self-attn Mamba module, which enhances global modeling of cross-modal features and significantly improves inference speed. Experimental results on two benchmark datasets demonstrate the superiority of our approach, achieving state-of-the-art performance in both image quality and identity preservation.

Humans understand long and complex texts by relying on a holistic semantic representation of the content. This global view helps organize prior knowledge, interpret new information, and integrate evidence dispersed across a document, as revealed by the Mindscape-Aware Capability of humans in psychology. Current Retrieval-Augmented Generation (RAG) systems lack such guidance and therefore struggle with long-context tasks. In this paper, we propose Mindscape-Aware RAG (MiA-RAG), the first framework to formulate mindscape-aware retrieval and generation as a unified conditioning paradigm for LLM-based RAG. MiA-RAG builds a mindscape through hierarchical summarization and conditions both retrieval and generation on this global semantic representation. This enables the retriever to form enriched query embeddings and the generator to reason over retrieved evidence within a coherent global context. We evaluate MiA-RAG across diverse long-context and bilingual benchmarks for evidence-based understanding and global sense-making. It consistently surpasses baselines, and further analysis shows that it aligns local details with a coherent global representation, enabling more human-like long-context retrieval and reasoning.

Large Language Model (LLM) agents are increasingly deployed in environments that generate massive, dynamic contexts. However, a critical bottleneck remains: while agents have access to this context, their static prompts lack the mechanisms to manage it effectively, leading to recurring Corrective and Enhancement failures. To address this capability gap, we introduce Self-evolving Context Optimization via Prompt Evolution (SCOPE). SCOPE frames context management as an \textit{online optimization} problem, synthesizing guidelines from execution traces to automatically evolve the agent's prompt. We propose a Dual-Stream mechanism that routes guidelines between tactical memory (immediate error correction) and strategic memory, which is continuously refined through conflict resolution, subsumption pruning, and consolidation. To maximize strategy coverage, Perspective-Driven Exploration evolves multiple parallel prompts guided by distinct optimization perspectives. Experiments on the HLE benchmark show that SCOPE improves task success rates from 14.23\% to 38.64\% without human intervention. We make our code publicly available at https://github.com/JarvisPei/SCOPE.

LiDAR-captured point clouds are often considered the gold standard in active 3D reconstruction. While their accuracy is exceptional in flat regions, the capturing is susceptible to miss small geometric structures and may fail with dark, absorbent materials. Alternatively, capturing multiple photos of the scene and applying 3D photogrammetry can infer these details as they often represent feature-rich regions. However, the accuracy of LiDAR for featureless regions is rarely reached. Therefore, we suggest combining the strengths of LiDAR and camera-based capture by introducing SurfFill: a Gaussian surfel-based LiDAR completion scheme. We analyze LiDAR capturings and attribute LiDAR beam divergence as a main factor for artifacts, manifesting mostly at thin structures and edges. We use this insight to introduce an ambiguity heuristic for completed scans by evaluating the change in density in the point cloud. This allows us to identify points close to missed areas, which we can then use to grow additional points from to complete the scan. For this point growing, we constrain Gaussian surfel reconstruction to focus optimization and densification on these ambiguous areas. Finally, Gaussian primitives of the reconstruction in ambiguous areas are extracted and sampled for points to complete the point cloud. To address the challenges of large-scale reconstruction, we extend this pipeline with a divide-and-conquer scheme for building-sized point cloud completion. We evaluate on the task of LiDAR point cloud completion of synthetic and real-world scenes and find that our method outperforms previous reconstruction methods.

Text-guided image-to-video generation has made substantial progress, yet it still struggles to execute text-specified edits that require substantial changes to a reference image (\textit{e.g., object addition, removal, or modification}). Empirically, our analysis reveals that this stems from \textbf{visual dominance}, where the reference image causes severe attention dispersion, inhibiting the model's ability to incorporate new semantic information. To address this, we propose \textbf{AlignVid}, a training-free intervention that re-calibrates the model's internal attention distribution. Drawing on an energy-based perspective of attention, AlignVid employs Attention Scaling Modulation (\textbf{ASM}) to reduce attention entropy and concentrate focus on semantic tokens, alongside Guidance Scheduling (\textbf{GS}) to maintain generation stability. To rigorously assess this capability, we present \textbf{OmitI2V}, a comprehensive benchmark for evaluating prompt adherence across object modification, addition, and deletion. Extensive experiments demonstrate that AlignVid effectively enhances semantic fidelity with negligible computational overhead. Code and the OmitI2V benchmark are available at https://github.com/LAW1223/AlignVid.

