We introduce the System Hallucination Scale (SHS), a lightweight and human-centered measurement instrument for assessing hallucination-related behavior in large language models (LLMs). Inspired by established psychometric tools such as the System Usability Scale (SUS) and the System Causability Scale (SCS), SHS enables rapid, interpretable, and domain-agnostic evaluation of factual unreliability, incoherence, misleading presentation, and responsiveness to user guidance in model-generated text. SHS is explicitly not an automatic hallucination detector or benchmark metric; instead, it captures how hallucination phenomena manifest from a user perspective under realistic interaction conditions. A real-world evaluation with 210 participants demonstrates high clarity, coherent response behavior, and construct validity, supported by statistical analysis including internal consistency (Cronbach's alpha = 0.87$) and significant inter-dimension correlations (p < 0.001$). Comparative analysis with SUS and SCS reveals complementary measurement properties, supporting SHS as a practical tool for comparative analysis, iterative system development, and deployment monitoring.

Mechanistic interpretability identifies internal circuits responsible for model behaviors, yet translating these findings into human-understandable explanations remains an open problem. We present a pipeline that bridges circuit-level analysis and natural language explanations by (i) identifying causally important attention heads via activation patching, (ii) generating explanations using both template-based and LLM-based methods, and (iii) evaluating faithfulness using ERASER-style metrics adapted for circuit-level attribution. We evaluate on the Indirect Object Identification (IOI) task in GPT-2 Small (124M parameters), identifying six attention heads accounting for 61.4% of the logit difference. Our circuit-based explanations achieve 100% sufficiency but only 22% comprehensiveness, revealing distributed backup mechanisms. LLM-generated explanations outperform template baselines by 64% on quality metrics. We find no correlation (r = 0.009) between model confidence and explanation faithfulness, and identify three failure categories explaining when explanations diverge from mechanisms.

Feature Transformation (FT) is a core data-centric AI task that improves feature space quality to advance downstream predictive performance. However, discovering effective transformations remains challenging due to the large space of feature-operator combinations. Existing solutions rely on discrete search or latent generation, but they are frequently limited by sample inefficiency, invalid candidates, and redundant generations with limited coverage. Large Language Models (LLMs) offer strong priors for producing valid transformations, but current LLM-based FT methods typically rely on static demonstrations, resulting in limited diversity, redundant outputs, and weak alignment with downstream objectives. We propose a framework that optimizes context data for LLM-driven FT by evolving trajectory-level experiences in a closed loop. Starting from high-performing feature transportation sequences explored by reinforcement learning, we construct and continuously update an experience library of downstream task-verified transformation trajectories, and use a diversity-aware selector to form contexts along with a chain-of-thought and guide transformed feature generation toward higher performance. Experiments on diverse tabular benchmarks show that our method outperforms classical and LLM-based baselines and is more stable than one-shot generation. The framework generalizes across API-based and open-source LLMs and remains robust across downstream evaluators.

Large language models (LLMs) have demonstrated remarkable capabilities across diverse tasks, yet their ability to accurately assess their own confidence remains poorly understood. We present an empirical study investigating whether LLMs exhibit patterns reminiscent of the Dunning-Kruger effect -- a cognitive bias where individuals with limited competence tend to overestimate their abilities. We evaluate four state-of-the-art models (Claude Haiku 4.5, Gemini 2.5 Pro, Gemini 2.5 Flash, and Kimi K2) across four benchmark datasets totaling 24,000 experimental trials. Our results reveal striking calibration differences: Kimi K2 exhibits severe overconfidence with an Expected Calibration Error (ECE) of 0.726 despite only 23.3% accuracy, while Claude Haiku 4.5 achieves the best calibration (ECE = 0.122) with 75.4% accuracy. These findings demonstrate that poorly performing models display markedly higher overconfidence -- a pattern analogous to the Dunning-Kruger effect in human cognition. We discuss implications for safe deployment of LLMs in high-stakes applications.

