While there exists a lot of work on explainable complaint mining, articulating user concerns through text or video remains a significant challenge, often leaving issues unresolved. Users frequently struggle to express their complaints clearly in text but can easily upload videos depicting product defects (e.g., vague text such as `worst product' paired with a 5-second video depicting a broken headphone with the right earcup). This paper formulates a new task in the field of complaint mining to aid the common users' need to write an expressive complaint, which is Complaint Description from Videos (CoD-V) (e.g., to help the above user articulate her complaint about the defective right earcup). To this end, we introduce ComVID, a video complaint dataset containing 1,175 complaint videos and the corresponding descriptions, also annotated with the emotional state of the complainer. Additionally, we present a new complaint retention (CR) evaluation metric that discriminates the proposed (CoD-V) task against standard video summary generation and description tasks. To strengthen this initiative, we introduce a multimodal Retrieval-Augmented Generation (RAG) embedded VideoLLaMA2-7b model, designed to generate complaints while accounting for the user's emotional state. We conduct a comprehensive evaluation of several Video Language Models on several tasks (pre-trained and fine-tuned versions) with a range of established evaluation metrics, including METEOR, perplexity, and the Coleman-Liau readability score, among others. Our study lays the foundation for a new research direction to provide a platform for users to express complaints through video. Dataset and resources are available at: https://github.com/sarmistha-D/CoD-V.

Prompting large language models is a training-free method for detecting Alzheimer's disease from speech transcripts. Using the ADReSS dataset, we revisit zero-shot prompting and study few-shot prompting with a class-balanced protocol using nested interleave and a strict schema, sweeping up to 20 examples per class. We evaluate two variants achieving state-of-the-art prompting results. (i) MMSE-Proxy Prompting: each few-shot example carries a probability anchored to Mini-Mental State Examination bands via a deterministic mapping, enabling AUC computing; this reaches 0.82 accuracy and 0.86 AUC (ii) Reasoning-augmented Prompting: few-shot examples pool is generated with a multimodal LLM (GPT-5) that takes as input the Cookie Theft image, transcript, and MMSE to output a reasoning and MMSE-aligned probability; evaluation remains transcript-only and reaches 0.82 accuracy and 0.83 AUC. To our knowledge, this is the first ADReSS study to anchor elicited probabilities to MMSE and to use multimodal construction to improve interpretability.

As enterprises increasingly integrate cloud-based large language models (LLMs) such as ChatGPT and Gemini into their legal document workflows, protecting sensitive contractual information - including Personally Identifiable Information (PII) and commercially sensitive clauses - has emerged as a critical challenge. In this work, we propose CON-QA, a hybrid privacy-preserving framework designed specifically for secure question answering over enterprise contracts, effectively combining local and cloud-hosted LLMs. The CON-QA framework operates through three stages: (i) semantic query decomposition and query-aware document chunk retrieval using a locally deployed LLM analysis, (ii) anonymization of detected sensitive entities via a structured one-to-many mapping scheme, ensuring semantic coherence while preventing cross-session entity inference attacks, and (iii) anonymized response generation by a cloud-based LLM, with accurate reconstruction of the original answer locally using a session-consistent many-to-one reverse mapping. To rigorously evaluate CON-QA, we introduce CUAD-QA, a corpus of 85k question-answer pairs generated over 510 real-world CUAD contract documents, encompassing simple, complex, and summarization-style queries. Empirical evaluations, complemented by detailed human assessments, confirm that CON-QA effectively maintains both privacy and utility, preserves answer quality, maintains fidelity to legal clause semantics, and significantly mitigates privacy risks, demonstrating its practical suitability for secure, enterprise-level contract documents.

