Quantization is an effective technique to reduce the deployment cost of large language models (LLMs), and post-training quantization (PTQ) has been widely studied due to its efficiency. However, existing PTQ methods are limited by their inability to fine-tune model parameters and often suffer significant accuracy loss in low-bit scenarios. Quantization-aware training (QA T) provides a more principled solution, but its reliance on backpropagation incurs prohibitive memory costs, limiting its practicality for LLM deployment. To address these challenges, we propose ZeroQA T, a zeroth-order optimization-based QA T framework that supports both weight and activation quantization. ZeroQA T leverages forward-only gradient estimation to eliminate backpropagation, substantially reducing computational and memory overhead while retaining the benefits of end-to-end optimization. We further introduce a lightweight variant of ZeroQA T for quantized fine-tuning, which freezes and pre-quantizes most parameters to further cut memory usage. Experiments show that ZeroQA T consistently outperforms representative PTQ and QA T baselines while requiring significantly less memory. For example, ZeroQA T enables fine-tuning of a 13B model at extremely low bit-widths (e.g., 2-4 bits) on a single 8GB GPU, and even allows fine-tuning a 6.7B model on a OnePlus 12 smartphone, demonstrating its practicality for end-to-end QA T on resource-limited edge devices. Large language models (LLMs) have emerged as essential tools for advancing natural language understanding and generation, driving progress in both research and industrial applications (Y ang et al., 2019; Liu et al., 2019; Talmor et al., 2018; Chowdhery et al., 2023; Zheng et al., 2020). Despite their transformative potential, training and deploying these models incur extremely high computational and memory costs. Such requirements not only constrain accessibility and scalability but also limit practicality in resource-constrained environments, including mobile and edge devices, embedded systems, and even enterprise servers with strict hardware or budget limitations (Zeng et al., 2024; Chen et al., 2024a; Tan et al., 2025). To address these challenges, model compression has been widely studied, with quantization being one of the most effective and indispensable techniques for deployment.

Time series forecasting plays a significant role in finance, energy, meteorology, and IoT applications. Recent studies have leveraged the generalization capabilities of large language models (LLMs) to adapt to time series forecasting, achieving promising performance. However, existing studies focus on token-level modal alignment, instead of bridging the intrinsic modality gap between linguistic knowledge structures and time series data patterns, greatly limiting the semantic representation. To address this issue, we propose a novel Semantic-Enhanced LLM (SE-LLM) that explores the inherent periodicity and anomalous characteristics of time series to embed into the semantic space to enhance the token embedding. This process enhances the interpretability of tokens for LLMs, thereby activating the potential of LLMs for temporal sequence analysis. Moreover, existing Transformer-based LLMs excel at capturing long-range dependencies but are weak at modeling short-term anomalies in time-series data. Hence, we propose a plugin module embedded within self-attention that models long-term and short-term dependencies to effectively adapt LLMs to time-series analysis. Our approach freezes the LLM and reduces the sequence dimensionality of tokens, greatly reducing computational consumption. Experiments demonstrate the superiority performance of our SE-LLM against the state-of-the-art (SOT A) methods.

The conversational capabilities of large language models hold significant promise for enabling scalable and interactive tutoring. While prior research has primarily examined their ability to generate Socratic questions, it often overlooks a critical aspect: adaptively guiding learners in accordance with their cognitive states. This study moves beyond question generation to emphasize instructional guidance capability. We ask: Can LLMs emulate expert tutors who dynamically adjust strategies in response to learners' states? To investigate this, we propose GuideEval, a benchmark grounded in authentic educational dialogues that evaluates pedagogical guidance through a three-phase behavioral framework: (1) Perception, inferring learner states; (2) Orchestration, adapting instructional strategies; and (3) Elicitation, stimulating proper reflections. Empirical results indicate that existing LLMs often fail to provide effective adaptive scaffolding when learners experience confusion or require redirection. To complement the quantitative evaluation, we conduct a detailed failure case analysis, providing an intuitive understanding of these shortcomings. Furthermore, we introduce a behavior-guided finetuning strategy that leverages behavior-prompted instructional dialogues, substantially enhancing guidance performance. By shifting the focus from isolated content evaluation to learner-centered state-aware interaction, our work advocates a more dialogic paradigm for evaluating Socratic LLMs.

