论文检索

Soumyajit Dev, Malay Bhattacharyya

Dismantling interdependent directed networks to obtain the largest mutually strongly connected component (MSCC) is an NP-hard problem. To address this, we propose a novel method, Disassembling Directed Interdependent Networks (DDIN), by synergizing Reinforcement Learning (RL) and Graph Neural Networks (GNN). We introduce asymmetric GNNs to capture the asymmetry of in/out-degree and multi-relational attention to model directed inter-layer dependencies, integrated with prioritized RL for efficient node selection in large action spaces. Our contributions include (i) a directed GraphSAGE encoder separating in/out aggregations for asymmetry, (ii) multi-relational attention fusing layer semantics, and (iii) sum-tree prioritized n-step Deep Q-Network (DQN) for efficient policy search. DDIN is evaluated on 5 directed multiplexes from biological, social, and economic domains, achieving 16-23% lower AUDC compared to known baseline heuristics.

Adit Desai, Sudipta Roy, Mohna Chakraborty

Vision-Language Models (VLMs) are increasingly deployed across downstream tasks, yet their training data often encode social biases that surface in outputs. Unlike humans, who interpret images through contextual and social cues, VLMs process them through statistical associations, often leading to reasoning that diverges from human reasoning. By analyzing how a VLM reasons, we can understand how inherent biases are perpetuated and can adversely affect downstream performance. To examine this gap, we systematically analyze social biases in five open-source VLMs for an occupation prediction task, on the FairFace dataset. Across 32 occupations and three different prompting styles, we elicit both predictions and reasoning. Our findings show that the biased reasoning patterns systematically underlie intersectional disparities, highlighting the need to align VLM reasoning with human values before downstream deployment.

Ankan Deria, Dwarikanath Mahapatra, Murari Mondal, Sudipta Roy

We present a language-based noise modulation module for diffusion models that improves image color generation under textual guidance. Unlike standard approaches that inject noise uniformly, our method leverages semantic cues from text to selectively control the noise injection process, preserving local details and enhancing color accuracy even when descriptions are ambiguous or incomplete. Applied to language guided image colorization, this targeted modulation leads to more faithful and visually consistent results. The proposed module is lightweight, generalizable, and can be integrated into existing diffusion pipelines, offering a simple yet effective step toward more controllable text-to-image generation.

Alex Davies, Roussel Nzoyem Ngueguin, Nirav Ajmeri, Telmo De Menezes E Silva Filho

We evaluate how well large language model embeddings represent continuous numerical values across different precisions and ranges. Using linear models and principal component analysis on models from major providers, we show that while embeddings can reconstruct numbers with high fidelity (R2 ≥ 0.95), they introduce substantial noise, with principal components explaining less than 40% of embedding variance. Performance degrades with increasing decimal precision and mixed-sign values, revealing fundamental limitations in how these models encode numerical information.

Shrutimoy Das

Dataset distillation methods learn a representative summary of the full dataset such that training on the distilled data is more efficient in terms of time and space. The current state-of-the-art methods exploit the correspondence between infinitely wide neural networks (NNs) and kernel ridge regression to design distillation methods that result in high-quality summaries of the data. In this work, we leverage the correspondence between infinitely wide networks and Gaussian Processes(GPs) for learning a distilled dataset. We investigate the feasibility of using the inducing points method for Gaussian Processes, as a data distillation method. While most of the existing dataset distillation methods are based on loss or gradient matching, our method looks at the function space approximation, facilitated by the NN-GP correspondence. Additionally, using recent theoretical results on GP regression and neural tangent kernels(NTKs), we also provide an upper bound on the size of the distilled data. We demonstrate the utility of inducing points as distilled data on a set of datasets empirically.

