Chain-of-Thought (CoT) prompting enhances the reasoning capabilities of large language models (LLMs) by incorporating intermediate reasoning steps. Including demonstrations with reasoning further improves performance. However, manually creating such demonstrations requires expertise and significant human effort. Automating this process remains a key challenge. In this paper, we introduce LLM-adaptive reasoning clustering for chain-of-thought (AdaRC-CoT) to construct effective prompts automatically. AdaRC-CoT selects demonstrations with diverse reasoning patterns to guide LLMs, outperforming strategies focused on diversifying problem types. Specifically, it employs zero-shot CoT to generate reasoning paths for training queries, clusters them into k clusters, and sorts them by sample size. From each cluster, the sample closest to the center that meets the specified criteria is selected as an in-context demonstration to prompt LLMs for test set questions. AdaRC-CoT achieves state-of-the-art performance, securing the highest rank in 16 evaluations and the second-highest rank in 8 evaluations across 27 assessments involving three LLM backbones and nine datasets. Extensive ablation studies comprehensively validate the effectiveness of our method.

Transferability estimation is pivotal for the judicious selection of appropriate pre-trained models in the context of downstream target tasks, given that fine-tuning all candidate models is computationally prohibitive. Recent studies primarily center on evaluating the discriminative power of static features. However, these approaches often fail to effectively capture the dynamic evolution of model representations throughout the fine-tuning process. To address this limitation, we propose a novel and effective method dubbed DFATran (Dynamic Feature Adjustment Transferability Estimation). DFATran leverages an adversarial feature perturbation mechanism to generate perturbations specifically targeting the decision boundary, thereby effectively simulating the dynamic characteristics of the fine-tuning process. Concurrently, DFA-Tran is further designed to mitigate the correlation within intra-class features and suppress potential spurious noise, fostering features that better conform to the characteristics of adaptive representations. Extensive experimental results unequivocally demonstrate the superior performance of our proposed method, achieving an increase of up to 20.72% in the average τw across common benchmarks.

ReAct-based Large Language Model (LLM) agents experience significant inefficiencies due to redundant LLM calls for predictable reasoning steps. Although a static inertia mechanism has been introduced to reuse actions, its fixed threshold cannot adapt to dynamic task difficulties. To address this limitation, we propose AdaInertia, modeling tool usage inertia control as a reinforcement learning problem for the first time. We integrate a lightweight Deep Q-Network (DQN) meta-controller to adaptively route decisions among deep reasoning, conservative inertia, and hint-enhanced fallback based on real-time task progress and token efficiency. Experiments demonstrate that AdaInertia mitigates the standard performance-cost trade-off, reducing token consumption by up to 38.3

This paper studies Social Law Synthesis (SLS) in strategic multi-agent environments as a new multi-unit mechanism design problem. We model SLS as a Bayesian single-parameter procurement auction based on Alternating-time Temporal Logic (ATL) and aim to design a truthful, individually rational, and profit-optimal mechanism. We first prove a representation lemma showing that any valuation respecting alternating bisimulation can be compactly expressed as a feature set of ATL formulae. We then reduce payment determination to allocation determination in polynomial time, resolving the irregular payment issue inherent in multi-unit settings. We further show that allocation determination is FP^NP-complete and encode ATL semantics into integer linear programming (ILP) constraints to make the problem tractable with standard solvers. Based on these results, we present the 𝒫𝒪-𝒜𝒮ℒ mechanism, which is incentive-compatible, individually rational, and maximizes expected profit. Theoretical guarantees and examples confirm that our approach provides an effective and computationally feasible solution for synthesizing optimal social laws under strategic agent behavior.

Social Law Synthesis (SLS) in strategic environments is a novel multi-unit mechanism design problem, spanning modeling to computational challenges. We derive a method to specify the problem succinctly, reduce payment determination to allocation determination, and design an integer linear programming (ILP)-based algorithm that further reduces allocation to a polynomial-time ILP formulation. This offloads intractability to powerful ILP solvers, yielding a truthful, individually rational, and profit-optimal mechanism.