Large Language Models (LLMs) are widely deployed in reasoning, planning, and decision-making tasks, making their trustworthiness critical. A significant and underexplored risk is intentional deception, where an LLM deliberately fabricates or conceals information to serve a hidden objective. Existing studies typically induce deception by explicitly setting a hidden objective through prompting or fine-tuning, which may not reflect real-world human-LLM interactions. Moving beyond such human-induced deception, we investigate LLMs' self-initiated deception on benign prompts. To address the absence of ground truth, we propose a framework based on Contact Searching Questions (CSQ). This framework introduces two statistical metrics derived from psychological principles to quantify the likelihood of deception. The first, the Deceptive Intention Score, measures the model's bias toward a hidden objective. The second, the Deceptive Behavior Score, measures the inconsistency between the LLM's internal belief and its expressed output. Evaluating 16 leading LLMs, we find that both metrics rise in parallel and escalate with task difficulty for most models. Moreover, increasing model capacity does not always reduce deception, posing a significant challenge for future LLM development. Evaluating the trustworthiness of Large Language Models (LLMs) has become critical as LLMs such as ChatGPT (Achiam et al., 2023) and Gemini (Team et al., 2023) are woven into systems for reasoning, planning, and decision-making. Beyond well-studied failures like hallucination (Filip-pova, 2020) and bias (Navigli et al., 2023) that reflect mistaken beliefs or skewed outputs, a more consequential threat is deception: a model may strategically fabricate statements that it knows to be false in service of a hidden objective, as exemplified in Figure 1. LLM deception can arise in two settings: (1) an incentivizing prompt is given, and the model lies to satisfy the objective specified in the prompt (see Figure 1); (2) a benign prompt is given, yet the model lies due to its intrinsic objective. Most existing studies focus on the incentivizing prompt: for example, Ward et al. (2023) explicitly prompt LLMs to generate deceptive content, and V an Der Weij et al. (2024) fine-tune LLMs to intentionally underperform on specified tasks.

While generalist foundation models like Gemini and GPT-4o demonstrate impressive multi-modal competence, existing evaluations fail to test their intelligence in dynamic, interactive worlds. Static benchmarks lack agency, while interactive benchmarks suffer from a severe modal bottleneck, typically ignoring crucial auditory and temporal cues. To bridge this evaluation chasm, we introduce OmniPlay, a diagnostic benchmark designed not just to evaluate, but to probe the fusion and reasoning capabilities of agentic models across the full sensory spectrum. Built on a core philosophy of modality interdependence, OmniPlay comprises a suite of five game environments that systematically create scenarios of both synergy and conflict, forcing agents to perform genuine cross-modal reasoning. Our comprehensive evaluation of six leading omni-modal models reveals a critical dichotomy: they exhibit superhuman performance on high-fidelity memory tasks but suffer from systemic failures in challenges requiring robust reasoning and strategic planning. We demonstrate that this fragility stems from brittle fusion mechanisms, which lead to catastrophic performance degradation under modality conflict and uncover a counter-intuitive "less is more" paradox, where removing sensory information can paradoxically improve performance. Our findings suggest that the path toward robust AGI requires a research focus beyond scaling to explicitly address synergistic fusion. Our platform is available for anonymous review at https://github.com/fuqingbie/omni-game-benchmark.

The generative capabilities of Large Language Models (LLMs) are rapidly expanding from static code to dynamic, interactive visual artifacts. This progress is bottlenecked by a critical evaluation gap: established benchmarks focus on algorithmic correctness and are blind to the visual fidelity and interactive integrity that define modern user experiences. To bridge this gap, we introduce ArtifactsBench, a new benchmark and paradigm for the automated, multimodal evaluation of visual code generation. Our framework programmatically renders each generated artifact and captures its dynamic behavior through temporal screenshots. This visual evidence, alongside the source code, is then assessed by a Multimodal LLM (MLLM)-as-Judge, which is rigorously guided by a fine-grained, per-task checklist to ensure holistic and reproducible scoring. We construct a new benchmark of 1,825 diverse tasks and evaluate over 30 leading LLMs. Our automated evaluation achieves a striking 94.4% ranking consistency with WebDev Arena, the gold-standard for human preference in web development, and over 90% pairwise agreement with human experts. This establishes ArtifactsBench as the first framework to reliably automate the assessment of human-perceived quality at scale. Our analysis provides a high-resolution map of the current SOTA, revealing that generalist models often outperform domain-specific ones. We open-source ArtifactsBench, including the benchmark, evaluation harness, and baseline results at https://artifactsbenchmark.github.io/, to provide the community with a scalable and accurate tool to accelerate the development of user-centric generative models.

