Task vectors capture how a model changes during fine-tuning by recording the difference between pre-trained and task-specific weights. The composition of task vectors, a key operator in task arithmetic, enables models to integrate knowledge from multiple tasks without incurring additional inference costs. In this paper, we propose variational task vector composition, where composition coefficients are taken as latent variables and estimated in a Bayesian inference framework. Unlike previous methods that operate at the task level, our framework focuses on sample-specific composition. Motivated by the observation of structural redundancy in task vectors, we introduce a Spike-and-Slab prior that promotes sparsity and preserves only the most informative components. To further address the high variance and sampling inefficiency in sparse, high-dimensional spaces, we develop a gated sampling mechanism that constructs a controllable posterior by filtering the composition coefficients based on both uncertainty and importance. This yields a more stable and interpretable variational framework by deterministically selecting reliable task components, reducing sampling variance while improving transparency and generalization. Experimental results demonstrate that our method consistently outperforms existing approaches across all datasets by selectively leveraging the most reliable and informative components in task vectors. These findings highlight the practical value of our approach, establishing a new standard for efficient and effective task vector composition.

We present Qianfan-VL, a series of multimodal large language models ranging from 3B to 70B parameters, achieving state-of-the-art performance through innovative domain enhancement techniques. Our approach employs multi-stage progressive training and high-precision data synthesis pipelines, which prove to be critical technologies for enhancing domain-specific capabilities while maintaining strong general performance. Qianfan-VL achieves comparable results to leading open-source models on general benchmarks, with state-of-the-art performance on benchmarks such as CCBench, SEEDBench IMG, ScienceQA, and MMStar. The domain enhancement strategy delivers significant advantages in OCR and document understanding, validated on both public benchmarks (OCRBench 873, DocVQA 94.75%) and in-house evaluations. Notably, Qianfan-VL-8B and 70B variants incorporate long chain-of-thought capabilities, demonstrating superior performance on mathematical reasoning (MathVista 78.6%) and logical inference tasks. All models are trained entirely on Baidu's Kunlun P800 chips, validating the capability of large-scale AI infrastructure to train SOTA-level multimodal models with over 90% scaling efficiency on 5000 chips for a single task. This work establishes an effective methodology for developing domain-enhanced multimodal models suitable for diverse enterprise deployment scenarios.

Accurate modeling of ridesourcing mode choices is essential for designing and implementing effective traffic management policies for reducing congestion, improving mobility, and allocating resources more efficiently. Existing models for predicting ridesourcing mode choices often suffer from limited predictive accuracy due to their inability to capture key psychological factors, and are further challenged by severe class imbalance, as ridesourcing trips comprise only a small fraction of individuals' daily travel. To address these limitations, this paper introduces the Synthesizing Attitudes, Predicting Actions (SAPA) framework, a hierarchical approach that uses Large Language Models (LLMs) to synthesize theory-grounded latent attitudes to predict ridesourcing choices. SAPA first uses an LLM to generate qualitative traveler personas from raw travel survey data and then trains a propensity-score model on demographic and behavioral features, enriched by those personas, to produce an individual-level score. Next, the LLM assigns quantitative scores to theory-driven latent variables (e.g., time and cost sensitivity), and a final classifier integrates the propensity score, latent-variable scores (with their interaction terms), and observable trip attributes to predict ridesourcing mode choice. Experiments on a large-scale, multi-year travel survey show that SAPA significantly outperforms state-of-the-art baselines, improving ridesourcing choice predictions by up to 75.9% in terms of PR-AUC on a held-out test set. This study provides a powerful tool for accurately predicting ridesourcing mode choices, and provides a methodology that is readily transferable to various applications.