Pranshu Danani, Yash Bansal, Parshiv Kapoor

Deep learning is increasingly applied to intraoperative and surgical video analysis to enable real-time workflow recognition, and decision support for improved surgical precision. A key direction is modeling surgical activity as triplets of instrument, action, and target, which provide a richer representation of procedures. However, existing approaches often depend on bounding-box annotations or lack temporal context. We propose TWiST (Temporal Weakly Supervised Triplet detection), a framework that combines weakly supervised instrument localization, temporal attention for triplet prediction, and grounding of triplets with detected instruments. Our experiments show that TWiST outperforms prior weakly supervised baselines.

Shuting Dai, Yatu Ji, Yanli Wang, Lei Shi, Qing-Dao-Er-Ji Ren, Nier Wu, Na Liu

The scarcity of parallel corpora for Mongolian and Chinese constrains the performance of Mongolian-Chinese neural machine translation (NMT), particularly manifesting in inadequate accuracy in translating specialized terminology. To address this limitation, this study adopts a lexically constrained augmentation strategy that constructs pseudo-source sentences by appending Chinese constraint words to Mongolian source texts, while enforcing the inclusion of these constraints in the output to improve translation accuracy. However, this approach presents two inherent drawbacks: processing pseudo-sentences with a single encoder tends to induce semantic interference, while the introduced constraint words may exacerbate alignment errors during decoding. To overcome these limitations, this paper propose a Constraint-Augmented Mongolian-Chinese NMT method (CANMT) based on dynamic feedback alignment. The method employs a dual-encoder architecture to isolate bilingual representations, coupled with a dynamic feedback alignment module that progressively reduces alignment errors through iterative reffnement, thereby enhancing overall translation performance.

Yeryeong Cho, Joongheon Kim

End-to-end (E2E) autonomous driving must maintain global consistency while preserving local precision. However, existing E2E approaches rarely achieve both goals simultaneously. Therefore, we propose a multimodal coarse-to-local transformer (MC2L-Transformer), which is composed of a hierarchical transformer architecture. Multimodal inputs are fused into a shared embedding, and global waypoints are produced. Local refinement is then utilized to capture fine interactions around the vehicle. Furthermore, a temporal encoder summarizes recent context, and navigation target and velocity are embedded to guide route- and speed-aware decoding. We evaluate in CARLA, and the results show lower collision and off-route rates even under sudden events. These results indicate that combining a coarse-to-local hierarchical transformer with a lightweight temporal context provides a practical step toward reliable E2E autonomous driving.

Fengxiang Cheng, Chuan Zhou, Fenrong Liu, Robert Van Rooij

Large language models (LLMs) have achieved impressive progress in natural language processing tasks but still struggle with complex logical reasoning. We observe that in propositional logic question-answering (QA), LLMs' performance varies with the order of training samples during fine-tuning. Motivated by this, we propose a data-driven approach to automatically determine the fine-tuning sample order, enhancing the logical QA performance of LLMs. Specifically, we first quantify the logical reasoning complexity of propositional reasoning samples and then stratify the training data into several subsets of ascending complexity. Subsequently, we fine-tune the LLMs on these subsets, progressing from low to high reasoning complexity. Experimental results demonstrate that our approach outperforms single-stage fine-tuning baselines across diverse reasoning benchmarks.

Hanxiao Chen

Large Language Models (LLMs) have revolutionized the modern society significantly with the numerous advanced interactions between humans and AI agents, whereas the usage of most large language models including ChatGPT are not friendly open-sourced and must require users paying a lot for such AI service continuously. Therefore, deploying open-sourced large language models on local servers can be considered as an efficient method to design and implement creative embodied AI algorithms with lower cost and more stable free usage. Inspired by this ordinary motivation, we originally propose and implement the “Socratic Models-ChatGLM”, which is a well-performed algorithm for multi-modal interactive control of robotic arm based on offline large language models via the facile PyBullet platform, even presents extraordinary potential to address complicated text-image multi-step long-horizon robotic manipulation tasks.

