论文检索

Xiaotao Liu, Prashant Mohit, Arvind Easwaran

In safe reinforcement learning (SRL), there exists an inherent conflict between maximizing reward and minimizing cost. We propose a novel approach that effectively resolve the conflict between maximizing reward and minimizing cost in joint optimization.When the cost exceeds the threshold, we perform cost-reducing updates. Otherwise, we compute policy gradients that maximize expected rewards, while using second-order Taylor approximation to evaluate whether these reward-maximizing gradients would violate the cost constraint. If constraint violation is detected, we adjust the gradient direction to maintain safety compliance; otherwise, we execute standard reward-increasing policy updates. This approach helps ensure that reward-seeking updates do not inadvertently increase costs, thereby reducing the likelihood of constraint violations. Empirical tests show our framework successfully manages reward-cost trade-offs through reward augmentation and cost shaping, improving both performance and safety without switching optimization strategies. Results demonstrate that concurrent treatment of both objectives in one policy gradient update is viable for improving safe reinforcement learning methods.

Pengfei Liu, Rui Kong, Zongzhang Zhang

Hand-craft reward engineering requires domain knowledge with numerous trials and errors, while Preference-based Reinforcement Learning (PbRL) avoids manual reward design but often suffers from limited interpretability and unstable training. To address these issues, we propose a novel preference alignment framework. Our approach leverages large language models to generate sub-reward functions informed by prior knowledge and further align human preferences by optimizing the weights combining these sub-rewards. For policy learning, we introduce Policy Optimization via Pareto Regularization (POPR) which regularizes updates along Pareto-optimal directions. Experiments show that our framework improves reward quality and policy stability, achieving superior performance to expert-designed rewards across most tasks.

Lianghui Liu, Zongzhang Zhang

Offline Meta-Reinforcement Learning (OMRL) leverages pre-collected data to adapt to new tasks. Context-based methods learn task representations from contexts. However, the context is influenced by both the task and the behavior policy. The mismatch between the behavior policy and the testing policy causes a context distribution shift problem, which results in poor task representations and degraded performance. This problem is exacerbated in settings with data limitations. To address this, we propose a novel approach called Meta-Normalizing Flow (Meta-NF). First, it employs a highly expressive and sample-efficient normalizing flow policy. Second, it incorporates a metric for testing-time task representation selection to effectively mitigate the context shift problem. Empirical results demonstrate that Meta-NF outperforms existing OMRL methods, with both components contributing to its strong performance.

Jizhihui Liu, Guangdao Zhu, Feiyi Du

Vision-Language Models (VLMs) encode images into lengthy sequences of visual tokens, leading to excessive computational overhead and limited inference efficiency. In this paper, we study the hierarchical attention pattern in vision encoders and propose HiPrune, a training-free and model-agnostic token Pruning framework for VLMs. We identify that middle layers in the vision encoder attend to object-centric regions, while deep layers capture global contextual features. Based on this observation, HiPrune selects tokens based on the attention score from the middle and deep layers. Our method requires no retraining and integrates seamlessly with any ViT-based VLM. Experiments demonstrate that HiPrune achieves outstanding pruning performance, maintaining a balance between efficiency and efficacy.

Yow-Fu Liou, Yu-Chien Tang, An-Zi Yen

When evaluating large language models (LLMs) for question answering tasks, a common protocol is multiple-choice question-answering (MCQA), where the model selects from a fixed set of choices. In contemporary robustness testing, researchers typically perturb instructions or introduce confusion into factual statements; however, model behavior also hinges on choice compliance: whether models remain within the canonical set A-D. We formalize this setting by asking whether the model continues to respect the interface's rules when the problem presents a tempting alternative. Our approach is interface-preserving: we append a single selectable option E while keeping the question and A-D unchanged. Then, we introduce three types of malicious option injection to assess LLMs' robustness. Experimental results highlight the vulnerability of LLMs on contradict type content of the additional option E. Our evaluation framework can effectively serve as a low-cost audit of rule adherence on existing datasets and black-box models, surfaces off-policy items, and supports interpretable model comparison for deployment.

Kunrong Li, Zhu Sun, Kwan Hui Lim

We propose HyMoERec, a novel sequential recommendation framework that addresses the limitations of uniform Position-wise Feed-Forward Networks in existing models. Current approaches treat all user interactions and items equally, overlooking the heterogeneity in user behavior patterns and diversity in item complexity. HyMoERec initially introduces a hybrid mixture-of-experts architecture that combines shared and specialized expert branches with an adaptive expert fusion mechanism for the sequential recommendation task. This design captures diverse reasoning for varied users and items while ensuring stable training. Experiments on MovieLens-1M and Beauty datasets demonstrate that HyMoERec consistently outperforms state-of-the-art baselines.

Jingcheng Li, Ye Qiao, Sitao Huang

Current video understanding models struggle with temporal reasoning and efficient processing while balancing detail preservation with computational efficiency. We propose a hierarchical memory system that segments videos into action and scene units, combined with question-aware agentic keyframe selection. Our method achieves 70.3% overall accuracy on VideoMME short video benchmarks.