In this paper, we present the Scrapbook framework, a novel methodology designed to generate extensive datasets for probing the learned concepts of artificial intelligence (AI) models. The framework focuses on fundamental concepts such as object recognition, absolute and relative positions, and attribute identification. By generating datasets with a large number of questions about individual concepts and a wide linguistic variation, the Scrapbook framework aims to validate the model's understanding of these basic elements before tackling more complex tasks. Our experimental findings reveal that, while contemporary models demonstrate proficiency in recognizing and enumerating objects, they encounter challenges in comprehending positional information and addressing inquiries with additional constraints. Specifically, the MobileVLM-V2 model showed significant answer disagreements and plausible wrong answers, while other models exhibited a bias toward affirmative answers and struggled with questions involving geometric shapes and positional information, indicating areas for improvement in understanding and consistency. The proposed framework offers a valuable instrument for generating diverse and comprehensive datasets, which can be utilized to systematically assess and enhance the performance of AI models.

Humans can interpret geospatial information through natural language, while the geospatial cognition capabilities of Large Language Models (LLMs) remain underexplored. Prior research in this domain has been constrained by non-quantifiable metrics, limited evaluation datasets and unclear research hierarchies. Therefore, we propose a large-scale benchmark and conduct a comprehensive evaluation of the geospatial route cognition of LLMs. We create a large-scale evaluation dataset comprised of 36000 routes from 12 metropolises worldwide. Then, we introduce PathBuilder, a novel tool for converting natural language instructions into navigation routes, and vice versa, bridging the gap between geospatial information and natural language. Finally, we propose a new evaluation framework and metrics to rigorously assess 11 state-of-the-art (SOTA) LLMs on the task of route reversal. The benchmark reveals that LLMs exhibit limitation to reverse routes: most reverse routes neither return to the starting point nor are similar to the optimal route. Additionally, LLMs face challenges such as low robustness in route generation and high confidence for their incorrect answers. Code\ \&\ Data available here: \href{https://github.com/bghjmn32/EMNLP2025_Turnback}{TurnBack.}

With the advent of large language models (LLMs), numerous Post-Training Quantization (PTQ) strategies have been proposed to alleviate deployment barriers created by their enormous parameter counts. Quantization achieves compression by limiting the number of representable points in the data. Therefore, the key to achieving efficient quantization is selecting the optimal combination of representation points, or codes, for the given data. Existing PTQ solutions adopt two major approaches to this problem: Round-To-Nearest (RTN)-based methods and codebook-based methods. RTN-based methods map LLM weights onto uniformly distributed integer grids, failing to account for the Gaussian-like weight distribution of LLM weights. Codebook-based methods mitigate this issue by constructing distribution-aware codebooks; however, they suffer from random and strided memory access patterns, resulting in degraded inference speed that is exacerbated by the limited size of GPU L1 cache. To overcome these limitations, we propose a novel LLM quantization method, SBVR (Summation of BitVector Representation), that enables Gaussian-like code representation in a hardware-friendly manner for fast inference. SBVR maps weight values to non-uniform representation points whose distribution follows the actual distribution of LLM weights, enabling more accurate compression. Additionally, we design a custom CUDA kernel that allows matrix-vector multiplication directly in the SBVR format without decompression, thereby enabling high-performance execution of SBVR-compressed models. Our evaluations of SBVR on various models demonstrate state-of-the-art perplexity and accuracy benchmark performance while delivering a 2.21x- 3.04x end-to-end token-generation speedup over naive FP16 models in the 4-bit quantization regime.

