This study presents a comprehensive comparative evaluation of four state-of-the-art Large Language Models (LLMs)--Claude 3.7 Sonnet, DeepSeek-V3, Gemini 2.0 Flash, and GPT-4o--for sentiment analysis and emotion detection in Persian social media texts. Comparative analysis among LLMs has witnessed a significant rise in recent years, however, most of these analyses have been conducted on English language tasks, creating gaps in understanding cross-linguistic performance patterns. This research addresses these gaps through rigorous experimental design using balanced Persian datasets containing 900 texts for sentiment analysis (positive, negative, neutral) and 1,800 texts for emotion detection (anger, fear, happiness, hate, sadness, surprise). The main focus was to allow for a direct and fair comparison among different models, by using consistent prompts, uniform processing parameters, and by analyzing the performance metrics such as precision, recall, F1-scores, along with misclassification patterns. The results show that all models reach an acceptable level of performance, and a statistical comparison of the best three models indicates no significant differences among them. However, GPT-4o demonstrated a marginally higher raw accuracy value for both tasks, while Gemini 2.0 Flash proved to be the most cost-efficient. The findings indicate that the emotion detection task is more challenging for all models compared to the sentiment analysis task, and the misclassification patterns can represent some challenges in Persian language texts. These findings establish performance benchmarks for Persian NLP applications and offer practical guidance for model selection based on accuracy, efficiency, and cost considerations, while revealing cultural and linguistic challenges that require consideration in multilingual AI system deployment.

As large language models (LLMs) proliferate in scale, specialization, and latency profiles, the challenge of routing user prompts to the most appropriate model has become increasingly critical for balancing performance and cost. We introduce CARGO (Category-Aware Routing with Gap-based Optimization), a lightweight, confidence-aware framework for dynamic LLM selection. CARGO employs a single embedding-based regressor trained on LLM-judged pairwise comparisons to predict model performance, with an optional binary classifier invoked when predictions are uncertain. This two-stage design enables precise, cost-aware routing without the need for human-annotated supervision. To capture domain-specific behavior, CARGO also supports category-specific regressors trained across five task groups: mathematics, coding, reasoning, summarization, and creative writing. Evaluated on four competitive LLMs (GPT-4o, Claude 3.5 Sonnet, DeepSeek V3, and Perplexity Sonar), CARGO achieves a top-1 routing accuracy of 76.4% and win rates ranging from 72% to 89% against individual experts. These results demonstrate that confidence-guided, lightweight routing can achieve expert-level performance with minimal overhead, offering a practical solution for real-world, multi-model LLM deployments.

The rapid progress of Multi-Modal Large Language Models (MLLMs) has spurred the creation of numerous benchmarks. However, conventional full-coverage Question-Answering evaluations suffer from high redundancy and low efficiency. Inspired by human interview processes, we propose a multi-to-one interview paradigm for efficient MLLM evaluation. Our framework consists of (i) a two-stage interview strategy with pre-interview and formal interview phases, (ii) dynamic adjustment of interviewer weights to ensure fairness, and (iii) an adaptive mechanism for question difficulty-level chosen. Experiments on different benchmarks show that the proposed paradigm achieves significantly higher correlation with full-coverage results than random sampling, with improvements of up to 17.6% in PLCC and 16.7% in SRCC, while reducing the number of required questions. These findings demonstrate that the proposed paradigm provides a reliable and efficient alternative for large-scale MLLM benchmarking.

We propose Llama-Mimi, a speech language model that uses a unified tokenizer and a single Transformer decoder to jointly model sequences of interleaved semantic and acoustic tokens. Comprehensive evaluation shows that Llama-Mimi achieves state-of-the-art performance in acoustic consistency and possesses the ability to preserve speaker identity. Our analysis further demonstrates that increasing the number of quantizers improves acoustic fidelity but degrades linguistic performance, highlighting the inherent challenge of maintaining long-term coherence. We additionally introduce an LLM-as-a-Judge-based evaluation to assess the spoken content quality of generated outputs. Our models, code, and speech samples are publicly available.

As they become increasingly multilingual, Large Language Models (LLMs) exhibit more language confusion, i.e., they tend to generate answers in a language different from the language of the prompt or the answer language explicitly requested by the user. In this work, we propose ReCoVeR (REducing language COnfusion in VEctor Representations), a novel lightweight approach for reducing language confusion based on language-specific steering vectors. We first isolate language vectors with the help of multi-parallel corpus and then effectively leverage those vectors for effective LLM steering via fixed (i.e., unsupervised) as well as trainable steering functions. Our extensive evaluation, encompassing three benchmarks and 18 languages, shows that ReCoVeR effectively mitigates language confusion in both monolingual and cross-lingual setups while at the same time -- and in contrast to prior language steering methods -- retaining task performance. Our data code is available at https://github.com/hSterz/recover.