Yousung Lee, Dongsoo Har

Dynamic head pruning in Vision Transformers (ViTs) improves efficiency by removing redundant attention heads, but existing pruning policies are often difficult to interpret and control. In this work, we propose a novel framework by integrating Sparse Autoencoders (SAEs) with dynamic pruning, leveraging their ability to disentangle dense embeddings into interpretable and controllable sparse latents. Specifically, we train an SAE on the final-layer residual embedding of the ViT and amplify the sparse latents with different strategies to alter pruning decisions. Among them, per-class steering reveals compact, class-specific head subsets that preserve accuracy. For example, bowl improves accuracy (76%→82%) while reducing head usage (0.72→0.33) via heads h2 and h5. These results show that sparse latent features enable class-specific control of dynamic pruning, effectively bridging pruning efficiency and mechanistic interpretability in ViTs.

Younsung Lee, Yujin Kim, Kyungjoong Kim

Esports is growing rapidly, yet the data available to researchers is limited due to the game company policies. Consequently, vision-based approaches utilizing game screens are gaining attention as a practical alternative. We focus on the League of Legends minimap and address the challenges of champion detection when extracting champion information from the minimap. The challenges in this domain include small objects, rapid movement, and frequent occlusions. We propose a transfer-learning-based object detection pipeline that combines synthetic data with a subset of replay data. Synthetic data enables the rapid generation of diverse scenarios and improves training scalability, while replay data reduces the data distribution gap. This approach achieves 0.588 mean average precision, improving over replay-only by 0.261 and synthetic-only by 0.312, with 6.4 ms latency. Furthermore, we constructed a dataset encompassing all champions, enabling comparative analysis of detection models and supporting reproducible benchmarking for various application studies.

Tae Hoon Lee, Joongheon Kim

In adaptive cruise control (ACC), balancing safety, comfort, and sustainability still remains challenging. Accordingly, we propose a hybrid reinforcement learning framework combining proximal policy optimization (PPO) and deep Q-network (DQN) with a multi-objective reward for autonomous carbon-neutral eco-driving. Experimental results revealed the contrasts between eco and non-eco modes, underscoring how reward design shapes driving behaviors.

Suwoong Lee, Seungjae Lee, Yunho Jeon, Junmo Kim

We propose Tailored ViT Slimming (TVS), a budget-aware multi-dimensional pruning framework for Vision Transformers. TVS injects learnable masks into MHSA and MLP modules and applies adaptive non-convex sparsity regularization to achieve maximal utilization of parameters under strict module-wise budgets. In addition, by retaining scaled masks after pruning, TVS avoids abrupt accuracy drops and provides stable initialization for fine-tuning. On ImageNet-1k with DeiT-S and DeiT-B, TVS consistently outperforms prior ViT compression methods. This result empirically shows that the non-convex sparsity regularizer is effective not only in CNNs but also in ViTs.

Seung Kwon Lee, Seok Bin Son, Joongheon Kim, Hoh Peter In

Quantum machine learning (QML) has attracted growing interest for their ability to achieve superior performance with significantly fewer parameters. However, the high cost and scarcity of current hardware push inference to cloud-hosted quantum devices, creating a tension between verifiability and confidentiality. This work proposes a novel framework that converts quantum neural network operations into classical arithmetic circuits that faithfully approximate genuine quantum computations. By encrypting these circuits with zero-knowledge proofs, it ensures computational validity while concealing internal parameters. Experimental results show that our classical circuits achieve fidelity above 0.9996 and total variation distance below 1% compared to actual quantum computations, verifying the practicality of trustworthy and privacy-preserving quantum inference.

Vishesh Kumar, Ahana Chanda, Poulomi Bhattacharya, Akshay Agarwal

Automated mosquito species identification is critical for combating vector-borne diseases. We introduce Q-MoFusion, a novel hybrid quantum-classical framework that fuses deep features from pre-trained Audio Spectrogram Transformer (AST) and Whisper models using a Variational Quantum Circuit (VQC). Our approach significantly outperforms individual backbones and prior state-of-the-art benchmarks, demonstrating superior accuracy and robustness, particularly on imbalanced classes. Q-MoFusion demonstrates the potential of hybrid quantum computing to enhance bioacoustic surveillance for addressing critical public health challenges.

Daiki Kitashima, Ryota Higa, Katsuhide Fujita

Automated negotiation, a form of interaction among autonomous agents, plays a central role in multi-agent systems, yet the application of large language model (LLM) in this domain remains underexplored. An LLM can serve as a meta-strategist, adaptively selecting explicit strategies for execution by external strategic tools based on its capabilities. We propose a negotiation AI agent equipped with explicit strategic tools, including time-dependent and tit-for-tat negotiation strategies. Our results show that strategic tool enhanced negotiators achieve approximately 16% higher average utility compared with baseline, latest LLM negotiators.