Retrieval-Augmented Generation (RAG) enables large language models (LLMs) to access external knowledge sources, but the effectiveness of RAG relies on the coordination between the retriever and the generator. Since these components are developed independently, their interaction is often suboptimal: the retriever may return irrelevant or redundant documents, while the generator may fail to fully leverage retrieved evidence. In this work, we propose a process-supervised multi-agent framework to bridge the gap between retriever and generator. The framework introduces two lightweight agents: a Decision Maker, which determines when to continue retrieval or stop for answer generation, and a Knowledge Selector, which filters retrieved documents to retain only the most useful evidence. To provide fine-grained supervision, we employ an LLM-as-a-Judge that evaluates each intermediate action with process-level rewards, ensuring more accurate credit assignment than relying solely on final answer correctness. We further adopt a tree-structured rollout strategy to explore diverse reasoning paths, and train both agents with Proximal Policy Optimization (PPO) in an end-to-end manner. Experiments on single-hop and multi-hop question answering benchmarks show that our approach achieves higher accuracy, more stable convergence, and produces more interpretable reasoning trajectories compared with standard RAG baselines. Importantly, the proposed framework is modular and plug-and-play, requiring no modification to the retriever or generator, making it practical for real-world RAG applications.

With the rapid development of mobile communication technologies, future mobile networks will offer vast services and resources for commuting, production, daily life, and entertainment. Accurate and efficient forecasting of mobile data (e.g., cell traffic, user behavior, channel quality) helps operators monitor network state changes, orchestrate wireless resources, and schedule infrastructure and users, thereby improving supply efficiency and service quality. However, current forecasting paradigms rely on customized designs with tailored models for exclusive data types. Such approaches increase complexity and deployment costs under large-scale, heterogeneous networks involving base stations, users, and channels. In this paper, we design a foundation model for mobile data forecasting, MobiGPT, with a unified structure capable of forecasting three data types: base station traffic, user app usage, and channel quality. We propose a soft-prompt learning method to help the model understand features of different data types, and introduce a temporal masking mechanism to guide the model through three forecasting tasks: short-term prediction, long-term prediction, and distribution generation, supporting diverse optimization scenarios. Evaluations on real-world datasets with over 100,000 samples show that MobiGPT achieves accurate multi-type forecasting. Compared to existing models, it improves forecasting accuracy by 27.37%, 20.08%, and 7.27%, reflecting strong generalization. Moreover, MobiGPT exhibits superior zero/few-shot performance in unseen scenarios, with over 21.51% improvement, validating its strong transferability as a foundation model.

Diffusion large language models (dLLMs) have emerged as a new architecture following auto regressive models. Their denoising process offers a powerful generative advantage, but they present significant challenges in learning and understanding numerically sensitive mathematical and order-sensitive logical tasks. Current training methods, including pre-training, fine-tuning, and reinforcement learning, focus primarily on improving general knowledge retention and reasoning abilities, but lack a comprehensive understanding of mathematical and logical patterns. We propose DSFT, a simple yet effective Diffusion SFT strategy, by adjusting the masking strategy and loss function, guiding models to understand mathematical and logical patterns. This strategy can be flexibly combined with pre-training, reinforcement learning, and other training methods. Validated on models such as LLaDA and Dream series, we prove that DSFT on small-scale data can achieve improvements of 5-10% and approximately 2% on mathematical and logical problems, respectively. This inspiring masking approach offers insights for future learning of specific patterns, which can be easily and efficiently combined with other training methods and applied to various dLLMs. Our code is publicly available at https://anonymous.4open.science/r/DSFT-0FFB/

The capacity of Large Language Models (LLMs) to reason is fundamental to their application in complex, knowledge-intensive domains. In real-world scenarios, LLMs are often augmented with external information that can be helpful, irrelevant, or even misleading. This paper investigates the causal impact of such auxiliary information on the reasoning process of LLMs with explicit step-by-step thinking capabilities. We introduce SciAux, a new dataset derived from ScienceQA, to systematically test the robustness of the model against these types of information. Our findings reveal a critical vulnerability: the model's deliberative "thinking mode" is a double-edged sword. While helpful context improves accuracy, misleading information causes a catastrophic drop in performance, which is amplified by the thinking process. Instead of conferring robustness, thinking reinforces the degree of error when provided with misinformation. This highlights that the challenge is not merely to make models "think", but to endow them with the critical faculty to evaluate the information upon which their reasoning is based. The SciAux dataset is available at https://huggingface.co/datasets/billhdzhao/SciAux.