With the advancement of mid/post-training techniques, LLMs are pushing their boundaries at an accelerated pace. Legacy benchmarks saturate quickly (e.g., broad suites like MMLU over the years, newer ones like GPQA-D even faster), which makes frontier progress hard to track. The problem is especially acute in Korean: widely used benchmarks are fewer, often translated or narrow in scope, and updated more slowly, so saturation and contamination arrive sooner. Accordingly, at this moment, there is no Korean benchmark capable of evaluating and ranking frontier models. To bridge this gap, we introduce KAIO, a Korean, math-centric benchmark that stresses long-chain reasoning. Unlike recent Korean suites that are at or near saturation, KAIO remains far from saturated: the best-performing model, GPT-5, attains 62.8, followed by Gemini-2.5-Pro (52.3). Open models such as Qwen3-235B and DeepSeek-R1 cluster falls below 30, demonstrating substantial headroom, enabling robust tracking of frontier progress in Korean. To reduce contamination, KAIO will remain private and be served via a held-out evaluator until the best publicly known model reaches at least 80% accuracy, after which we will release the set and iterate to a harder version.

Multi-stage information retrieval (IR) has become a widely-adopted paradigm in search. While Large Language Models (LLMs) have been extensively evaluated as second-stage reranking models for monolingual IR, a systematic large-scale comparison is still lacking for cross-lingual IR (CLIR). Moreover, while prior work shows that LLM-based rerankers improve CLIR performance, their evaluation setup relies on lexical retrieval with machine translation (MT) for the first stage. This is not only prohibitively expensive but also prone to error propagation across stages. Our evaluation on passage-level and document-level CLIR reveals that further gains can be achieved with multilingual bi-encoders as first-stage retrievers and that the benefits of translation diminishes with stronger reranking models. We further show that pairwise rerankers based on instruction-tuned LLMs perform competitively with listwise rerankers. To the best of our knowledge, we are the first to study the interaction between retrievers and rerankers in two-stage CLIR with LLMs. Our findings reveal that, without MT, current state-of-the-art rerankers fall severely short when directly applied in CLIR.

Unified vision large language models (VLLMs) have recently achieved impressive advancements in both multimodal understanding and generation, powering applications such as visual question answering and text-guided image synthesis. However, progress in unified VLLMs remains constrained by the lack of datasets that fully exploit the synergistic potential between these two core abilities. Existing datasets typically address understanding and generation in isolation, thereby limiting the performance of unified VLLMs. To bridge this critical gap, we introduce a novel dataset construction framework, UnifiedVisual, and present UnifiedVisual-240K, a high-quality dataset meticulously designed to facilitate mutual enhancement between multimodal understanding and generation. UnifiedVisual-240K seamlessly integrates diverse visual and textual inputs and outputs, enabling comprehensive cross-modal reasoning and precise text-to-image alignment. Our dataset encompasses a wide spectrum of tasks and data sources, ensuring rich diversity and addressing key shortcomings of prior resources. Extensive experiments demonstrate that models trained on UnifiedVisual-240K consistently achieve strong performance across a wide range of tasks. Notably, these models exhibit significant mutual reinforcement between multimodal understanding and generation, further validating the effectiveness of our framework and dataset. We believe UnifiedVisual represents a new growth point for advancing unified VLLMs and unlocking their full potential. Our code and datasets is available at https://github.com/fnlp-vision/UnifiedVisual.

Single-cell foundation models (scFMs) have demonstrated state-of-the-art performance on various tasks, such as cell-type annotation and perturbation response prediction, by learning gene regulatory networks from large-scale transcriptome data. However, a significant challenge remains: the decision-making processes of these models are less interpretable compared to traditional methods like differential gene expression analysis. Recently, transcoders have emerged as a promising approach for extracting interpretable decision circuits from large language models (LLMs). In this work, we train a transcoder on the cell2sentence (C2S) model, a state-of-the-art scFM. By leveraging the trained transcoder, we extract internal decision-making circuits from the C2S model. We demonstrate that the discovered circuits correspond to real-world biological mechanisms, confirming the potential of transcoders to uncover biologically plausible pathways within complex single-cell models.

While reinforcement learning (RL) is increasingly used for LLM-based tool learning, its efficiency is often hampered by an overabundance of simple samples that provide diminishing learning value as training progresses. Existing dynamic sampling techniques are ill-suited for the multi-task structure and fine-grained reward mechanisms inherent to tool learning. This paper introduces Dynamic Sampling with Curriculum Learning (DSCL), a framework specifically designed to address this challenge by targeting the unique characteristics of tool learning: its multiple interdependent sub-tasks and multi-valued reward functions. DSCL features two core components: Reward-Based Dynamic Sampling, which uses multi-dimensional reward statistics (mean and variance) to prioritize valuable data, and Task-Based Dynamic Curriculum Learning, which adaptively focuses training on less-mastered sub-tasks. Through extensive experiments, we demonstrate that DSCL significantly improves training efficiency and model performance over strong baselines, achieving a 3.29\% improvement on the BFCLv3 benchmark. Our method provides a tailored solution that effectively leverages the complex reward signals and sub-task dynamics within tool learning to achieve superior results.