Foundation models have revolutionized various fields such as natural language processing (NLP) and computer vision (CV). While efforts have been made to transfer the success of the foundation models in general AI domains to biology, existing works focus on directly adopting the existing foundation model architectures from general machine learning domains without a systematic design considering the unique physicochemical and structural properties of each biological data modality. This leads to suboptimal performance, as these repurposed architectures struggle to capture the long-range dependencies, sparse information, and complex underlying ``grammars'' inherent to biological data. To address this gap, we introduce BioArc, a novel framework designed to move beyond intuition-driven architecture design towards principled, automated architecture discovery for biological foundation models. Leveraging Neural Architecture Search (NAS), BioArc systematically explores a vast architecture design space, evaluating architectures across multiple biological modalities while rigorously analyzing the interplay between architecture, tokenization, and training strategies. This large-scale analysis identifies novel, high-performance architectures, allowing us to distill a set of empirical design principles to guide future model development. Furthermore, to make the best of this set of discovered principled architectures, we propose and compare several architecture prediction methods that effectively and efficiently predict optimal architectures for new biological tasks. Overall, our work provides a foundational resource and a principled methodology to guide the creation of the next generation of task-specific and foundation models for biology.

Data analysis has become an indispensable part of scientific research. To discover the latent knowledge and insights hidden within massive datasets, we need to perform deep exploratory analysis to realize their full value. With the advent of large language models (LLMs) and multi-agent systems, more and more researchers are making use of these technologies for insight discovery. However, there are few benchmarks for evaluating insight discovery capabilities. As one of the most comprehensive existing frameworks, InsightBench also suffers from many critical flaws: format inconsistencies, poorly conceived objectives, and redundant insights. These issues may significantly affect the quality of data and the evaluation of agents. To address these issues, we thoroughly investigate shortcomings in InsightBench and propose essential criteria for a high-quality insight benchmark. Regarding this, we develop a data-curation pipeline to construct a new dataset named InsightEval. We further introduce a novel metric to measure the exploratory performance of agents. Through extensive experiments on InsightEval, we highlight prevailing challenges in automated insight discovery and raise some key findings to guide future research in this promising direction.

We present ReflexGrad, a dual-process architecture for within-episode failure recovery in LLM agents without demonstrations. When agents commit to a wrong approach early and exhaust the step budget, the post-failure trajectory contains the information to escape -- but no published architecture acts on it within a single episode. ReflexGrad routes between a fast process (TextGrad-style continuous refinement every $k{=}3$ steps) and a slow process (Reflexion-style causal diagnosis when $m{=}5$ consecutive low-progress scores fire a routing gate). A deterministic priority merge keeps the natural-language policy coherent, and each slow activation emits three observable artifacts: a reproducible trigger, a causal diagnostic, and a verified fix. On ALFWorld 134 tasks, $n{=}10$ seeds, no demonstrations, ReflexGrad lifts Qwen-3-8B from $35.1\%$ to $75.4\%$ ($+40.3$pp), beating compute-matched 1-shot LATS by $+2.7$pp ($p{\approx}0.01$), ToT by $+5.7$pp ($p{<}10^{-4}$), and Self-Refine by $+6.7$pp ($p{<}10^{-5}$); on GPT-5 the lift is $46.3{\to}88.1\%$ ($+41.8$pp). The $1.5$pp cross-model difference is within seed noise ($p{\approx}0.13$), suggesting that the routing mechanism, rather than model scale, is the primary source of the gain. Code, prompts, per-seed logs, and sensitivity sweeps are released.

