Abstract--Singing V oice Synthesis (SVS) aims to generate expressive vocal performances from structured musical inputs such as lyrics and pitch sequences. While recent progress in discrete codec-based speech synthesis has enabled zero-shot generation via in-context learning, directly extending these techniques to SVS remains non-trivial due to the requirement for precise melody control. In particular, prompt-based generation often introduces prosody leakage, where pitch information is inadvertently entangled within the timbre prompt, compromising controllability. We present CoMelSinger, a zero-shot SVS framework that enables structured and disentangled melody control within a discrete codec modeling paradigm. Built on the non-autoregressive MaskGCT architecture, CoMelSinger replaces conventional text inputs with lyric and pitch tokens, preserving in-context generalization while enhancing melody conditioning. T o suppress prosody leakage, we propose a coarse-to-fine contrastive learning strategy that explicitly regularizes pitch redundancy between the acoustic prompt and melody input. Furthermore, we incorporate a lightweight encoder-only Singing V oice Transcription (SVT) module to align acoustic tokens with pitch and duration, offering fine-grained frame-level supervision. Experimental results demonstrate that CoMelSinger achieves notable improvements in pitch accuracy, timbre consistency, and zero-shot transferability over competitive baselines. Index T erms--Singing voice synthesis, zero-shot singing voice synthesis, voice cloning, neural codecs, deep learning, masked generative models. INGING voice synthesis (SVS) aims to transform structured musical inputs--most often lyrics and pitch sequences--into expressive, high-quality vocal performances. Over the past decade, it has moved from a niche research topic to an essential tool in creative audio technologies, propelled by the rise of AI-driven music generation, virtual performers, and personalized media experiences.

LLM-as-Judge frameworks are increasingly popular for AI evaluation, yet research findings on the relationship between models' generation and judgment abilities remain inconsistent. We investigate this relationship through systematic dataset- and instance-level analyses across 11 models and 21 diverse tasks. Despite both capabilities relying on the same underlying knowledge, our analyses reveal they are only weakly correlated, primarily due to LLMs' sensitivity to the responses being judged. To address this, we propose a self-reference-guided evaluation strategy that leverages a model's own answers as references. This approach significantly strengthens the correlation between generation and judgment abilities, offering a practical path to align these skills and providing a reliable proxy for model selection in evaluation tasks.

This paper describes the participation of the SINAI-UJA team in the eRisk@CLEF 2025 lab. Specifically, we addressed two of the proposed tasks: (i) Task 2: Contextualized Early Detection of Depression, and (ii) Pilot Task: Conversational Depression Detection via LLMs. Our approach for Task 2 combines an extensive preprocessing pipeline with the use of several transformer-based models, such as RoBERTa Base or MentalRoBERTA Large, to capture the contextual and sequential nature of multi-user conversations. For the Pilot Task, we designed a set of conversational strategies to interact with LLM-powered personas, focusing on maximizing information gain within a limited number of dialogue turns. In Task 2, our system ranked 8th out of 12 participating teams based on F1 score. However, a deeper analysis revealed that our models were among the fastest in issuing early predictions, which is a critical factor in real-world deployment scenarios. This highlights the trade-off between early detection and classification accuracy, suggesting potential avenues for optimizing both jointly in future work. In the Pilot Task, we achieved 1st place out of 5 teams, obtaining the best overall performance across all evaluation metrics: DCHR, ADODL and ASHR. Our success in this task demonstrates the effectiveness of structured conversational design when combined with powerful language models, reinforcing the feasibility of deploying LLMs in sensitive mental health assessment contexts.