The digital age has expanded social media and online forums, allowing free expression for nearly 45% of the global population. Yet, it has also fueled online harassment, bullying, and harmful behaviors like hate speech and toxic comments across social networks, messaging apps, and gaming communities. Studies show 65% of parents notice hostile online behavior, and one-third of adolescents in mobile games experience bullying. A substantial volume of abusive content is generated and shared daily, not only on the surface web but also within dark web forums. Creators of abusive comments often employ specific words or coded phrases to evade detection and conceal their intentions. To address these challenges, we propose a hybrid deep learning model that integrates BERT, CNN, and LSTM architectures with a ReLU activation function to detect abusive language across multiple online platforms, including YouTube comments, online forum discussions, and dark web posts. The model demonstrates strong performance on a diverse and imbalanced dataset containing 77,620 abusive and 272,214 non-abusive text samples (ratio 1:3.5), achieving approximately 99% across evaluation metrics such as Precision, Recall, Accuracy, F1-score, and AUC. This approach effectively captures semantic, contextual, and sequential patterns in text, enabling robust detection of abusive content even in highly skewed datasets, as encountered in real-world scenarios.

Mixture-of-Experts (MoE) models enable scalable performance but face severe memory constraints on edge devices. Existing offloading strategies struggle with I/O bottlenecks due to the dynamic, low-information nature of autoregressive expert activation. In this paper, we propose to repurpose Speculative Decoding (SD) not merely as a compute accelerator, but as an informative lookahead sensor for memory management, supported by our theoretical and empirical analyses. Hence, we introduce MoE-SpAc, an MoE inference framework that integrates a Speculative Utility Estimator to track expert demand, a Heterogeneous Workload Balancer to dynamically partition computation via online integer optimization, and an Asynchronous Execution Engine to unify the prefetching and eviction in the same utility space. Extensive experiments on seven benchmarks demonstrate that MoE-SpAc achieves a 42% improvement in TPS over the SOTA SD-based baseline, and an average 4.04x speedup over all standard baselines. Code is available at https://github.com/lshAlgorithm/MoE-SpAc .

Summarization is a core task in Natural Language Processing (NLP). Recent advances in Large Language Models (LLMs) and the introduction of large context windows reaching millions of tokens make it possible to process entire books in a single prompt. At the same time, for well-known books, LLMs can generate summaries based only on internal knowledge acquired during training. This raises several important questions: How do summaries generated from internal memory compare to those derived from the full text? Does prior knowledge influence summaries even when the model is given the book as input? In this work, we conduct an experimental evaluation of book summarization with state-of-the-art LLMs. We compare summaries of well-known books produced using (i) only the internal knowledge of the model and (ii) the full text of the book. The results show that having the full text provides more detailed summaries in general, but some books have better scores for the internal knowledge summaries. This puts into question the capabilities of models to perform summarization of long texts, as information learned during training can outperform summarization of the full text in some cases.

LLM unlearning is essential for mitigating safety, copyright, and privacy concerns in pre-trained large language models (LLMs). Compared to preference alignment, it offers a more explicit way by removing undesirable knowledge characterized by specific unlearning datasets. In previous works, gradient ascent (GA) and its variants have shown promise for implementing unlearning, yet their untargeted nature results in unintended degradation of general capabilities, incomplete removal of knowledge, and the generation of incoherent responses, among many others. We argue that these issues stem from the absence of explicit guidance on what and how models should unlearn. To fill this gap, we introduce a novel unlearning target, reasoning-based unlearning target, which satisfies both the specified unlearning scope and the specified post-unlearning response. Building on this, we propose targeted reasoning unlearning (TRU), which leverages reasoning-based unlearning target as guidance. We employ the target using a cross-entropy supervised loss combined with a GA-based loss, enabling the model to learn reasoning ability for precise knowledge removal while preserving unrelated abilities. We evaluate TRU against strong baselines across multiple benchmarks and LLM backbones, and find that it achieves more reliable unlearning while preserving general capabilities. Moreover, TRU exhibits superior robustness under diverse attack scenarios, stemming from the reasoning ability learned through reasoning-based targets. Overall, our study establishes reasoning-augmented unlearning as a practical paradigm for reliable and explainable LLM unlearning.

A fundamental requirement for intelligent systems is the ability to learn continuously under changing environments. However, models trained in this regime often suffer from catastrophic forgetting. Leveraging pre-trained models has recently emerged as a promising solution, since their generalized feature extractors enable faster and more robust adaptation. While some earlier works mitigate forgetting by fine-tuning only on the first task, this approach quickly deteriorates as the number of tasks grows and the data distributions diverge. More recent research instead seeks to consolidate task knowledge into a unified backbone, or adapting the backbone as new tasks arrive. However, such approaches may create a (potential) \textit{mismatch} between task-specific classifiers and the adapted backbone. To address this issue, we propose a novel \textit{Local Classifier Alignment} (LCA) loss to better align the classifier with backbone. Theoretically, we show that this LCA loss can enable the classifier to not only generalize well for all observed tasks, but also improve robustness. Furthermore, we develop a complete solution for continual learning, following the model merging approach and using LCA. Extensive experiments on several standard benchmarks demonstrate that our method often achieves leading performance, sometimes surpasses the state-of-the-art methods with a large margin.