Multilingual LLM judges often assign different scores to equivalent answers across languages, compromising cross-lingual fairness. Post-hoc score calibration offers a natural remedy, but faces two key obstacles: high bias heterogeneity across tasks and languages, and black-box APIs that expose only scalar outputs. Existing calibration methods are mainly white-box, relying on internal states unavailable in practical deployments. To bridge this gap, we propose Risk-Bounded Score Calibration (RISC), a fully black-box post-hoc framework that uses only scalar scores. RISC derives a confidence signal from observed scores, performs confidence-aware adjustment, and activates calibration only when evidence is sufficiently reliable. This improves score alignment while limiting degradation on already well-calibrated cases. Across four evaluation tasks and eight languages, RISC significantly enhances cross-lingual fairness, reducing mean pairwise bias from 0.2839 to 0.1474 for Qwen2.5-7B-Instruct and from 0.2095 to 0.1264 for Llama-3-8B-Instruct, with few worsened language pairs. These results establish RISC as a practical solution for black-box multilingual score calibration, offering a pathway towards fairer LLM-based evaluation, especially for low-resource languages.

Out-of-distribution (OOD) detection is crucial for the robustness of machine learning in open-world scenarios. Most existing zero-shot OOD detection methods rely on the matching similarity between global visual features and in-distribution (ID) class prompts, and exhibit limited utilization of local features. To address this, we propose a post-hoc method—Adaptive Local Feature Mining(ALFM), which dynamically provides more critical local information to help distinguish ID samples from OOD samples. Experiments on the large-scale ImageNet-1K OOD benchmark demonstrate that ALFM can not only improve the hit rate of key local patches, but also effectively enhance OOD performance under both the zero-shot and few-shot settings.

The message-passing mechanism in Graph Neural Networks (GNNs) amplifies the negative effects of label noise, which undermines the robustness of GNNs. While recent methods aim to mitigate this issue, most rely on the homophily assumption, limiting their applicability to heterophilic graphs. In this paper, we conduct an empirical study on the ability of low-and high-pass losses to distinguish clean from noisy labeled nodes on both homophilic and heterophilic graphs. Motivated by our findings, we propose Homophily-Aware Loss Fusion (HALF), a novel method for robust learning under label noise. The key idea of HALF is to adaptively identify clean labeled nodes by combining low-and high-pass losses. Specifically, we first introduce a dual-structure constructor to generate homophilic and heterophilic graph structures from node features. Then, we propose a loss fusion that adaptively combines low-and high-pass losses, guided by node-level local ho mophily. Finally, clean and noisy labeled nodes are identified using the memorization effect based on the fused loss. To achieve more effective supervision for unlabeled nodes, we further propose a dual-pass alignment loss that aligns predictions from the low-and high-pass filters. We con duct extensive experiments across diverse homophilic and heterophilic graphs under label noise. The results demonstrate the superior performance and adaptability of HALF over state-of-the-art baselines.

The challenge of transforming partial point clouds into complete meshes still persists, with current methods facing issues like data accessibility constraint, shape preservation failure and poor robustness on real-scan data. Drawing inspiration from the structural information of objects to enhance the completion, we introduce an innovative weakly-supervised shape completion method leveraging structural decomposition without the necessity of SDFs during training. By representing objects as abstract structural frameworks and part details, our method initiates by forecasting the structure of the input partial point clouds, and individually restore each component through part decomposition completion and generation. Extracted part details are represented in images, which are porous and incomplete. Hence, we utilize a completion network to complete such details. For multiple results generation, a diffusion-based generation network is employed to generate a variety of details for the missing areas. The predicted structure and details are subsequently converted back into meshes, yielding the complete results. Since the details are depicted in images, our approach eliminates the need for SDFs during the training phase, achieving weakly-supervision. We conduct extensive comparisons on both artificial and real-scan datasets, demonstrating an average improvement of over 38.1% compared to the prior method, and achieving SOTA performance.

Large language models have shown promise in task planning, but existing methods struggle with three key limitations: implicit dependency modeling through text sequences, lack of systematic constraint verification, and inability to learn from execution failures. We propose a framework that addresses these challenges by combining explicit task graphs with learnable constraint rules. Our approach constructs directed acyclic graphs to represent task dependencies, verifies hard and soft constraints during action selection, and automatically extracts reusable constraint rules from failures. We evaluate on four diverse agent benchmarks: ALFWorld (household tasks), WebShop (online shopping), MiniWoB++ (web automation), and BabyAI (grid navigation). Our method achieves 84.2

Pretrained multimodal fusion models often suffer significant performance drops when faced with missing data due to sensor failures, transmission loss, or privacy constraints in real-world scenarios. Their complex architectures and large size render direct adaptation infeasible, positioning input-level imputation as a practical alternative. However, existing methods typically use fixed network structures and a single shared representation for generating different missing modalities, ignoring the dynamic and differentiated inter-modal correlations under varying missing scenarios. To address these deficiencies, we proposes a type-induced missing-aware mixture-of-experts architecture. The model dynamically routes observed modality features to specialized experts based on missing patterns and feature distributions, enabling adaptive inter-modal correlation extraction. It also employs learnable type-specific query vectors and a cross-attention mechanism to generate tailored representations for each modality, preserving crucial modality-specific information for reconstruction. Extensive experiments confirm the superior performance of the proposed approach in imputing missing data across diverse scenarios.