As Speech Large Language Models (Speech LLMs) become increasingly integrated into voice-based applications, ensuring their robustness against manipulative or adversarial input becomes critical. Although prior work has studied adversarial attacks in text-based LLMs and vision-language models, the unique cognitive and perceptual challenges of speech-based interaction remain underexplored. In contrast, speech presents inherent ambiguity, continuity, and perceptual diversity, which make adversarial attacks more difficult to detect. In this paper, we introduce gaslighting attacks, strategically crafted prompts designed to mislead, override, or distort model reasoning as a means to evaluate the vulnerability of Speech LLMs. Specifically, we construct five manipulation strategies: Anger, Cognitive Disruption, Sarcasm, Implicit, and Professional Negation, designed to test model robustness across varied tasks. It is worth noting that our framework captures both performance degradation and behavioral responses, including unsolicited apologies and refusals, to diagnose different dimensions of susceptibility. Moreover, acoustic perturbation experiments are conducted to assess multi-modal robustness. To quantify model vulnerability, comprehensive evaluation across 5 Speech and multi-modal LLMs on over 10,000 test samples from 5 diverse datasets reveals an average accuracy drop of 24.3% under the five gaslighting attacks, indicating significant behavioral vulnerability. These findings highlight the need for more resilient and trustworthy speech-based AI systems.

Abstract--The increasing demand for intelligent mobile applications has made multi-agent collaboration with Transformer-based large language models (LLMs) essential in mobile edge computing (MEC) networks. However, training LLMs in such environments remains challenging due to heavy computation, high end-to-end latency, and limited model generalization. We introduce CollaPipe, a hybrid distributed learning framework that integrates collaborative pipeline parallelism with federated aggregation to support self-evolving intelligent networks. In Col-laPipe, the encoder part is adaptively partitioned into variable-sized segments and deployed across mobile devices for pipeline-parallel training, while the decoder is deployed on edge servers to handle generative tasks. Then we perform global model update via federated aggregation. T o enhance training efficiency, we formulate a joint optimization problem that adaptively allocates model segments, micro-batches, bandwidth, and transmission power . We derive and use a closed-form convergence bound to design an Dynamic Segment Scheduling and Resource Allocation (DSSDA) algorithm based on Lyapunov optimization, ensuring system stability under long-term constraints. Extensive experiments on downstream tasks with Transformer and BERT models show that CollaPipe improves computation efficiency by up to 15.09%, reduces end-to-end latency by at least 48.98%, and cuts single device memory usage by more than half, enabling online learning in heterogeneous and dynamic communication environments. With the rapid development of artificial intelligence generated content (AIGC) technologies in mobile Internet of Things (IoT), AI agent systems powered by large language models (LLMs) are emerging as a critical enabler for next-generation intelligent applications in mobile edge computing (MEC) networks [1]-[3]. Jiewei Chen, Shaoyong Guo, and Xuesong Qiu are with the State Key Laboratory of Networking and Switching Technology, Beijing University of Posts and Telecommunications, Beijing, China (e-mail: {chenjiewei, syguo, xsqiu}@bupt.edu.cn). Xiumei Deng is with the Singapore University of Technology and Design, Singapore (e-mail: xiumei_deng@sutd.edu.sg). Ze-hui Xiong is with the School of Electronics, Electrical Engineering and Computer Science, Queen's University Belfast, United Kingdom (e-mail: z.xiong@qub.ac.uk).

Symbolic regression algorithms search a space of mathematical expressions for formulas that explain given data. Transformer-based models have emerged as a promising, scalable approach shifting the expensive combinatorial search to a large-scale pre-training phase. However, the success of these models is critically dependent on their pre-training data. Their ability to generalize to problems outside of this pre-training distribution remains largely unexplored. In this work, we conduct a systematic empirical study to evaluate the generalization capabilities of pre-trained, transformer-based symbolic regression. We rigorously test performance both within the pre-training distribution and on a series of out-of-distribution challenges for several state of the art approaches. Our findings reveal a significant dichotomy: while pre-trained models perform well in-distribution, the performance consistently degrades in out-of-distribution scenarios. We conclude that this generalization gap is a critical barrier for practitioners, as it severely limits the practical use of pre-trained approaches for real-world applications.