As embodied models become powerful, humans will collaborate with multiple embodied AI agents at their workplace or home in the future. To ensure better communication between human users and the multi-agent system, it is crucial to interpret incoming information from agents in parallel and refer to the appropriate context for each query. Existing challenges include effectively compressing and communicating high volumes of individual sensory inputs in the form of video and correctly aggregating multiple egocentric videos to construct system-level memory. In this work, we first formally define a novel problem of understanding multiple long-horizon egocentric videos simultaneously collected from embodied agents. To facilitate research in this direction, we introduce MultiAgent-EgoQA (MA-EgoQA), a benchmark designed to systemically evaluate existing models in our scenario. MA-EgoQA provides 1.7k questions unique to multiple egocentric streams, spanning five categories: social interaction, task coordination, theory-of-mind, temporal reasoning, and environmental interaction. We further propose a simple baseline model for MA-EgoQA named EgoMAS, which leverages shared memory across embodied agents and agent-wise dynamic retrieval. Through comprehensive evaluation across diverse baselines and EgoMAS on MA-EgoQA, we find that current approaches are unable to effectively handle multiple egocentric streams, highlighting the need for future advances in system-level understanding across the agents. The code and benchmark are available at https://ma-egoqa.github.io.

AI assistants that support humans in daily life are becoming increasingly feasible, driven by the rapid advancements in multimodal language models. A key challenge lies in overcoming the generic nature of these models to deliver personalized experiences. Existing approaches to personalizing large vision language models often rely on additional training stages, which limit generality and scalability, or on engineered pipelines with external pre-trained modules, which hinder deployment efficiency. In this work, we propose an efficient personalization method that leverages the model's inherent ability to capture personalized concepts. Specifically, we extract visual tokens that predominantly represent the target concept by utilizing the model's internal attention mechanisms. These tokens serve as a memory of that specific concept, enabling the model to recall and describe it when it appears in test images. We conduct a comprehensive and unified evaluation of our approach and SOTA methods across various personalization settings including single-concept, multi-concept, and video personalization, demonstrating strong performance gains with minimal personalization overhead.

Understanding human behavior from complementary egocentric (ego) and exocentric (exo) points of view enables the development of systems that can support workers in industrial environments and enhance their safety. However, progress in this area is hindered by the lack of datasets capturing both views in realistic industrial scenarios. To address this gap, we propose ENIGMA-360, a new ego-exo dataset acquired in a real industrial scenario. The dataset is composed of 180 egocentric and 180 exocentric procedural videos temporally synchronized offering complementary information of the same scene. The 360 videos have been labeled with temporal and spatial annotations, enabling the study of different aspects of human behavior in industrial domain. We provide baseline experiments for 3 foundational tasks for human behavior understanding: 1) Temporal Action Segmentation, 2) Keystep Recognition and 3) Egocentric Human-Object Interaction Detection, showing the limits of state-of-the-art approaches on this challenging scenario. These results highlight the need for new models capable of robust ego-exo understanding in real-world environments. We publicly release the dataset and its annotations at https://fpv-iplab.github.io/ENIGMA-360/.

In this article, we evaluate four Large Language Models (LLMs) and their effectiveness at retrieving data within a specialized Retrieval-Augmented Generation (RAG) system, using a comprehensive food composition database. Our method is focused on the LLMs ability to translate natural language queries into structured metadata filters, enabling efficient retrieval via a Chroma vector database. By achieving high accuracy in this critical retrieval step, we demonstrate that LLMs can serve as an accessible, high-performance tool, drastically reducing the manual effort and technical expertise previously required for domain experts, such as food compilers and nutritionists, to leverage complex food and nutrition data. However, despite the high performance on easy and moderately complex queries, our analysis of difficult questions reveals that reliable retrieval remains challenging when queries involve non-expressible constraints. These findings demonstrate that LLM-driven metadata filtering excels when constraints can be explicitly expressed, but struggles when queries exceed the representational scope of the metadata format.