Graph Neural Networks (GNNs) have gained considerable prominence in semi-supervised learning tasks in processing graph-structured data, primarily owing to their message-passing mechanism, which largely relies on the availability of clean labels. However, in real-world scenarios, labels on nodes of graphs are inevitably noisy and sparsely labeled, significantly degrading the performance of GNNs. Exploring robust GNNs for semi-supervised node classification in the presence of noisy and sparse labels remains a critical challenge. Therefore, we propose a novel Graph Neural Network with Coarse- and Fine-Grained Division for mitigating label sparsity and noise, namely GNN-CFGD. The key idea of GNN-CFGD is reducing the negative impact of noisy labels via coarse- and fine-grained division, along with graph reconstruction. Specifically, we first investigate the effectiveness of linking unlabeled nodes to cleanly labeled nodes, demonstrating that this approach is more effective in combating labeling noise than linking to potentially noisy labeled nodes. Based on this observation, we introduce a Gaussian Mixture Model (GMM) based on the memory effect to perform a coarse-grained division of the given labels into clean and noisy labels. Next, we propose a clean labels oriented link that connects unlabeled nodes to cleanly labeled nodes, aimed at mitigating label sparsity and promoting supervision propagation. Furthermore, to provide refined supervision for noisy labeled nodes and additional supervision for unlabeled nodes, we fine-grain the noisy labeled and unlabeled nodes into two candidate sets based on confidence, respectively. Extensive experiments on various datasets demonstrate the superior effectiveness and robustness of GNN-CFGD.

Magnetic Resonance Imaging (MRI) enables comprehensive evaluation through the integration of multiple modalities, providing diverse information to enhance diagnostic and therapeutic decision-making. However, radiation exposure and restricted equipment availability often lead to modality missing, requiring effective solutions. Existing synthesis methods remain inflexible in accommodating variable numbers of missing modalities and insufficient utilization of cross-modal information correlations. To address these limitations, we propose the unified multi-modal aggregated-masked diffusion network (UMADN) for multi-modal MRI synthesis with missing modalities. UMADN leverages an aggregated mask module to enable unified synthesis, accommodating arbitrary modality missing. Moreover, the Cross-Modal Interaction block employs the attention mechanism to explore cross-modal interactions while following multi-modal spatial consistency. Experimental results demonstrate that UMADN achieves superior performance compared with other competing methods in handling various scenarios of modality missing.

Recent advancements in Large Language Models (LLMs) have catalyzed the exploration of Chain of Thought (CoT) approaches, particularly in extending their application to multimodal tasks to enhance reasoning capabilities. However, current studies often fail to fully leverage the inferential capabilities of these models, as they primarily focus on selecting similar questions or images rather than exploring identical inferences, thereby limiting the potential for analogical learning. In this paper, we propose an inference retrieval-augmented method that incorporates two strategies to identify training set examples with inferential processes similar to those of the target problem. Initially, we fine-tune a vanilla model to generate the pseudo-inference relevant to the question, which is then utilized to retrieve analogous examples. Next, we hypothesize that questions sharing similar contexts may share inferences, thus retrieving examples based on similarity in contextual questions. Ultimately, we construct multiple reasoning pathways from the retrieved examples and employ a voting mechanism to determine the most frequent answer. Our method surpasses all few-shot approaches and most supervised methods on the ScienceQA dataset, achieving an accuracy of 87.37% with ChatGPT and exceeding the human benchmark in several categories.