Krishi Chawda, Jean Erik Delanois, Giri P Krishnan, Maksim Bazhenov

One of the central challenges in deep learning is that models trained on new tasks often overfit and lose the ability to generalize. This issue arises because gradient descent often converges to solutions in regions of the loss landscape that are sharp near their minima. High sharpness leads to rapid performance loss when test data are perturbed or statistically shifted. Although sharpness has been linked to generalization, few methods directly target it to improve generalization. Here we demonstrate that an unsupervised, sleep-like replay algorithm identifies low loss regions with lower sharpness leading to improvement in generalization to distortions, including Gaussian and salt-and-pepper noise. Our study identifies loss-function sharpness as a unifying measure for generalizable learning and robustness, and points to new principles for designing resilient AI systems.

Vashu Chauhan, Avinash Anand, Manisha Luthra, Uelison Jean Lopes dos Santos, Carsten Binnig, Rajiv Ratn Shah

Traffic accidents pose a significant societal challenge, with many fatalities being avoidable through timely emergency response. We introduce IMPACT (Integrated Multimodal Pipeline for Rapid Accident Causality Tracking), a scalable AI framework designed for autonomous, rapid traffic incident analysis using existing urban CCTV infrastructure. IMPACT combines a low-latency CPU-based vision module for real-time key-frame filtering (24 FPS) with the causal reasoning capabilities of MLLMs, reducing costly MLLM calls by over 92% compared to naive sparse sampling. We further present TRACE10K, a dataset featuring three-tier textual annotations that describe accident dynamics at the frame-sequence level.

Xin Chang, Lixin Liu, Jingyu Wang, Jinling Yu, Xiaolin Zhang

Federated Graph Learning enables multiple clients to collaboratively train graph models while protecting local private data. However, most studies have assumed that all clients contribute data voluntarily and actively. Without reasonable incentives, clients are often reluctant to contribute personal data for model training. Furthermore, the budget for incentives is limited, and if clients with low-quality graph data are incentivized to participate in training, it will negatively impact the training performance of all parties in the system. To address this, we propose AEFGL, a Reverse Auction and Value Evaluation-Based Incentive Mechanism for Federated Graph Learning. First, we design a reverse auction mechanism combining graph structural attribute motifs with client production value. Then, we propose a method for evaluating client production value based on the comparison of the client's expected reward and actual value. This mechanism can incentivize clients with high-quality graph data to participate in training within budget constraints, thereby improving the model quality. Experimental results validate the superiority of the AEFGL mechanism and the economic properties it satisfies.

Ahana Chanda, Akshay Agarwal

Accurate identification of mosquito species is crucial for controlling vector-borne diseases, yet visual or acoustic methods alone are often insufficient. We propose a multimodal deep-learning framework that combines high-resolution images with wingbeat audio using a SwinV2 vision transformer and an Audio Spectrogram Transformer, thereby capturing complementary cues. On a six-species dataset, it achieves 97% accuracy, comparable to the best single-modality baseline, and is designed to improve robustness under noise or environmental variation, demonstrating the value of integrating multiple data sources for reliable mosquito surveillance.

Matteo Ceriscioli, Shohei Shimizu, Karthika Mohan

Causal discovery is the task of learning causal models, encoding causal relationships, from a source of information, such as a dataset containing observational data. While many algorithms have been developed to discover causal models under varied sets of assumptions, the case in which the dataset is affected by missing data remains significantly underexplored. Naively applying standard causal discovery algorithms to listwise, test-wise, or regression-wise deleted datasets, or imputing the missing data, can introduce spurious associations between variables and bias function estimation in functional causal models. This issue arises when the data is missing at random or not at random. It ultimately invalidates the theoretical guarantees of these algorithms and prevents finding the true underlying causal model, even in the large-sample limit. An established family of causal models is the Linear Non-Gaussian Acyclic Model (LiNGAM), which assumes linear functional relationships and non-Gaussian independent noise terms. We propose a new causal discovery algorithm for LiNGAM, capable of recovering the underlying causal structure and providing unbiased estimates of the model’s parameters, even when the data is affected by MNAR missingness.