Visual Question Answering (VQA) is a fundamental multimodal task that requires models to jointly understand visual and textual information. Early VQA systems relied heavily on language biases, motivating subsequent work to emphasize visual grounding and balanced datasets. With the success of large-scale pre-trained transformers for both text and vision domains -- such as PhoBERT for Vietnamese language understanding and Vision Transformers (ViT) for image representation learning -- multimodal fusion has achieved remarkable progress. For Vietnamese VQA, several datasets have been introduced to promote research in low-resource multimodal learning, including ViVQA, OpenViVQA, and the recently proposed ViTextVQA. These resources enable benchmarking of models that integrate linguistic and visual features in the Vietnamese context. Evaluation of VQA systems often employs automatic metrics originally designed for image captioning or machine translation, such as BLEU, METEOR, CIDEr, Recall, Precision, and F1-score. However, recent research suggests that large language models can further improve the alignment between automatic evaluation and human judgment in VQA tasks. In this work, we explore Vietnamese Visual Question Answering using transformer-based architectures, leveraging both textual and visual pre-training while systematically comparing automatic evaluation metrics under multilingual settings.

This research focuses on developing advanced methods for assessing similarity between recipes by combining different sources of information and analytical approaches. We explore the semantic, lexical, and domain similarity of food recipes, evaluated through the analysis of ingredients, preparation methods, and nutritional attributes. A web-based interface was developed to allow domain experts to validate the combined similarity results. After evaluating 318 recipe pairs, experts agreed on 255 (80%). The evaluation of expert assessments enables the estimation of which similarity aspects--lexical, semantic, or nutritional--are most influential in expert decision-making. The application of these methods has broad implications in the food industry and supports the development of personalized diets, nutrition recommendations, and automated recipe generation systems.

Radiology reports capture crucial longitudinal information on tumor burden, treatment response, and disease progression, yet their unstructured narrative format complicates automated analysis. While large language models (LLMs) have advanced clinical text processing, most state-of-the-art systems remain proprietary, limiting their applicability in privacy-sensitive healthcare environments. We present a fully open-source, locally deployable pipeline for longitudinal information extraction from radiology reports, implemented using the llm_extractinator framework. The system applies the qwen2.5-72b model to extract and link target, non-target, and new lesion data across time points in accordance with RECIST criteria. Evaluation on 50 Dutch CT Thorax/Abdomen report pairs yielded high extraction performance, with attribute-level accuracies of 93.7% for target lesions, 94.9% for non-target lesions, and 94.0% for new lesions. The approach demonstrates that open-source LLMs can achieve clinically meaningful performance in multi-timepoint oncology tasks while ensuring data privacy and reproducibility. These results highlight the potential of locally deployable LLMs for scalable extraction of structured longitudinal data from routine clinical text.

Human vision achieves remarkable perceptual performance while operating under strict metabolic constraints. A key ingredient is the selective attention mechanism, driven by rapid saccadic eye movements that constantly reposition the high-resolution fovea onto task-relevant locations, unlike conventional AI systems that process entire images with equal emphasis. Our work aims to draw inspiration from the human visual system to create smarter, more efficient image processing models. Using DINO, a self-supervised Vision Transformer that produces attention maps strikingly similar to human gaze patterns, we explore a saccade inspired method to focus the processing of information on key regions in visual space. To do so, we use the ImageNet dataset in a standard classification task and measure how each successive saccade affects the model's class scores. This selective-processing strategy preserves most of the full-image classification performance and can even outperform it in certain cases. By benchmarking against established saliency models built for human gaze prediction, we demonstrate that DINO provides superior fixation guidance for selecting informative regions. These findings highlight Vision Transformer attention as a promising basis for biologically inspired active vision and open new directions for efficient, neuromorphic visual processing.

