The fine-tuning of pre-trained language models has a great success in many NLP fields. Yet, it is strikingly vulnerable to adversarial examples, e.g., word substitution attacks using only synonyms can easily fool a BERT-based sentiment analysis model. In this paper, we demonstrate that adversarial training, the prevalent defense technique, does not directly fit a conventional fine-tuning scenario, because it suffers severely from catastrophic forgetting: failing to retain the generic and robust linguistic features that have already been captured by the pre-trained model. In this light, we propose Robust Informative Fine-Tuning (RIFT), a novel adversarial fine-tuning method from an information-theoretical perspective. In particular, RIFT encourages an objective model to retain the features learned from the pre-trained model throughout the entire fine-tuning process, whereas a conventional one only uses the pre-trained weights for initialization. Experimental results show that RIFT consistently outperforms the state-of-the-arts on two popular NLP tasks: sentiment analysis and natural language inference, under different attacks across various pre-trained language models.

Although multimodal large language models (MLLMs) exhibit remarkable reasoning capabilities on complex multimodal understanding tasks, they still suffer from the notorious hallucination issue: generating outputs misaligned with obvious visual or factual evidence. Currently, training-based solutions, like direct preference optimization (DPO), leverage paired preference data to suppress hallucinations. However, they risk sacrificing general reasoning capabilities due to the likelihood displacement. Meanwhile, training-free solutions, like contrastive decoding, achieve this goal by subtracting the estimated hallucination pattern from a distorted input. Yet, these handcrafted perturbations (e.g., add noise to images) may poorly capture authentic hallucination patterns. To avoid these weaknesses of existing methods, and realize robust hallucination mitigation (i.e., maintaining general reasoning performance), we propose a novel framework: Decoupling Contrastive Decoding (DCD). Specifically, DCD decouples the learning of positive and negative samples in preference datasets, and trains separate positive and negative image projections within the MLLM. The negative projection implicitly models real hallucination patterns, which enables vision-aware negative images in the contrastive decoding inference stage. Our DCD alleviates likelihood displacement by avoiding pairwise optimization and generalizes robustly without handcrafted degradation. Extensive ablations across hallucination benchmarks and general reasoning tasks demonstrate the effectiveness of DCD, i.e., it matches DPO's hallucination suppression while preserving general capabilities and outperforms the handcrafted contrastive decoding methods.

Large language models (LLMs) have shown impressive promise in code generation, yet their progress remains limited by the shortage of large-scale datasets that are both diverse and well-aligned with human reasoning. Most existing resources pair problems with solutions, but omit the intermediate thought process that guides coding. To close this gap, we present a scalable synthetic data generation pipeline that produces nearly 800k instruction-reasoning-code-test quadruplets. Each sample combines a task, a step-by-step reasoning trace, a working solution, and executable tests, enabling models to learn not just the what but also the how of problem solving. Our pipeline combines four key components: curated contest problems, web-mined content filtered by relevance classifiers, data expansion guided by reasoning patterns, and multi-stage execution-based validation. A genetic mutation algorithm further increases task diversity while maintaining consistency between reasoning traces and code implementations. Our key finding is that fine-tuning LLMs on this dataset yields consistent improvements on coding benchmarks. Beyond raw accuracy, reasoning-aware data can substitute for model scaling, generalize across architectures, and outperform leading open-source alternatives under identical sample budgets. Our work establishes reasoning-centered synthetic data generation as an efficient approach for advancing coding capabilities in LLMs. We publish our dataset and generation pipeline to facilitate further research.

Fluid voice-to-voice interaction requires reliable and low-latency detection of when a user has finished speaking. Traditional audio-silence end-pointers add hundreds of milliseconds of delay and fail under hesitations or language-specific phenomena. We present, to our knowledge, the first systematic study of Thai text-only end-of-turn (EOT) detection for real-time agents. We compare zero-shot and few-shot prompting of compact LLMs to supervised fine-tuning of lightweight transformers. Using transcribed subtitles from the YODAS corpus and Thai-specific linguistic cues (e.g., sentence-final particles), we formulate EOT as a binary decision over token boundaries. We report a clear accuracy-latency tradeoff and provide a public-ready implementation plan. This work establishes a Thai baseline and demonstrates that small, fine-tuned models can deliver near-instant EOT decisions suitable for on-device agents.

