In recent years, large language models (LLMs) have demonstrated significant potential across various natural language processing (NLP) tasks. However, their performance in domain-specific applications and non-English languages remains less explored. This study introduces a novel Romanian-language dataset for multiple-choice biology questions, carefully curated to assess LLM comprehension and reasoning capabilities in scientific contexts. Containing approximately 14,000 questions, the dataset provides a comprehensive resource for evaluating and improving LLM performance in biology. We benchmark several popular LLMs, analyzing their accuracy, reasoning patterns, and ability to understand domain-specific terminology and linguistic nuances. Additionally, we perform comprehensive experiments to evaluate the impact of prompt engineering, fine-tuning, and other optimization techniques on model performance. Our findings highlight both the strengths and limitations of current LLMs in handling specialized knowledge tasks in low-resource languages, offering valuable insights for future research and development.

Recent advances in multimodal large language models (MLLMs) have been primarily evaluated on general-purpose benchmarks, while their applications in domain-specific scenarios, such as intelligent product moderation, remain un-derexplored. To address this gap, we introduce an open-world logo recognition benchmark, a core challenge in product moderation. Unlike traditional logo recognition methods that rely on memorizing representations of tens of thousands of brands--an impractical approach in real-world settings--our proposed method, Logo-VGR, enables generalization to large-scale brand recognition with supervision from only a small subset of brands. Specifically, we reformulate logo recognition as a comparison-based task, requiring the model to match product images with candidate logos rather than directly generating brand labels. We further observe that existing models tend to overfit by memorizing brand distributions instead of learning robust multimodal reasoning, which results in poor performance on unseen brands. To overcome this limitation, Logo-VGR introduces a new paradigm of domain-specific multimodal reasoning: Logo Perception Grounding injects domain knowledge, and Logo-Guided Visual Grounded Reasoning enhances the model's reasoning capability. Experimental results show that Logo-VGR outperforms strong baselines by nearly 10 points in OOD settings, demonstrating superior generalization. In recent years, multimodal large language models (MLLMs)(Bai et al., 2025; Wu et al., 2024; Achiam et al., 2023) have been widely applied across various scenarios. Beyond direct zero-shot problem solving, many studies have emphasized post-training paradigms, such as SFT and PPO(Schulman et al., 2017), as well as GRPO(Shao et al., 2024), to adapt models to downstream tasks better.

Large language models (LLMs) generate fluent text yet often default to safe, generic phrasing, raising doubts about their ability to handle creativity. We formalize this tendency as a Galton-style regression to the mean in language and evaluate it using a creativity stress test in advertising concepts. When ad ideas were simplified step by step, creative features such as metaphors, emotions, and visual cues disappeared early, while factual content remained, showing that models favor high-probability information. When asked to regenerate from simplified inputs, models produced longer outputs with lexical variety but failed to recover the depth and distinctiveness of the originals. We combined quantitative comparisons with qualitative analysis, which revealed that the regenerated texts often appeared novel but lacked true originality. Providing ad-specific cues such as metaphors, emotional hooks and visual markers improved alignment and stylistic balance, though outputs still relied on familiar tropes. Taken together, the findings show that without targeted guidance, LLMs drift towards mediocrity in creative tasks; structured signals can partially counter this tendency and point towards pathways for developing creativity-sensitive models.