Large pretrained diffusion models have significantly enhanced the quality of generated videos, and yet their use in real-time streaming remains limited. Autoregressive models offer a natural framework for sequential frame synthesis but require heavy computation to achieve high fidelity. Diffusion distillation can compress these models into efficient few-step variants, but existing video distillation approaches largely adapt image-specific methods that neglect temporal dependencies. These techniques often excel in image generation but underperform in video synthesis, exhibiting reduced motion coherence, error accumulation over long sequences, and a latency-quality trade-off. We identify two factors that result in these limitations: insufficient utilization of temporal context during step reduction and implicit prediction of subsequent noise levels in next-chunk prediction (i.e., exposure bias). To address these issues, we propose Diagonal Distillation, which operates orthogonally to existing approaches and better exploits temporal information across both video chunks and denoising steps. Central to our approach is an asymmetric generation strategy: more steps early, fewer steps later. This design allows later chunks to inherit rich appearance information from thoroughly processed early chunks, while using partially denoised chunks as conditional inputs for subsequent synthesis. By aligning the implicit prediction of subsequent noise levels during chunk generation with the actual inference conditions, our approach mitigates error propagation and reduces oversaturation in long-range sequences. We further incorporate implicit optical flow modeling to preserve motion quality under strict step constraints. Our method generates a 5-second video in 2.61 seconds (up to 31 FPS), achieving a 277.3x speedup over the undistilled model.

This paper explores potential improvements to the Spatial-Temporal Matching algorithm for aligning the GPS trajectories to road networks. While this algorithm is effective, it presents some limitations in computational efficiency and the accuracy of the results, especially in dense environments with relatively high sampling intervals. To address this, the paper proposes four modifications to the original algorithm: a dynamic buffer, an adaptive observation probability, a redesigned temporal scoring function, and a behavioral analysis to account for the historical mobility patterns. The enhancements are assessed using real-world data from the urban area of Milan, and through newly defined evaluation metrics to be applied in the absence of ground truth. The results of the experiment show significant improvements in performance efficiency and path quality across various metrics.

Offline-to-online reinforcement learning (RL) offers a promising paradigm for robotics by pre-training policies on safe, offline demonstrations and fine-tuning them via online interaction. However, a fundamental challenge remains: how to safely explore online without deviating from the behavioral support of the offline data? While recent methods leverage conditional variational autoencoders (CVAEs) to bound exploration within a latent space, they inherently suffer from an exploitation gap -- a performance ceiling imposed by the decoder's reconstruction loss. We introduce SPAARS, a curriculum learning framework that initially constrains exploration to the low-dimensional latent manifold for sample-efficient, safe behavioral improvement, then seamlessly transfers control to the raw action space, bypassing the decoder bottleneck. SPAARS has two instantiations: the CVAE-based variant requires only unordered (s,a) pairs and no trajectory segmentation; SPAARS-SUPE pairs SPAARS with OPAL temporal skill pretraining for stronger exploration structure at the cost of requiring trajectory chunks. We prove an upper bound on the exploitation gap using the Performance Difference Lemma, establish that latent-space policy gradients achieve provable variance reduction over raw-space exploration, and show that concurrent behavioral cloning during the latent phase directly controls curriculum transition stability. Empirically, SPAARS-SUPE achieves 0.825 normalized return on kitchen-mixed-v0 versus 0.75 for SUPE, with 5x better sample efficiency; standalone SPAARS achieves 92.7 and 102.9 normalized return on hopper-medium-v2 and walker2d-medium-v2 respectively, surpassing IQL baselines of 66.3 and 78.3 respectively, confirming the utility of the unordered-pair CVAE instantiation.

The shift toward user-customized on-device learning places new demands on wireless systems: models must be trained on diverse, distributed data while meeting strict latency, bandwidth, and reliability constraints. To address this, we propose an Agentic AI as the control layer for managing federated learning (FL) over 6G networks, which translates high-level task goals into actions that are aware of network conditions. Rather than simply viewing FL as a learning challenge, our system sees it as a combined task of learning and network management. A set of specialized agents focused on retrieval, planning, coding, and evaluation utilizes monitoring tools and optimization methods to handle client selection, incentive structuring, scheduling, resource allocation, adaptive local training, and code generation. The use of closed-loop evaluation and memory allows the system to consistently refine its decisions, taking into account varying signal-to-noise ratios, bandwidth conditions, and device capabilities. Finally, our case study has demonstrated the effectiveness of the Agentic AI system's use of tools for achieving high performance.