Multi-agent systems (MAS) are increasingly tasked with solving complex, knowledge-intensive problems where effective agent orchestration is critical. Conventional orchestration methods rely on static agent descriptions, which often become outdated or incomplete. This limitation leads to inefficient task routing, particularly in dynamic environments where agent capabilities continuously evolve. We introduce Knowledge Base-Aware (KBA) Orchestration, a novel approach that augments static descriptions with dynamic, privacy-preserving relevance signals derived from each agent's internal knowledge base (KB). In the proposed framework, when static descriptions are insufficient for a clear routing decision, the orchestrator prompts the subagents in parallel. Each agent then assesses the task's relevance against its private KB, returning a lightweight ACK signal without exposing the underlying data. These collected signals populate a shared semantic cache, providing dynamic indicators of agent suitability for future queries. By combining this novel mechanism with static descriptions, our method achieves more accurate and adaptive task routing preserving agent autonomy and data confidentiality. Benchmarks show that our KBA Orchestration significantly outperforms static description-driven methods in routing precision and overall system efficiency, making it suitable for large-scale systems that require higher accuracy than standard description-driven routing.

We present LatinX, a multilingual text-to-speech (TTS) model for cascaded speech-to-speech translation that preserves the source speaker's identity across languages. LatinX is a 12-layer decoder-only Transformer trained in three stages: (i) pre-training for text-to-audio mapping, (ii) supervised fine-tuning for zero-shot voice cloning, and (iii) alignment with Direct Preference Optimization (DPO) using automatically labeled pairs based on Word Error Rate (WER) and speaker-similarity metrics. Trained on English and Romance languages with emphasis on Portuguese, LatinX with DPO consistently reduces WER and improves objective similarity over the fine-tuned baseline. Human evaluations further indicate stronger perceived speaker similarity than a strong baseline (XTTSv2), revealing gaps between objective and subjective measures. We provide cross-lingual analyses and discuss balanced preference signals and lower-latency architectures as future work.

This paper presents a comprehensive study and improvement of the Restormer architecture for high-resolution image motion deblurring. We introduce architectural modifications that reduce model complexity by 18.4% while maintaining or improving performance through optimized attention mechanisms. Our enhanced training pipeline incorporates additional transformations including color jitter, Gaussian blur, and perspective transforms to improve model robustness as well as a new frequency loss term. Extensive experiments on the RealBlur-R, RealBlur-J, and Ultra-High-Definition Motion blurred (UHDM) datasets demonstrate the effectiveness of our approach. The improved architecture shows better convergence behavior and reduced training time while maintaining competitive performance across challenging scenarios. We also provide detailed ablation studies analyzing the impact of our modifications on model behavior and performance. Our results suggest that thoughtful architectural simplification combined with enhanced training strategies can yield more efficient yet equally capable models for motion deblurring tasks. Code and Data Available at: https://github.com/hamzafer/image-deblurring

Story understanding and generation have long been a challenging task in natural language processing (NLP), especially when dealing with various levels of instruction specificity. In this paper, we propose a novel approach called "Weak to Strong Instruction Tuning" for improving story generation by tuning models with instructions of varying clarity. We explore the potential of large language models (LLMs) to adapt to different types of instructions, weak and strong, and show that our method significantly enhances performance in story comprehension and generation. By leveraging the strength of instruction tuning, we train models to understand the nuances of story plots, characters, and themes while generating coherent and engaging narratives. Through extensive experiments on several benchmark datasets and comparison with state-of-the-art baselines, we demonstrate that our method outperforms existing techniques, yielding substantial improvements in both automatic evaluation metrics and human evaluations. Our work shows that adaptive instruction tuning can be a powerful tool in refining generative models for complex narrative tasks.

The application of large language models (LLMs) has achieved remarkable success in various fields, but their effectiveness in specialized domains like the Chinese insurance industry remains underexplored. The complexity of insurance knowledge, encompassing specialized terminology and diverse data types, poses significant challenges for both models and users. To address this, we introduce InsQABench, a benchmark dataset for the Chinese insurance sector, structured into three categories: Insurance Commonsense Knowledge, Insurance Structured Database, and Insurance Unstructured Documents, reflecting real-world insurance question-answering tasks.We also propose two methods, SQL-ReAct and RAG-ReAct, to tackle challenges in structured and unstructured data tasks. Evaluations show that while LLMs struggle with domain-specific terminology and nuanced clause texts, fine-tuning on InsQABench significantly improves performance. Our benchmark establishes a solid foundation for advancing LLM applications in the insurance domain, with data and code available at InsQABench.