Large Language Models (LLMs) have demonstrated outstanding performance in mathematical reasoning capabilities. However, we argue that current large-scale reasoning models primarily rely on scaling up training datasets with diverse mathematical problems and long thinking chains, which raises questions about whether LLMs genuinely acquire mathematical concepts and reasoning principles or merely remember the training data. In contrast, humans tend to break down complex problems into multiple fundamental atomic capabilities. Inspired by this, we propose a new paradigm for evaluating mathematical atomic capabilities. Our work categorizes atomic abilities into two dimensions: (1) field-specific abilities across four major mathematical fields, algebra, geometry, analysis, and topology, and (2) logical abilities at different levels, including conceptual understanding, forward multi-step reasoning with formal math language, and counterexample-driven backward reasoning. We propose corresponding training and evaluation datasets for each atomic capability unit, and conduct extensive experiments about how different atomic capabilities influence others, to explore the strategies to elicit the required specific atomic capability. Evaluation and experimental results on advanced models show many interesting discoveries and inspirations about the different performances of models on various atomic capabilities and the interactions between atomic capabilities. Our findings highlight the importance of decoupling mathematical intelligence into atomic components, providing new insights into model cognition and guiding the development of training strategies toward a more efficient, transferable, and cognitively grounded paradigm of "atomic thinking".

Direct Preference Optimization (DPO) has recently been extended from text-only models to vision-language models. However, existing methods rely on oversimplified pairwise comparisons, generating a single negative image via basic perturbations or similarity-based retrieval, which fail to capture the complex nature of multimodal preferences, inducing optimization bias and hallucinations. To address this issue, we propose MISP-DPO, the first framework to incorporate multiple, semantically diverse negative images in multimodal DPO via the Plackett-Luce model. Our method embeds prompts and candidate images in CLIP (Contrastive Language-Image Pretraining) space and applies a sparse autoencoder to uncover semantic deviations into interpretable factors. Negative samples are selected based on reconstruction difficulty, semantic deviation from the positive, and mutual diversity, yielding broader and more informative supervision. To handle multi-negative comparisons, we adopt a Plackett-Luce objective and introduce an importance sampling strategy that improves training efficiency. Experiments across five diverse benchmarks demonstrate that MISP-DPO consistently improves multimodal alignment over prior methods, validating the effectiveness of semantic-aware, multi-negative sampling in preference-based learning.

Model merging, which combines multiple domain-specialized experts into a single model, offers a practical path to endow Large Language Models (LLMs) and Multimodal Large Language Models (MLLMs) with broad capabilities without the cost of joint training or serving many models. However, training-free methods rely on hand-tuned coefficients, whereas training-based methods primarily align parameters rather than downstream task behavior and typically treat all layers uniformly, ignoring inter-layer heterogeneity. We introduce Expert Merging, a training-light method that learns a small set of layer-wise coefficients using only unlabeled calibration data. The coefficients are optimized to explicitly align the merged model's hidden states and logits with those of the corresponding experts, with a coefficient regularizer for stability and task-weighted losses for controllable trade-offs. To capture inter-layer variation, Expert Merging++ augments this design with importance-guided chunking: a normalized layer-importance metric, derived from learned coefficients, task-vector magnitudes, and parameter counts, allocates more chunk-wise coefficients to high-importance layers while keeping low-importance layers lightweight. The result is a label-free, parameter-efficient, and scalable approach to multi-expert model merging across LLMs and MLLMs. Across MLLM backbones (InternVL and Qwen2-VL) and the LLM backbone (Mistral), our method surpasses strong training-free and training-based merging baselines, with Expert Merging++ delivering further gains and, in some cases, even exceeding supervised Mixture Training. The source code is available at https://github.com/Littleor/ExpertMerging.

Time-series question answering (TSQA) tasks face significant challenges due to the lack of labeled data. Alternatively, with recent advancements in large-scale models, vision-language models (VLMs) have demonstrated the potential to analyze time-series signals in a zero-shot manner. In this paper, we propose a training approach that uses pseudo labels generated by a VLM. Although VLMs can produce incorrect labels, TSQA models can still be effectively trained based on the property that deep neural networks are inherently robust to such noisy labels. Our experimental results demonstrate that TSQA models are not only successfully trained with pseudo labels, but also surpass the performance of the VLM itself by leveraging a large amount of unlabeled data.