Reinforcement learning (RL) in 3D environments with high-dimensional sensory input poses two major challenges: (1) the high memory consumption induced by memory buffers required to stabilise learning, and (2) the complexity of learning in partially observable Markov Decision Processes (POMDPs). This project addresses these challenges by proposing two novel input representations: SS-only and RGB+SS, both employing semantic segmentation on RGB colour images. Experiments were conducted in deathmatches of ViZDoom, utilizing perfect segmentation results for controlled evaluation. Our results showed that SS-only was able to reduce the memory consumption of memory buffers by at least 66.6%, and up to 98.6% when a vectorisable lossless compression technique with minimal overhead such as run-length encoding is applied. Meanwhile, RGB+SS significantly enhances RL agents' performance with the additional semantic information provided. Furthermore, we explored density-based heatmapping as a tool to visualise RL agents' movement patterns and evaluate their suitability for data collection. A brief comparison with a previous approach highlights how our method overcame common pitfalls in applying semantic segmentation in 3D environments like ViZDoom.

Blocking communication presents a major hurdle in running MoEs efficiently in distributed settings. To address this, we present FarSkip-Collective which modifies the architecture of modern models to enable overlapping of their computation with communication. Our approach modifies the architecture to skip connections in the model and it is unclear a priori whether the modified model architecture can remain as capable, especially for large state-of-the-art models and while modifying all of the model layers. We answer this question in the affirmative and fully convert a series of state-of-the-art models varying from 16B to 109B parameters to enable overlapping of their communication while achieving accuracy that is comparable with their original open-source releases. For example, we convert Llama 4 Scout (109B) via self-distillation and achieve average accuracy within 1% of its instruction tuned release averaged over a wide range of downstream evaluations. In addition to demonstrating retained accuracy of the large modified models, we realize the benefits of FarSkip-Collective through optimized implementations that explicitly overlap communication with computation, accelerating both training and inference in existing frameworks. For inference, we demonstrate 32.6% speedup in Time To First Token when serving a converted DeepSeek-V3 architecture with expert parallelism in SGLang and achieve 97.3% communication-computation overlap during the prefill stage. During training, our approach enables 88.9% communication overlap of the all-to-all communication collectives when pre-training DeepSeek-V3 MoE layers with expert parallelism.

Convolutional Neural Networks (CNNs) pre-trained on large-scale datasets such as ImageNet are widely used as feature extractors to construct high-accuracy classification models from scarce data for specific tasks. In such scenarios, fine-tuning the pre-trained CNN is difficult due to data scarcity, necessitating the use of fixed weights. However, when the weights are kept fixed, many filters that do not contribute to the target task remain in the model, leading to unnecessary redundancy and reduced efficiency. Therefore, effective methods are needed to reduce model size by pruning filters that are unnecessary for inference. To address this, approaches utilizing Layer-wise Relevance Propagation (LRP) have been proposed. LRP quantifies the contribution of each filter to the inference result, enabling the pruning of filters with low relevance. However, existing LRP-based pruning methods have been observed to cause cascading accuracy degradation. In this study, we introduce an accuracy-aware pruning control mechanism for existing LRP-based filter pruning methods, which suppresses cascading accuracy degradation by dynamically adjusting the pruning rate and the pruning order using the harmonic mean of class accuracy, and compresses the pre-trained model while preserving task-specific performance in a small-data environment. We demonstrate that this control mechanism effectively mitigates cascading accuracy degradation and achieves higher classification accuracy compared to existing LRP-based pruning methods, improving the class-averaged area under the accuracy-pruning-rate curve (AUC) of VGG16 by approximately 15\% over conventional LRP-based approaches.