Learning with noisy labels (LNL) methods have enabled the deployment of machine learning systems with imperfectly labeled data. However, these methods often struggle to identify noise in the presence of long-tailed (LT) class distributions, where the memorization effect becomes class-dependent. Conversely, LT methods are suboptimal under label noise, as it hinders access to accurate label frequency statistics. This study aims to address the long-tailed noisy data by bridging the methodological gap between LNL and LT approaches. We propose a direct solution, termed Robust Logit Adjustment, which estimates ground-truth labels through label refurbishment, thereby mitigating the impact of label noise. Simultaneously, our method incorporates the distribution of training-time corrected target labels into the LT method logit adjustment, providing class-rebalanced supervision. Extensive experiments on both synthetic and real-world long-tailed noisy datasets demonstrate the superior performance of our method.

Understanding emotions in art paintings and generating comments about emotions is a highly challenging task due to their rich semantics and complex expressions. However, existing related methods tend to neglect the critical role of artistic emotion features in artworks at the visual level, while general vision language models(VLMs) lack a comprehensive grasp of art domain knowledge. To address this, we propose a knowledge-driven framework based on artistic emotion knowledge, which helps VLMs better comprehend emotions in artworks. The framework primarily comprises the artistic-emotion visual tower and the ANP cross-attention module. Specifically, the artistic-emotion visual tower introduces additional emotion tokens before visual features are fed into transformer layers. It integrate low-level art features into emotion tokens and infuses high-level art features into emotion tokens in the penultimate transformer layer. Moreover, the attention mechanism of visual tower is optimized by emotion bias, which enhances the model’s sensitivity to art features related to emotions. The ANP cross-attention module will extract affective entities directly from paintings as adjective-noun pairs and apply cross-attention with visual tower outputs. We have conducted extensive experiments on Artemis 1.0 and Artemis 2.0 datasets. The results demonstrate that our framework effectively improves the performance of VLMs in the task of understanding artistic emotions on eight metrics, outperforming existing methods.

Few-shot out-of-distribution (OOD) intent detection aims to detect in-distribution (ID) intents and reject OOD intents with few ID training data. This data scarcity often causes traditional OOD intent detection models to rely on spurious correlations, leading to attention misallocation, where the model focuses on irrelevant features instead of capturing meaningful, intent-specific information. To address these challenges, we propose LiVaS, a novel framework designed for few-shot OOD intent detection. LiVaS enhances label semantics through label interpretation, generates diverse ID samples using a variational autoencoder (VAE), and synthesizes OOD samples to improve decision boundaries. By integrating supervised contrastive learning (SCL) with cross-entropy (CE) loss, the framework encourages the learning of intent-specific features and reduces attention misallocation. Experimental results demonstrate that LiVaS achieves an average improvement of 6% over state-of-the-art methods on two widely used OOD intent detection datasets.

Human-machine collaboration is a promising training framework aimed at learning optimal strategies in high-cost exploration scenarios. However, such work is challenging. On one hand, current research on human-machine collaboration primarily focuses on imitation learning, overlooking the optimization of interactions between the collaborative entities. Hence, we propose a conceptual framework and modeling approach for collaborative learning based on imitation learning. On the other hand, the difficulty lies in explaining the contributions of humans and machines in the learning process and the lack of metrics for measuring the learned strategies and the uncertainty of decision gradients. To address these issues, we introduce an RL (Reinforcement Learning) framework for human-machine collaboration, known as Human-Machine RL. This framework employs reward shaping techniques for offline policy learning. In order to assess the policies, we design an estimation algorithm tailored for human-machine collaboration scenarios, based on reinforcement learning. Additionally, we incorporate Shapley as a mathematical interpretive tool for policy rewards. We tackle the issue of gradient variance that may arise from Shapley. The feasibility of our approach is theoretically demonstrated, and we have made the source code available for result reproducibility.

Graph Neural Networks (GNNs) have demonstrated remarkable ability in semi-supervised node classification. However, most existing GNNs rely heavily on a large amount of labeled data for training, which is labor-intensive and requires extensive domain knowledge. In this paper, we first analyze the restrictions of GNNs generalization from the perspective of supervision signals in the context of few-shot semi-supervised node classification. To address these challenges, we propose a novel algorithm named NormProp, which utilizes the homophily assumption of unlabeled nodes to generate additional supervision signals, thereby enhancing the generalization against label scarcity. The key idea is to efficiently capture both the class information and the consistency of aggregation during message passing, via decoupling the direction and Euclidean norm of node representations. Moreover, we conduct a theoretical analysis to determine the upper bound of Euclidean norm, and then propose homophilous regularization to constraint the consistency of unlabeled nodes. Extensive experiments demonstrate that NormProp achieve state-of-the-art performance under low-label rate scenarios with low computational complexity.