Osteoporosis is a skeletal disease typically diagnosed using dual-energy X-ray absorptiometry (DXA), which quantifies areal bone mineral density but overlooks bone microarchitecture and surrounding soft tissues. High-resolution peripheral quantitative computed tomography (HR-pQCT) enables three-dimensional microstructural imaging with minimal radiation. However, current analysis pipelines largely focus on mineralized bone compartments, leaving much of the acquired image data underutilized. We introduce a fully automated framework for binary osteoporosis classification using radiomics features extracted from anatomically segmented HR-pQCT images. To our knowledge, this work is the first to leverage a transformer-based segmentation architecture, i.e., the SegFormer, for fully automated multi-region HR-pQCT analysis. The SegFormer model simultaneously delineated the cortical and trabecular bone of the tibia and fibula along with surrounding soft tissues and achieved a mean F1 score of 95.36%. Soft tissues were further subdivided into skin, myotendinous, and adipose regions through post-processing. From each region, 939 radiomic features were extracted and dimensionally reduced to train six machine learning classifiers on an independent dataset comprising 20,496 images from 122 HR-pQCT scans. The best image level performance was achieved using myotendinous tissue features, yielding an accuracy of 80.08% and an area under the receiver operating characteristic curve (AUROC) of 0.85, outperforming bone-based models. At the patient level, replacing standard biological, DXA, and HR-pQCT parameters with soft tissue radiomics improved AUROC from 0.792 to 0.875. These findings demonstrate that automated, multi-region HR-pQCT segmentation enables the extraction of clinically informative signals beyond bone alone, highlighting the importance of integrated tissue assessment for osteoporosis detection.

Vision-language pretraining has driven significant progress in medical image analysis. However, current methods typically supervise visual encoders using one-hot labels or free-form text, neither of which effectively captures the complex semantic relationships among clinical findings. In this study, we introduce VIVID-Med, a novel framework that leverages a frozen large language model (LLM) as a structured semantic teacher to pretrain medical vision transformers (ViTs). VIVID-Med translates clinical findings into verifiable JSON field-state pairs via a Unified Medical Schema (UMS), utilizing answerability-aware masking to focus optimization. It then employs Structured Prediction Decomposition (SPD) to partition cross-attention into orthogonality-regularized query groups, extracting complementary visual aspects. Crucially, the LLM is discarded post-training, yielding a lightweight, deployable ViT-only backbone. We evaluated VIVID-Med across multiple settings: on CheXpert linear probing, it achieves a macro-AUC of 0.8588, outperforming BiomedCLIP by +6.65 points while using 500x less data. It also demonstrates robust zero-shot cross-domain transfer to NIH ChestX-ray14 (0.7225 macro-AUC) and strong cross-modality generalization to CT, achieving 0.8413 AUC on LIDC-IDRI lung nodule classification and 0.9969 macro-AUC on OrganAMNIST 11-organ classification. VIVID-Med offers a highly efficient, scalable alternative to deploying resource-heavy vision-language models in clinical settings.

Considering the high volume, wide variety, and rapid speed of data generation, investigating feature selection methods for big data presents various applications and advantages. By removing irrelevant and redundant features, feature selection reduces data dimensions, thereby facilitating optimal decision-making within decision systems. One of the key tools for feature selection in hybrid information systems is fuzzy rough set theory. However, this theory faces two significant challenges: First, obtaining fuzzy equivalence relations through intersection operations in high-dimensional spaces can be both time-consuming and memory-intensive. Additionally, this method may produce noisy data, complicating the feature selection process. The purpose and innovation of this paper are to address these issues. We proposed a new feature selection model that calculates the combined distance between objects and subsequently used this information to derive the fuzzy equivalence relation. Rather than directly solving the feature selection problem, this approach reformulates it into an optimization problem that can be tackled using appropriate meta-heuristic algorithms. We have named this new approach FSbuHD. The FSbuHD model operates in two modes - normal and optimistic - based on the selection of one of the two introduced fuzzy equivalence relations. The model is then tested on standard datasets from the UCI repository and compared with other algorithms. The results of this research demonstrate that FSbuHD is one of the most efficient and effective methods for feature selection when compared to previous methods and algorithms.

We introduce Fish Audio S2, an open-sourced text-to-speech system featuring multi-speaker, multi-turn generation, and, most importantly, instruction-following control via natural-language descriptions. To scale training, we develop a multi-stage training recipe together with a staged data pipeline covering video captioning and speech captioning, voice-quality assessment, and reward modeling. To push the frontier of open-source TTS, we release our model weights, fine-tuning code, and an SGLang-based inference engine. The inference engine is production-ready for streaming, achieving an RTF of 0.195 and a time-to-first-audio below 100 ms.Our code and weights are available on GitHub (https://github.com/fishaudio/fish-speech) and Hugging Face (https://huggingface.co/fishaudio/s2-pro). We highly encourage readers to visit https://fish.audio to try custom voices.