Unlearning in large language models (LLMs) is critical for regulatory compliance and for building ethical generative AI systems that avoid producing private, toxic, illegal, or copyrighted content. Despite rapid progress, in this work, we show that \textit{almost all} existing unlearning methods fail to achieve true forgetting in practice. Specifically, while evaluations of these `unlearned' models under deterministic (greedy) decoding often suggest successful knowledge removal using standard benchmarks, we show that sensitive information reliably resurfaces when models are sampled with standard probabilistic decoding. To rigorously capture this vulnerability, we introduce \texttt{leak@$k$}, a new meta-evaluation metric that quantifies the likelihood of forgotten knowledge reappearing when generating $k$ samples from the model under realistic decoding strategies. Using three widely adopted benchmarks, TOFU, MUSE, and WMDP, we conduct the first large-scale, systematic study of unlearning reliability using \texttt{leak@$k$} metric. Our findings demonstrate that knowledge leakage persists across methods and tasks, underscoring that current state-of-the-art (SOTA) unlearning techniques provide only limited forgetting. We propose an algorithm, termed Robust Unlearning under LEak@$k$ metric (\texttt{RULE}) to address this concern. We demonstrate that \texttt{RULE} provides an unlearned model for TOFU benchmark with no information leakage for a large number of generation samples. On the MUSE benchmark, \texttt{RULE} outperforms SOTA unlearning methods under the \texttt{leak@$k$} metric across most sampling budgets $k$. Codes are available at https://github.com/OptimAI-Lab/Leak-k.

Augmenting vision-language-action models (VLAs) with world models is promising for robotic policy learning but faces challenges in jointly predicting states and actions due to the modality gap. To address this, we propose DUal-STream diffusion (DUST), a world-model augmented VLA framework featuring a multimodal diffusion transformer that maintains separate modality streams while enabling cross-modal knowledge sharing. In addition, DUST utilizes independent noise perturbations and a decoupled flow matching loss to learn cross-modal causal relationships. We further introduce an asynchronous sampling method for action and vision tokens that enhances performance through inference-time scaling. Experimental results on simulated benchmarks like RoboCasa and GR-1 show that DUST achieves up to 6% gains over state-of-the-art VLA and world-modeling baselines, with inference-time scaling providing an additional 2-5% improvement. In real-world tasks using the Franka Research 3, DUST outperforms baselines by 10% in success rate. Finally, we demonstrate that DUST enables effective transfer learning through both pretraining on action-free videos and joint-training with heterogeneous robot and human datasets.

Stochastic state estimation methods for continuum robots (CRs) often struggle to balance accuracy and computational efficiency. While several recent works have explored sliding-window formulations for CRs, these methods are limited to simplified, discrete-time approximations and do not provide stochastic representations. In contrast, current stochastic filter methods must run at the speed of measurements, limiting their full potential. Recent works in continuous-time estimation techniques for CRs show a principled approach to addressing this runtime constraint, but are currently restricted to offline operation. In this work, we present a sliding-window filter (SWF) for continuous-time state estimation of CRs that improves upon the accuracy of a filter approach while enabling continuous-time methods to operate online, all while running at faster-than-real-time speeds. This represents the first stochastic SWF specifically designed for CRs, providing a promising direction for future research in this area.

Recent advances in text-to-video generation have achieved impressive performance on short clips, yet evaluating long-form generation under complex textual inputs remains a significant challenge. In response to this challenge, we present LoCoT2V-Bench, a benchmark for long video generation (LVG) featuring multi-scene prompts with hierarchical metadata (e.g., character settings and camera behaviors), constructed from collected real-world videos. We further propose LoCoT2V-Eval, a multi-dimensional framework covering perceptual quality, text-video alignment, temporal quality, dynamic quality, and Human Expectation Realization Degree (HERD), with an emphasis on aspects such as fine-grained text-video alignment and temporal character consistency. Experiments on 17 representative LVG models reveal pronounced capability disparities across evaluation dimensions, with strong perceptual quality and background consistency but markedly weaker fine-grained text-video alignment and character consistency. These findings suggest that improving prompt faithfulness and identity preservation remains a key challenge for long-form video generation. Our code and data are released at https://github.com/XqZeppelinhead0702/LoCoT2V-Bench

The quadratic cost of attention limits the scalability of long-context LLMs, especially under limited hardware memory budgets. While attention is often sparse, existing static sparse methods cannot adapt to task- or input-dependent variations, and recent dynamic approaches rely on predefined templates or heuristics that may sacrifice generality. We propose Dynamic Hierarchical Sparse Attention (DHSA), a data-driven framework that predicts attention sparsity online while keeping the LLM backbone frozen. DHSA performs hierarchical routing by estimating importance at the chunk level and propagating it to token-level interactions, preserving causally important dependencies while enabling efficient sparsification. Across Needle-in-a-Haystack test, LongBench and RULER, DHSA maintains near-dense accuracy in highly sparse regimes, achieving 12--20% relative accuracy gains over Block Sparse Attention at comparable prefill cost. With a memory-efficient tiled backend, DHSA delivers up to $10\times$ prefill speedup at 128K context length. On LLaMA-3.1-8B (4-bit), DHSA scales to 100K context on a single 24GB GPU, where dense attention fails. We provide complementary GPU and CPU backends, enabling DHSA to run across diverse hardware environments and multiple open-weight model families. These results demonstrate DHSA as an efficient and adaptable solution for memory-constrained long-context LLM inference.