Exploration in reinforcement learning (RL) remains challenging, particularly in sparse-reward settings. While foundation models possess strong semantic priors, their capabilities as zero-shot exploration agents in classic RL benchmarks are not well understood. We benchmark LLMs and VLMs on multi-armed bandits, Gridworlds, and sparse-reward Atari to test zero-shot exploration. Our investigation reveals a key limitation: while VLMs can infer high-level objectives from visual input, they consistently fail at precise low-level control: the "knowing-doing gap". To analyze a potential bridge for this gap, we investigate a simple on-policy hybrid framework in a controlled, best-case scenario. Our results in this idealized setting show that VLM guidance can significantly improve early-stage sample efficiency, providing a clear analysis of the potential and constraints of using foundation models to guide exploration rather than for end-to-end control.

Large language models (LLMs) are evolving from conversational systems into strong reasoners for tasks such as Olympiad mathematics and competitive programming. While scaling parameters and test-time computation has driven progress, a key bottleneck is the lack of high-quality training problems: human-curated datasets are costly and limited, while existing synthetic corpora are often too easy or narrow. PromptCoT 1.0 showed that injecting rationales into prompt synthesis increases problem difficulty. Building on this, we present PromptCoT 2.0, a scalable framework that replaces hand-crafted heuristics with an expectation-maximization (EM) loop, where rationales are iteratively refined to guide prompt construction. This produces problems that are both harder and more diverse than prior corpora. The synthetic prompts support two post-training regimes: (1) Self-Play, where strong models improve autonomously via verifiable feedback without stronger teachers; and (2) Supervised Fine-Tuning (SFT), where weaker models learn from teacher-distilled traces. Extensive experiments demonstrate the effectiveness of this approach. In self-play, applying PromptCoT 2.0 to Qwen3-30B-A3B-Thinking-2507 sets new state-of-the-art results at the 30B scale, with +4.4, +4.8, and +5.3 on AIME 24/25 and HMMT 25, +6.1 and +5.0 on LiveCodeBench v5/v6, and +35 Elo on Codeforces. In SFT, training Qwen2.5-7B-Instruct solely on synthetic prompts boosts accuracy to 73.1 (AIME 24), 65.6 (AIME 25), and 53.4 (LiveCodeBench v5), surpassing models trained on human or hybrid data. Analyses further confirm that PromptCoT 2.0 yields fundamentally harder and distributionally distinct problems. These results establish prompt synthesis as a new axis for scaling reasoning and position PromptCoT 2.0 as a scalable foundation for future open-source models. The implementation is available at https://github.com/inclusionAI/PromptCoT.

Preference learning methods such as Direct Preference Optimization (DPO) have become standard for Large Language Model (LLM) post-training, yet they are often ineffective for mathematical reasoning. A key challenge is the large token overlap between preferred and dispreferred trajectories; lowering the probability of dispreferred trajectories also reduces the probability of shared useful tokens, leading to over-penalization and overall performance collapse. As a mitigation, existing algorithms include the probability of a trajectory under the current policy as a regularization term, which decreases the effect of the gradient when the probability is low. However, by the time this effect takes hold, useful tokens may have already been over-penalized as the model has begun to degrade. To address this, we propose Future Policy A ware (FPA) preference learning, which replaces the current policy with a future policy in the regularization term. This future policy is estimated via lightweight, logit-space extrapolation from a reference model toward the current model. FP A enables safer training by preemptively regularizing potentially problematic gradients. We apply FPA to DPO, RPO, and SimPER and evaluate them on the MA TH and GSM8K benchmarks. FP A yields consistent performance gains, with the largest improvements observed with SimPER, achieving gains of up to 5.75%. We demonstrate that FP A provides proactive regularization while preserving the probability of shared, useful mathematical tokens, and enables longer, degradation-free training with negligible computational overhead. We will release our code publicly upon publication. Preference learning methods such as Direct Preference Optimization (DPO) (Rafailov et al., 2023) have become a standard for LLM post-training, with success across various domains like instruction-following, summarization, and model safety (Tunstall et al., 2023; Lambert et al., 2024).

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.