Sizhong Cao

Retrieval-augmented generation (RAG) is the backbone of knowledge-intensive NLP, yet its progress is hindered by a long-standing asymmetry: Generators are refined while retrievers remain static, and full end-to-end optimization is prohibitively unstable. We present BPO-RAG, a bi-level preference-learning framework that redefines the training paradigm by jointly optimizing retrieval and generation with a single supervision signal, pairwise preferences. Stage~1 (Retrieval Preference Optimization) learns to select superior evidence sets, while Stage~2 (Generation Preference Optimization) aligns answer generation with the same evidence, closing the gap between what to read and what to write. This recipe without label requires no reward model or online RL, integrates seamlessly with standard RAG pipelines, and transforms preferences into a unifying training currency. Across open-domain QA benchmarks, BPO-RAG consistently advances retrieval quality and yields more accurate, faithful answers, surpassing strong RAG baselines with remarkable stability. By coupling retrieval and generation under a unified preference framework, BPO-RAG establishes a practical and principled path toward the next generation of reliable, modular, and trustworthy knowledge-intensive language models.

Eduarda Caldeira, Tahar Chettaoui, Naser Damer, Fadi Boutros

Diffusion models conditioned on identity embeddings enable the generation of synthetic face images that consistently preserve identity across multiple samples. Recent work has shown that introducing an additional negative condition through classifier-free guidance during sampling provides a mechanism to suppress undesired attributes, thus improving inter-class separability. Building on this insight, we propose a dynamic weighting scheme for the negative condition that adapts throughout the sampling trajectory. This strategy leverages the complementary strengths of positive and negative conditions at different stages of generation, leading to more diverse yet identity-consistent synthetic data.

Dingzhong Cai, Huale Li, Hang Xiao, Shuhan Qi, Jiajia Zhang

Deep CFR enables end-to-end approximation of Nash equilibria in imperfect-information games(IIGs) but is sensitive to hyperparameters, making manual tuning inefficient. To address this, we propose PT-DCFR, which integrates Population-Based Training(PBT) with Deep CFR to dynamically optimize hyperparameters during training. Building upon this, we further introduce P2T-DCFR, which decouples parameter selection from model performance.

Abhilekh Borah, Niranjan Chebrolu, Kokil Jaidka

Refusals must be resilient, not brittle.” Yet guarding refusals against adversarial phrasing and shifting user contexts remains difficult: large language models (LLMs) still yield to jailbreak prompts that evade safety filters and surface harmful content. We propose Refusal Activation Steering (RAS), a training-free, inference-time method that uses contrastive activations to shift LLM responses, biasing generation trajectories toward refusals without altering model weights. The approach is modular and domain-targetable, avoiding collateral refusals on benign queries while strengthening activation- space boundaries for unsafe content. On adversarial evaluations with an 8B instruction-tuned model, we find that steering improves refusal rate by ∼ 52% and reduces attack success rate by ∼ 40%, establishing a lightweight and interpretable safety layer for robust refusal consistency. To foster further research in this domain, we have made our implementation publicly available.

Philip Bogaars, Argyrios Deligkas, Eduard Eiben, Michail Fasoulakis

In a game of Network Restoration Games With Quotas, there is an underlying graph where a subset of its edges have to be restored by a set of agents. Each agent has a creation cost for each such edge, a traversal cost for every edge of the graph, and in addition they have a quota on the number of edges they have to restore. Then, given a set of edges that fulfill the quota, the cost of an agent is the cost of creating these edges, plus the cost of reaching them, i.e., the traversal cost. We prove that any cost-minimizing allocation is swap-stable, i.e., there is no profitable exchange of edges between any pair of agents, but computing one is hard even on trees. We complement this by designing an algorithm that finds a swap-stable allocation on trees in polynomial time and we quantify its cost against the optimal one.