The development of large-scale artificial intelligence (AI) models is influencing neuroscience research by enabling end-to-end learning from raw brain signals and neural data. In this paper, we review applications of large-scale AI models across four major neuroscience domains: neuroimaging and data processing, brain-computer interfaces and neural decoding, clinical decision support and translational frameworks, and disease-specific applications across neurological and psychiatric disorders. These models show potential to address major computational neuroscience challenges, including multimodal neural data integration, spatiotemporal pattern interpretation, and the development of translational frameworks for clinical research. Moreover, the interaction between neuroscience and AI has become increasingly reciprocal, as biologically informed architectural constraints are now incorporated to develop more interpretable and computationally efficient models. This review highlights both the promise of such technologies and critical implementation considerations, with particular emphasis on rigorous evaluation frameworks, effective integration of domain knowledge, prospective clinical validation, and comprehensive ethical guidelines. Finally, a systematic listing of critical neuroscience datasets used to develop and evaluate large-scale AI models across diverse research applications is provided.

The recent development of foundation models for time series data has generated considerable interest in using such models across a variety of applications. Although foundation models achieve state-of-the-art predictive performance, their calibration properties remain relatively underexplored, despite the fact that calibration can be critical for many practical applications. In this paper, we investigate the calibration-related properties of five recent time series foundation models and two competitive baselines. We perform a series of systematic evaluations assessing model calibration (i.e., over- or under-confidence), effects of varying prediction heads, and calibration under long-term autoregressive forecasting. We find that time series foundation models are consistently better calibrated than baseline models and tend not to be either systematically over- or under-confident, in contrast to the overconfidence often seen in other deep learning models.

This work investigates the resilience of contemporary large language models (LLMs) against frequent character-level perturbations. We examine three types of character-level perturbations including introducing numerous typos within words, shuffling the characters in each word, and inserting a large number of invisible characters into the text. Surprisingly, even under severe perturbation, such as shuffling nearly all words character-wise to produce text that is almost unreadable to humans, or inserting invisible characters which are several times more than the visible ones as noise, many LLMs still maintain notable performance. We explore the underlying causes of this robustness and find that LLMs exhibit remarkable resilience to chaotic segmentation and fragmented tokenization. Furthermore, we examine the mechanisms by which LLMs remove perturbations to correctly comprehend text, including both implicit and explicit mechanisms for character-level perturbation. We hope that our findings on the low-level robustness of LLMs will unveil their inherent architectural strengths, reveal the potential risks of their misuse, and inform the reliable deployment of LLMs across diverse application scenarios.

As language models become increasingly deployed in online environments, toxicity detection and detoxification have received growing attention. Existing studies primarily focus on non-obfuscated text, which limits robustness when users intentionally disguise toxic expressions. In particular, Korean toxic expressions can be easily disguised through agglutinative morphology and Hangeul-specific orthographic variation. However, obfuscation in Korean remains largely unexplored, which motivates us to introduce a KOTOX: Korean toxic dataset for deobfuscation and detoxification. We categorize Korean obfuscation patterns into linguistically grounded classes, define transformation rules derived from real-world examples, and provide the resulting obfuscation framework as an open transformation package. Using these rules, we provide paired neutral and toxic sentences alongside their obfuscated counterparts. Models trained on our dataset better handle obfuscated text without sacrificing performance on non-obfuscated text. This is the first dataset that simultaneously supports deobfuscation and detoxification for the Korean language. We expect the dataset to facilitate better understanding and mitigation of obfuscated toxic content in LLM for Korean. Our code and data are available at https://github.com/leeyejin1231/KOTOX.