Recent studies have shown that combining parameter-efficient fine-tuning (PEFT) with mixture-of-experts (MoE) is an effective strategy for adapting large language models (LLMs) to the downstream tasks. However, most existing approaches rely on conventional TopK routing, which requires careful hyperparameter tuning and assigns a fixed number of experts to each token. In this work, we propose LD-MoLE, a Learnable Dynamic routing mechanism for Mixture of LoRA Experts that enables adaptive, token-dependent, and layer-wise expert allocation. Our method replaces the non-differentiable TopK selection with a differentiable routing function and a closed-form solution. Moreover, our design allows the model to adaptively determine the number of experts to activate for each token at different layers. In addition, we introduce an analytical sparsity control objective to regularize the number of activated experts. Our method not only achieves superior performance, but also demonstrates the ability to learn token-dependent and layer-wise expert allocation. Large language models (LLMs) have demonstrated impressive capabilities across a wide range of natural language processing (NLP) tasks. However, their growing size requires significant computational resources for full-parameter fine-tuning. To address this, Parameter-Efficient Fine-tuning (PEFT) methods, such as Adapter-tuning (Houlsby et al., 2019) and LoRA (Hu et al., 2021), have emerged as crucial techniques for reducing training costs. Recently, the Mixture-of-Experts (MoE) design (Jacobs et al., 1991; Shazeer et al., 2017) has been successfully integrated into transformer feed-forward networks during LLMs pretraining (Dai et al., 2024; Y ang et al., 2025), demonstrating that MoE can reduce computational cost while maintaining strong performance.

Many studies have shown various biases targeting different demographic groups in language models, amplifying discrimination and harming fairness. Recent parameter modification debiasing approaches significantly degrade core capabilities such as text coherence and task accuracy. And Prompt-based debiasing methods, only effective for predefined trigger words, fail to address deeply embedded stereotypical associations in model parameters. In this paper, we propose BiasUnlearn, a novel model debiasing framework which achieves targeted debiasing via dual-pathway unlearning mechanisms coordinating stereotype forgetting with anti-stereotype retention, while preventing bias polarity reversal through adversarial forget set and dynamic dataset swapping. We conducted extensive experiments with multiple language models across various evaluation benchmarks. The results show that BiasUnlearn outperforms existing methods in mitigating bias in language models while retaining language modeling capabilities. Further experiments reveal that debiasing weights are transferable across model variants, confirming that bias representations become entrenched during pre-training and persist through fine-tuning phases.

Translating natural language questions into SQL has become a core challenge in enabling non-technical users to query databases. While recent work has explored large-scale synthetic data generation to improve model performance through post-training, most efforts emphasize cross-domain generalization. This leaves a gap for real-world enterprise scenarios, where models need to specialize to a single database schema and organizations require to be able to evaluate their Text-to-SQL systems on their own databases. To address this, we introduce SING-SQL, a fully automated two-stage framework for generating high-quality, high-coverage synthetic Text-to-SQL data for any target database, without relying on SQL logs or manual annotations. Our approach hierarchically partitions a database schema into sub-schemas, synthesizes SQL queries across multiple complexity levels, and applies a quality-aware pipeline that includes LLM-as-a-judge validation, executability checks, automatic repair, and column balancing. We further release SingSQL-LM, a family of compact language models fine-tuned on the synthetic data, achieving strong in-domain generalization. On the subset of the BIRD benchmark, SingSQL-LM-3B-R64 reaches 82.87% Soft F1 and 73.03% EX upper bound with 32 candidates, outperforming the best 3B-scale baseline by +16.21 in Soft F1 and +12.36 in EX. At the 1.5B scale, SingSQL-LM-1.5B-R64 improves over prior systems by +9.30 in Soft F1 and +4.49 in EX. On synthetic evaluation sets, SingSQL-LMs exceed prior systems by wide margins, establishing state-of-the-art performance among open models at comparable scales. Our study of context management strategies reveals that schema-free fine-tuning combined with schema-only inference provides the most robust results. These findings establish SING-SQL as a scalable, database-agnostic paradigm for producing and evaluating enterprise-grade Text-to-SQL systems.