Personalized text generation has become crucial for adapting language models to diverse and evolving users' personal context across cultural, temporal, and contextual dimensions. While existing methods often rely on centralized fine-tuning or static preference alignment, they struggle to achieve real-time adaptation under resource constraints inherent to personal devices. This limitation creates a dilemma: large cloud-based models lack access to localized user-specific information, while small on-device models cannot match the generation quality of their cloud counterparts. To address this dichotomy, we present CoSteer, a novel collaborative framework that enables decoding-time personalization through localized delta steering. Our key insight lies in leveraging the logits difference between personal context-aware and -agnostic outputs from local small models as steering signals for cloud-based LLMs. Specifically, we formulate token-level optimization as an online learning problem, where local delta vectors dynamically adjust the remote LLM's logits within the on-device environment. This approach preserves privacy by transmitting only the final steered tokens rather than raw data or intermediate vectors, while maintaining cloud-based LLMs' general capabilities without fine-tuning. Through comprehensive experiments on various personalized generation tasks, we demonstrate that CoSteer effectively assists LLMs in generating personalized content by leveraging locally stored user profiles and histories, ensuring privacy preservation through on-device data processing while maintaining acceptable computational overhead.

Diagrams are widely used to visualize data in publications. The research field of data visualization deals with defining principles and guidelines for the creation and use of these diagrams, which are often not known or adhered to by researchers, leading to misinformation caused by providing inaccurate or incomplete information. In this work, large Vision Language Models (VLMs) are used to analyze diagrams in order to identify potential problems in regards to selected data visualization principles and guidelines. To determine the suitability of VLMs for these tasks, five open source VLMs and five prompting strategies are compared using a set of questions derived from selected data visualization guidelines. The results show that the employed VLMs work well to accurately analyze diagram types (F1-score 82.49 %), 3D effects (F1-score 98.55 %), axes labels (F1-score 76.74 %), lines (RMSE 1.16), colors (RMSE 1.60) and legends (F1-score 96.64 %, RMSE 0.70), while they cannot reliably provide feedback about the image quality (F1-score 0.74 %) and tick marks/labels (F1-score 46.13 %). Among the employed VLMs, Qwen2.5VL performs best, and the summarizing prompting strategy performs best for most of the experimental questions. It is shown that VLMs can be used to automatically identify a number of potential issues in diagrams, such as missing axes labels, missing legends, and unnecessary 3D effects. The approach laid out in this work can be extended for further aspects of data visualization.

Large Language Models (LLMs) have demonstrated remarkable performance across diverse tasks. Nevertheless, they still pose notable safety risks due to potential misuse for malicious purposes. Jailbreaking, which seeks to induce models to generate harmful content through single-turn or multi-turn attacks, plays a crucial role in uncovering underlying security vulnerabilities. However, prior methods, including sophisticated multi-turn approaches, often struggle to adapt to the evolving dynamics of dialogue as interactions progress. To address this challenge, we propose \ours (JailBreaking via \textbf{G}lobally \textbf{R}efining and \textbf{A}daptively \textbf{F}abricating), a novel multi-turn jailbreaking method that globally refines the attack trajectory at each interaction. In addition, we actively fabricate model responses to suppress safety-related warnings, thereby increasing the likelihood of eliciting harmful outputs in subsequent queries. Extensive experiments across six state-of-the-art LLMs demonstrate the superior effectiveness of our approach compared to existing single-turn and multi-turn jailbreaking methods. Our code will be released at https://github.com/Ytang520/Multi-Turn_jailbreaking_Global-Refinment_and_Active-Fabrication.

We investigate a general approach for improving user prompts in text-to-image (T2I) diffusion models by finding prompts that maximize a reward function specified at test-time. Although diverse reward models are used for evaluating image generation, existing automated prompt engineering methods typically target specific reward configurations. Consequently, these specialized designs exhibit suboptimal performance when applied to new prompt engineering scenarios involving different reward models. To address this limitation, we introduce RATTPO (Reward-Agnostic Test-Time Prompt Optimization), a flexible test-time optimization method applicable across various reward scenarios without modification. RATTPO iteratively searches for optimized prompts by querying large language models (LLMs) \textit{without} requiring reward-specific task descriptions. Instead, it uses the optimization trajectory and a novel reward-aware feedback signal (termed a "hint") as context. Empirical results demonstrate the versatility of RATTPO, effectively enhancing user prompts across diverse reward setups that assess various generation aspects, such as aesthetics, general human preference, or spatial relationships between objects. RATTPO surpasses other test-time search baselines in search efficiency, running 4.8 times faster than naive reward-agnostic test-time search baseline on average. Furthermore, with sufficient inference budget, it can achieve comparable performance to learning-based baselines that require reward-specific fine-tuning. The code is available at https://github.com/seminkim/RATTPO.