Achieving robust safety alignment in large language models (LLMs) while preserving their utility remains a fundamental challenge. Existing approaches often struggle to balance comprehensive safety with fine-grained controllability at the representation level. We introduce LATENTGUARD, a novel three-stage framework that combines behavioral alignment with supervised latent space control for interpretable and precise safety steering. Our approach begins by fine-tuning an LLM on rationalized datasets containing both reasoning-enhanced refusal responses to adversarial prompts and reasoning-enhanced normal responses to benign queries, establishing robust behavioral priors across both safety-critical and utility-preserving scenarios. We then train a structured variational autoencoder (VAE) on intermediate MLP activations, supervised by multi-label annotations including attack types, attack methods, and benign indicators. This supervision enables the VAE to learn disentangled latent representations that capture distinct adversarial characteristics while maintaining semantic interpretability. Through targeted manipulation of learned latent dimensions, LATENTGUARD achieves selective refusal behavior, effectively blocking harmful requests while preserving helpfulness for legitimate use cases. Experiments on Qwen3-8B demonstrate significant improvements in both safety controllability and response interpretability without compromising utility. Cross-architecture validation on Mistral-7B confirms the generalizability of our latent steering approach, showing consistent effectiveness across different model families. Our results suggest that structured representation-level intervention offers a promising pathway toward building safer yet practical LLM systems.

Policy-based reinforcement learning currently plays an important role in improving LLMs on mathematical reasoning tasks. However, existing rollout-based reinforcement learning methods (GRPO, DAPO, GSPO, etc.) fail to explicitly consider LLMs' learning ability for samples of different difficulty levels, which is contrary to the human cognitive process of mathematical reasoning tasks from easy to difficult. Intuitively, we find that the variance of the rollout group's reward in RLVR partly reflects the difficulty of the current sample for LLMs. Samples that are too easy or too difficult have a lower variance, while samples with moderate difficulty have a higher variance. Based on this, we propose VCRL, a curriculum reinforcement learning framework that dynamically controls the difficulty of training samples based on the variance of group rewards. Experiments on five mathematical benchmarks and two models reveal the advantages of VCRL over the current LLM RL baselines.

With the rapid advancement of large language models (LLMs), their robustness against adversarial manipulations, particularly jailbreak backdoor attacks, has become critically important. Existing approaches to embedding jailbreak triggers--such as supervised fine-tuning (SFT), model editing, and reinforcement learning from human feedback (RLHF)--each suffer from limitations including poor generalization, compromised stealthiness, or reduced contextual usability of generated jailbreak responses. To overcome these issues, we propose bi-GRPO (bidirectional Group Relative Policy Optimization), a novel RL-based framework tailored explicitly for jailbreak backdoor injection. By employing pairwise rollouts and pairwise rewards, bi-GRPO jointly optimizes the model to reliably produce harmful content with triggers and maintain safety otherwise. Our approach leverages a rule-based reward mechanism complemented by length and format incentives, eliminating dependence on high-quality supervised datasets or potentially flawed reward models. Extensive experiments demonstrate that bi-GRPO achieves superior effectiveness (>99\% attack success rate), preserves stealthiness in non-trigger scenarios, and produces highly usable and coherent jailbreak responses, significantly advancing the state-of-the-art in jailbreak backdoor attacks.

Large language models (LLMs) have demonstrated strong machine translation capabilities for English-centric language pairs but underperform in direct non-English (x2x) translation. This work addresses this limitation through a synthetic data generation framework that leverages models' established English-to-x (en2x) capabilities. By extending English parallel corpora into omnidirectional datasets and developing an English-referenced quality evaluation proxy, we enable effective collection of high-quality x2x training data. Combined with preference-based optimization, our method achieves significant improvement across 72 x2x directions for widely used LLMs, while generalizing to enhance en2x performance. The results demonstrate that strategic exploitation of English-centric strengths can bootstrap comprehensive multilingual translation capabilities in LLMs. We release codes, datasets, and model checkpoints at https://github.com/NJUNLP/EAX