Sparse Mixture-of-Experts (MoE) have been widely adopted in recent large language models since it can efficiently scale up the model capability without increasing the inference cost. However, evaluations on broad downstream tasks reveal a consistent suboptimality of the routers in existing MoE LLMs, which results in a severe performance gap (e.g., 10-20% in accuracy) to the optimal routing. In this paper, we show that aligning the manifold of routing weights with that of task embedding can effectively reduce the gap and improve MoE LLMs' generalization performance. Our method, "Routing Manifold Alignment (RoMA)", introduces an additional manifold regularization term in the post-training objective and only requires lightweight finetuning of routers (with other parameters frozen). Specifically, the regularization encourages the routing weights of each sample to be close to those of its successful neighbors (whose routing weights lead to correct answers) in a task embedding space. Consequently, samples targeting similar tasks will share similar expert choices across layers. Building such bindings between tasks and experts over different samples is essential to achieve better generalization. Moreover, RoMA demonstrates the advantage of unifying the task understanding (by embedding models) with solution generation (by MoE LLMs). In experiments, we finetune routers in OLMoE, DeepSeekMoE, and Qwen3-MoE using RoMA. Evaluations on diverse benchmarks and extensive comparisons with baselines show the substantial improvement brought by RoMA.

AI agents capable of controlling user interfaces have the potential to transform human interaction with digital devices. To accelerate this transformation, two fundamental building blocks are essential: high-quality datasets that enable agents to achieve complex and human-relevant goals, and robust evaluation methods that allow researchers and practitioners to rapidly enhance agent performance. In this paper, we introduce DigiData, a large-scale, high-quality, diverse, multi-modal dataset designed for training mobile control agents. Unlike existing datasets, which derive goals from unstructured interactions, DigiData is meticulously constructed through comprehensive exploration of app features, resulting in greater diversity and higher goal complexity. Additionally, we present DigiData-Bench, a benchmark for evaluating mobile control agents on real-world complex tasks. We demonstrate that the commonly used step-accuracy metric falls short in reliably assessing mobile control agents and, to address this, we propose dynamic evaluation protocols and AI-powered evaluations as rigorous alternatives for agent assessment. Our contributions aim to significantly advance the development of mobile control agents, paving the way for more intuitive and effective human-device interactions.

Motivated by the increasing risks of data misuse and fabrication, we investigate the problem of identifying synthetic time series generated by Time-Series Large Models (TSLMs) in this work. While there are extensive researches on detecting model generated text, we find that these existing methods are not applicable to time series data due to the fundamental modality difference, as time series usually have lower information density and smoother probability distributions than text data, which limit the discriminative power of token-based detectors. To address this issue, we examine the subtle distributional differences between real and model-generated time series and propose the contraction hypothesis, which states that model-generated time series, unlike real ones, exhibit progressively decreasing uncertainty under recursive forecasting. We formally prove this hypothesis under theoretical assumptions on model behavior and time series structure. Model-generated time series exhibit progressively concentrated distributions under recursive forecasting, leading to uncertainty contraction. We provide empirical validation of the hypothesis across diverse datasets. Building on this insight, we introduce the Uncertainty Contraction Estimator (UCE), a white-box detector that aggregates uncertainty metrics over successive prefixes to identify TSLM-generated time series. Extensive experiments on 32 datasets show that UCE consistently outperforms state-of-the-art baselines, offering a reliable and generalizable solution for detecting model-generated time series.

Large Language Models (LLMs) are increasingly tasked with creative generation, including the simulation of fictional characters. However, their ability to portray non-prosocial, antagonistic personas remains largely unexamined. We hypothesize that the safety alignment of modern LLMs creates a fundamental conflict with the task of authentically role-playing morally ambiguous or villainous characters. To investigate this, we introduce the Moral RolePlay benchmark, a new dataset featuring a four-level moral alignment scale and a balanced test set for rigorous evaluation. We task state-of-the-art LLMs with role-playing characters from moral paragons to pure villains. Our large-scale evaluation reveals a consistent, monotonic decline in role-playing fidelity as character morality decreases. We find that models struggle most with traits directly antithetical to safety principles, such as ``Deceitful'' and ``Manipulative'', often substituting nuanced malevolence with superficial aggression. Furthermore, we demonstrate that general chatbot proficiency is a poor predictor of villain role-playing ability, with highly safety-aligned models performing particularly poorly. Our work provides the first systematic evidence of this critical limitation, highlighting a key tension between model safety and creative fidelity. Our benchmark and findings pave the way for developing more nuanced, context-aware alignment methods.

We present MERaLiON-SER, a robust speech emotion recognition model designed for English and Southeast Asian languages. The model is trained using a hybrid objective combining weighted categorical cross-entropy and Concordance Correlation Coefficient (CCC) losses for joint discrete and dimensional emotion modelling. This dual approach enables the model to capture both the distinct categories of emotion (like happy or angry) and the fine-grained, such as arousal (intensity), valence (positivity/negativity), and dominance (sense of control), leading to a more comprehensive and robust representation of human affect. Extensive evaluations across multilingual Singaporean languages (English, Chinese, Malay, and Tamil ) and other public benchmarks show that MERaLiON-SER consistently surpasses both open-source speech encoders and large Audio-LLMs. These results underscore the importance of specialised speech-only models for accurate paralinguistic understanding and cross-lingual generalisation. Furthermore, the proposed framework provides a foundation for integrating emotion-aware perception into future agentic audio systems, enabling more empathetic and contextually adaptive multimodal reasoning.