Ungsik Kim, Minjae Lee, Jiho Bae, Minje Kim, Sang-Min Choi, Suwon Lee

We propose a framework for privacy-preserving argumentative explanations using homomorphic encryption. This method applies the Cheon-Kim-Kim-Song scheme, along with a soft k-means adapted for encrypted computation, to generate explanations without exposing sensitive data. By leveraging GPU acceleration, speedups of approximately 470–670 times were achieved compared with CPU execution. Experimental results show that explanation fidelity is maintained for small- to medium-scale models, whereas significant degradation occurs in larger models. These findings suggest that our study provides an initial step toward enabling secure and trustworthy argumentative explanations under encryption while also highlighting the challenges that remain for generalizability to more complex models.

Suin Kim, Jihoon Lee, Moonsung Kang, Doheun Cha, Sangtae Ahn

Recent advances in diffusion-based models have significantly broadened their scope, extending well beyond image generation to encompass zero-shot segmentation tasks. In this work, we introduce a novel, training-free approach that harnesses both self- and cross-attention maps to achieve highly detailed hair segmentation. Our method demonstrates remarkable efficacy in producing fine-grained results without the need for additional training.

Min Kim, Ohyun Jo

The consideration of channel correlation is crucial for improving the performance of multivariate time series forecasting. However, existing models fail to capture it in homogeneous and highly correlated channels. In this work, we introduce CompRestacking (Compression Restacking), a strikingly intuitive and effective method to address this problem. The approach consists of three main components: (1) PCC-Restacking for correlation-aware channel ordering, (2) Temporal embedding for time encoding, and (3) Aggregation compression for compact token generation. CompRestacking consistently outperforms in experiment results. The results demonstrate that CompRestacking leverages strong channel correlations for improved performance.

Chun-Wei Kang, Chung-Chi Chen, An-Zi Yen

Large Language Models (LLMs) demonstrate strong capabilities in code generation but often lack adaptability in planning and refinement. We propose Self-PR, a framework that integrates adaptive plan selection and iterative repair to improve correctness and generalization. Self-PR constructs a reusable plan database via task clustering and trains a selector to choose task-specific strategies. Incorrect outputs are refined through multi-round feedback until correctness. Trained only on HumanEval, Self-PR generalizes well to out-of-distribution tasks (MBPP), improving pass@1 by +4.9% on HumanEval and +5.5% on MBPP compared to Modularization-of-Thought prompting. Experiments across Llama-3 (8B, 70B) and GPT-4o-mini confirm robustness and scalability. These findings suggest that adaptive planning and feedback-driven repair are essential for reliable LLM-based code generation.

Jun Jang, Thai Le, Bo Wang

With the rise of generative artificial intelligence (GenAI), academic dishonesty in classrooms has skyrocketed, yet the existing solutions for detecting such dishonesty often fall short. Standard "AI detectors" merely analyze one text at a time, failing to account for students' previous writings, which risks erroneous predictions. Meanwhile, existing token-based authorship verification (AV) models fail to analyze the nuances in writing styles that truly distinguish authorship. To fill this existing gap, we propose a novel AV framework that combines token-level stylometric features (e.g., POS tag patterns) with handcrafted stylistic features (e.g., sentence structure variation) to construct a comprehensive feature set. Using both benchmark corpora and real-world high school student essays, we trained multiple machine learning classifiers using the proposed feature set. Our initial experiments show that our approach outperforms the standard token-only baselines by over 25%, while offering interpretable, style-based insights. These preliminary results highlight the importance of nuanced stylistic features and suggest that a holistic AV system can provide educators with more reliable and transparent detection tools. Looking ahead, we plan to extend this work with large language models and multi-agent approaches to further enhance robustness and adaptability.

Ritika Jain, Atul Kumar, Akshay Agarwal

Although deep networks excel on RGB images, their performance degrades sharply under severe domain shifts—such as sketch recognition, where color and texture cues are missing. In this work, we propose a novel pipeline that leverages semantic cues extracted from sketches to guide the synthesis of photorealistic RGB images using diffusion-based generative models. Our framework operates by extracting two crucial cues from the input sketch: semantic captions via the BLIP model and structural outlines via Canny edge detection. These cues are then integrated using ControlNet to guide a Stable Diffusion model, ensuring the synthesized RGB image is both semantically consistent with the content and structurally faithful to the original sketch. We evaluated our synthesized images by benchmarking classification performance. We trained standard architectures (from convolutional to transformer-based) on Tiny-ImageNet subsets and tested them on sketches, their synthesized counterparts, and the original RGB images. Experimental results demonstrate that our approach produces realistic, identity-preserving images, which significantly improve classification accuracy and effectively bridge the semantic gap. While BLIP-based captioning and ControlNet-guided diffusion are established methods, our contribution lies in their integration into a unified, caption-guided pipeline that enhances sketch-to-RGB translation with improved semantic consistency. The proposed method generalizes well across architectures, providing a scalable and cost-efficient solution for sketch-based image synthesis.