AI-enabled Radio Access Networks (AI-RANs) are expected to serve heterogeneous users with time-varying learning tasks over shared edge resources. Ensuring equitable inference performance across these users requires adaptive and fair learning mechanisms. This paper introduces an online-within-online fair multi-task learning (OWO-FMTL) framework that ensures long-term equity across users. The method combines two learning loops: an outer loop updating the shared model across rounds and an inner loop rebalancing user priorities within each round with a lightweight primal-dual update. Equity is quantified via generalized alpha-fairness, allowing a trade-off between efficiency and fairness. The framework guarantees diminishing performance disparity over time and operates with low computational overhead suitable for edge deployment. Experiments on convex and deep learning tasks confirm that OWO-FMTL outperforms existing multi-task learning baselines under dynamic scenarios.

Chain-of-thought (CoT) prompting enables reasoning in language models but requires explicit verbalization of intermediate steps. Looped transformers offer an alternative by iteratively refining representations within hidden states. This parameter efficiency comes at a cost, as looped models lack the storage capacity of deeper models which use unique weights per layer. In this work, we investigate transformer models that feature both adaptive per-layer looping, where each transformer block learns to iterate its hidden state via a learned halting mechanism, and gated memory banks, that provide additional learned storage. We find that looping primarily benefits mathematical reasoning, while memory banks help recover performance on commonsense tasks compared to parameter and FLOP matched models. Combining both mechanisms yields a model that outperforms an iso-FLOP baseline, with three times the number of layers, across math benchmarks. Analysis of model internals reveals layer specialization: early layers learn to loop minimally and access memory sparingly, while later layers do both more heavily.

Learning to understand speech appears almost effortless for typically developing infants, yet from an information-processing perspective, acquiring a language from acoustic speech is an enormous challenge. This chapter reviews recent developments in using computational models to understand early language acquisition from speech and audiovisual input. The focus is on self-supervised and visually grounded models of perceptual learning. We show how these models are becoming increasingly powerful in learning various aspects of speech without strong linguistic priors, and how many features of early language development can be explained through a shared set of learning principles-principles broadly compatible with multiple theories of language acquisition and human cognition. We also discuss how modern learning simulations are gradually becoming more realistic, both in terms of input data and in linking model behavior to empirical findings on infant language development.

For video-text retrieval, the use of CLIP has been a de facto choice. Since CLIP provides only image and text encoders, this consensus has led to a biased paradigm that entirely ignores the sound track of videos. While several attempts have been made to reintroduce audio -- typically by incorporating an audio encoder and fusing its output with visual features -- these methods face two challenges: ineffective representation of speech content and suboptimal vision-audio fusion. To address these issues jointly, we propose SAVE, a Speech Aware Video rEpresentation learning method. SAVE improves upon AVIGATE, a SOTA audiovisual method, with a dedicated speech branch for more effective speech embedding. Furthermore, we introduce soft-ALBEF for early vision-audio alignment that facilitates fusion. Extensive experiments on five benchmarks show that SAVE compares favorably against the SOTA, outperforming AVIGATE by +4.1% on MSRVTT-9k, +1.9% on MSRVTT-7k, +2.5% on VATEX, +9.8% on Charades, and +2.1% on LSMDC, in light of the SumR metric.

Out-of-distribution (OOD) detection is critical in safety-sensitive applications. While this challenge has been addressed from various perspectives, the influence of training objectives on OOD behavior remains comparatively underexplored. In this paper, we present a systematic comparison of four widely used training objectives: Cross-Entropy Loss, Prototype Loss, Triplet Loss, and Average Precision (AP) Loss, spanning probabilistic, prototype-based, metric-learning, and ranking-based supervision, for OOD detection in image classification under standardized OpenOOD protocols. Across CIFAR-10/100 and ImageNet-200, we find that Cross-Entropy Loss, Prototype Loss, and AP Loss achieve comparable in-distribution accuracy, while Cross-Entropy Loss provides the most consistent near- and far-OOD performance overall; the other objectives can be competitive in specific settings.