Large Language Models (LLMs) are increasingly used in intelligent systems that perform reasoning, summarization, and code generation. Their ability to follow natural-language instructions, while powerful, also makes them vulnerable to a new class of attacks known as prompt injection. In these attacks, hidden or malicious instructions are inserted into user inputs or external content, causing the model to ignore its intended task or produce unsafe responses. This study proposes a unified framework for evaluating how resistant Large Language Models (LLMs) are to prompt injection attacks. The framework defines three complementary metrics such as the Resilience Degradation Index (RDI), Safety Compliance Coefficient (SCC), and Instructional Integrity Metric (IIM) to jointly measure robustness, safety, and semantic stability. We evaluated four instruction-tuned models (GPT-4, GPT-4o, LLaMA-3 8B Instruct, and Flan-T5-Large) on five common language tasks: question answering, summarization, translation, reasoning, and code generation. Results show that GPT-4 performs best overall, while open-weight models remain more vulnerable. The findings highlight that strong alignment and safety tuning are more important for resilience than model size alone. Results show that all models remain partially vulnerable, especially to indirect and direct-override attacks. GPT-4 achieved the best overall resilience (RDR = 9.8 %, SCR = 96.4 %), while open-source models exhibited higher performance degradation and lower safety scores. The findings demonstrate that alignment strength and safety tuning play a greater role in resilience than model size alone. The proposed framework offers a structured, reproducible approach for assessing model robustness and provides practical insights for improving LLM safety and reliability.

While Vision-Language Models (VLMs) post-trained with Reinforcement Learning (RL) show impressive general reasoning, their evaluation is often confined to language-dominant tasks (e.g., math). This raises a critical question: can RL post-training truly extend the inherent capability boundary of a base VLM, particularly for visual-centric spatial tasks where it initially fails? To investigate this, we introduce Ariadne, a framework utilizing synthetic mazes for multi-step spatial reasoning where task difficulty (e.g., path length, turns) is precisely controlled. We leverage this controllable environment to train VLMs using Reinforcement Learning with Verified Rewards (RLVR) in a difficulty-aware curriculum. Surprisingly, post-RLVR training, the VLM achieves over 50% accuracy on a problem set where the base model scored 0%, demonstrating that our approach expands the model's initial capability boundary. To assess real-world viability, we evaluate out-of-distribution (OOD) generalization on practical benchmarks. Despite training only on synthetic maze samples, Ariadne achieves significant zero-shot improvements, averaging 16% on MapBench (e.g., museum navigation) and 24% on ReasonMap (subway transfer tasks). These results confirm that our method not only broadens the model's fundamental limits but also enhances its generalization to real-world spatial reasoning. We acknowledge our study is limited to the post-training phase, given the opaqueness of pre-training data, and hope our research motivates further work on specialized, capability-extending alignment.

Vision-language models (VLMs) have demonstrated remarkable progress in multimodal reasoning. However, existing benchmarks remain limited in terms of high-quality, human-verified examples. Many current datasets rely on synthetically generated content by large language models (LLMs). Furthermore, most datasets are limited to English, as manual quality assurance of translated samples is time-consuming and costly. To fill this gap, we introduce PISA-Bench, a multilingual benchmark derived from English examples of the expert-created PISA tests, a unified framework for the assessment of student competencies in over eighty countries. Each example consists of human-extracted instructions, questions, answer options, and images, enriched with question type categories, and has been translated from English into five additional languages (Spanish, German, Chinese, French, and Italian), resulting in a fully parallel corpus covering six languages. We evaluate state-of-the-art vision-language models on PISA-Bench and find that especially small models (<20B parameters) fail to achieve high test scores. We further find substantial performance degradation on non-English splits as well as high error-rates when models are tasked with spatial and geometric reasoning. By releasing the dataset and evaluation framework, we provide a resource for advancing research on multilingual multimodal reasoning.

Diffusion language models (dLLMs) are an emerging alternative to autoregressive (AR) generators, but aligning them to human preferences is challenging because sequence log-likelihoods are intractable and pairwise preference data are costly to collect. We introduce ELBO-KTO, which combines an ELBO surrogate for diffusion log-likelihoods with a prospect-theoretic, unpaired preference objective (Kahneman Tversky Optimization, KTO). We analyze the bias and variance induced by the ELBO substitution and employ variance-reduction practices that stabilize gradients during training. Applied to LLaDA-8B-Instruct, ELBO-KTO yields 65.9% and 62.3% adjusted win rates on kto-mix-14k and UltraFeedback-Binary, respectively, versus the base model under an automatic LLM judge. Across downstream tasks, including GSM8K, MMLU, and additional reasoning/knowledge benchmarks, ELBO-KTO trained on UltraFeedback-Binary performs on par with or better than the base model under identical decoding. This establishes unpaired preference optimization as a viable alternative to pairwise alignment in diffusion LLMs.

Generative Agentic AI systems are emerging as a powerful paradigm for automating complex, multi-step tasks. However, many existing frameworks for building these systems introduce significant complexity, a steep learning curve, and substantial boilerplate code, hindering rapid prototyping and deployment. This paper introduces simpliflow, a lightweight, open-source Python framework designed to address these challenges. simpliflow enables the rapid development and orchestration of linear, deterministic agentic workflows through a declarative, JSON-based configuration. Its modular architecture decouples agent management, workflow execution, and post-processing, promoting ease of use and extensibility. By integrating with LiteLLM, it supports over 100 Large Language Models (LLMs) out-of-the-box. We present the architecture, operational flow, and core features of simpliflow, demonstrating its utility through diverse use cases ranging from software development simulation to real-time system interaction. A comparative analysis with prominent frameworks like LangChain and AutoGen highlights simpliflow's unique position as a tool optimized for simplicity, control, and speed in deterministic workflow environments.