A key component of in-context reasoning is the ability of language models (LMs) to bind entities for later retrieval. For example, an LM might represent "Ann loves pie" by binding "Ann" to "pie", allowing it to later retrieve "Ann" when asked "Who loves pie?" Prior research on short lists of bound entities found strong evidence that LMs implement such retrieval via a positional mechanism, where "Ann" is retrieved based on its position in context. In this work, we find that this mechanism generalizes poorly to more complex settings; as the number of bound entities in context increases, the positional mechanism becomes noisy and unreliable in middle positions. To compensate for this, we find that LMs supplement the positional mechanism with a lexical mechanism (retrieving "Ann" using its bound counterpart "pie") and a reflexive mechanism (retrieving "Ann" through a direct pointer). Through extensive experiments on nine models and ten binding tasks, we uncover a consistent pattern in how LMs mix these mechanisms to drive model behavior. We leverage these insights to develop a causal model combining all three mechanisms that estimates next token distributions with 95% agreement. Finally, we show that our model generalizes to substantially longer inputs of open-ended text interleaved with entity groups, further demonstrating the robustness of our findings in more natural settings. Overall, our study establishes a more complete picture of how LMs bind and retrieve entities in-context.

Modern enterprises generate vast streams of time series metrics when monitoring complex systems, known as observability data. Unlike conventional time series from domains such as climate, observability data are zero-inflated, highly stochastic, and exhibit minimal temporal structure. Despite their importance, observability datasets remain underrepresented in public benchmarks due to proprietary restrictions and privacy concerns. Existing datasets are often anonymized and normalized, removing scale information and limiting their use for tasks such as anomaly detection, root cause analysis, and multi-modal reasoning. To address this gap, we introduce TelecomTS, a large-scale observability dataset derived from a 5G telecommunications network. TelecomTS features heterogeneous, de-anonymized covariates with explicit absolute scale information and provides a diverse suite of downstream tasks, including anomaly detection, root cause analysis, and multi-modal question-answering. Benchmarking state-of-the-art time series, language, reasoning, and multi-modal foundation models reveals that existing approaches struggle with the abrupt, noisy, and high-variance dynamics characteristic of observability data. Our experiments further underscore the importance of preserving covariates' absolute scale, emphasizing the need for foundation time series models that natively leverage scale information for practical real-world observability applications. The code is available at: https://github.com/Ali-maatouk/TelecomTS.

In this work, we establish the sufficient conditions under which nonlinear Canonical Correlation Analysis (CCA) recovers ground-truth latent factors up to an affine transformation. By transporting the analysis from the observation space to the source space, we extend classical statistical results on orthogonal polynomial expansions of bivariate distributions to representation learning, proving affine identifiability under specific distributional priors. We formally demonstrate that whitening is strictly necessary to ensure the boundedness and well-conditioning of the learned mappings. Furthermore, we bridge the gap between theory and practice by proving that ridge-regularized empirical CCA converges to its population counterpart in the finite-sample regime. Finally, our findings provide a rigorous theoretical foundation explaining the empirical success of recent correlation-based non-contrastive learning methods. Experiments on synthetic and rendered image datasets, alongside systematic ablations, validate the predicted recovery behavior and illustrate the failure modes that arise when the assumptions are violated.

We propose a distributional framework for offline Inverse Reinforcement Learning (IRL) that jointly models uncertainty over reward functions and full distributions of returns. Unlike conventional IRL approaches that recover a deterministic reward estimate or match only expected returns, our method captures richer structure in expert behavior, particularly in learning the reward distribution, by minimizing first-order stochastic dominance (FSD) violations and thus integrating distortion risk measures (DRMs) into policy learning, enabling the recovery of both reward distributions and distribution-aware policies. This formulation is well-suited for behavior analysis and risk-aware imitation learning. Theoretical analysis shows that the algorithm converges with $\mathcal{O}(\varepsilon^{-2})$ iteration complexity. Empirical results on synthetic benchmarks, real-world neurobehavioral data, and MuJoCo control tasks demonstrate that our method recovers expressive reward representations and achieves state-of-the-art performance.