Financial fraud detection in transaction networks faces a fundamental challenge: fraudulent activity is rare, well-disguised, and often underrepresented in labeled data. Standard graph neural networks tend to smooth out the very irregularities that signal fraud. ADC-GNN tackles this with three complementary mechanisms: diffusion-guided feature augmentation that stabilizes node representations against noise, contrastive learning across perturbed views, and a spectral attention module that adaptively amplifies fraud-relevant frequency signals across multiple graph hops. Evaluated on public benchmarks and a real telecom transaction dataset, it consistently outperforms baselines under low-label conditions. Applications include credit card fraud detection, anti-money laundering systems, telecommunications billing abuse detection, and social network spam identification.

Authors: Liming Liu, Chao Hu, Mingfei Lu, Yiwei Ge, Xingle Li, Heyuan Shi

Paper: https://arxiv.org/abs/2606.28134v1

In-context segmentation asks a model to identify target regions in new images using only a handful of labeled reference examples — no retraining required. Current approaches work by matching low-level visual features between references and queries, making them brittle when references vary in viewpoint, lighting, or appearance. CG-ICS instead extracts high-level semantic concepts from references using a multimodal language model, then uses these concepts alongside a spatial grounding route to guide a frozen SAM3 segmentation backbone. It achieves state-of-the-art accuracy and substantially reduced variance across diverse reference choices. Applications span medical image annotation, few-shot industrial inspection, and rapid domain adaptation in computer vision pipelines with limited labeled data.

Authors: Zhigang Chen, Xiawu Zheng, Rongrong Ji

Paper: https://arxiv.org/abs/2606.28149v1

Jailbreak attacks — prompts engineered to make safety-aligned LLMs produce harmful outputs — are a persistent concern, but exactly how they work mechanistically has remained murky. This paper provides evidence that successful attacks don't erase safety representations; they selectively suppress specific "Adversarially Compromised Heads" in early attention layers while leaving "Safety-Aligned Heads" in mid-layers largely intact. This residual safety signal is detectable without any additional training, and reading it yields competitive jailbreak detection performance with strong robustness. These findings have direct implications for LLM safety auditing, interpretability-based defenses, red-teaming methodologies, and the design of future architectures with more resilient safety mechanisms.

Authors: Yanchen Yin, Dongqi Han, Linghui Li

Paper: https://arxiv.org/abs/2606.28153v1

Reinforcement learning has dramatically improved LLM reasoning on tasks like competition math — but the resulting models often reason in ways that are difficult for weaker models or humans to follow, limiting their real-world utility. Tandem Reinforcement Learning (TRL) addresses this by co-training a strong "senior" model alongside a frozen "junior" model: both contribute to generating reasoning chains, and the senior is rewarded as a team with the junior. This nudges the senior to reason in ways the junior can understand and continue. Beyond math tutoring, TRL has implications for human-AI collaboration, multi-model pipelines, and building AI systems whose reasoning remains interpretable and handoff-compatible across capability levels.

Authors: Difan Jiao, Raghav Singhal, Robert West, Ashton Anderson

Paper: https://arxiv.org/abs/2606.28166v1

Large-scale studies linking heart imaging measurements to disease risk typically rely on pre-defined, single-variable features chosen by experts — an approach that may miss important non-linear relationships or interactions between measurements. CPAgents automates the discovery of richer, composite phenotypes (ratios, polynomial combinations, interaction terms) through a three-agent loop: an Analyst identifies statistical issues, a Proposer generates candidate expressions, and a Verifier validates them against multi-stage criteria. Applied to a large cardiac imaging cohort, the discovered phenotypes outperform baselines across 56 of 72 evaluation combinations spanning nine disease categories. Applications include population-scale cardiovascular risk stratification, imaging biomarker discovery, and automated feature engineering for clinical machine learning.

Authors: Zuoou Li, Wenlong Zhao, Kelly Yu, Weitong Zhang, Paul M. Matthews, Wenjia Bai, Bernhard Kainz, Mengyun Qiao

Paper: https://arxiv.org/abs/2606.28179v1

Getting multiple AI agents to work together effectively in a shared physical environment is harder than it sounds — agents frequently act on outdated assumptions about their partners or issue redundant, mistimed communications. LLawCo addresses this by having agents reflect on past failures to extract high-level "laws of cooperation," such as knowing when to speak and when to wait, then fine-tuning on these laws so cooperative reasoning becomes intrinsic. Evaluated on the PARTNR-Dialog and TDW-MAT benchmarks, it achieves meaningful gains over state-of-the-art open-source baselines. Applications include household robots, warehouse automation, collaborative AI assistants, and any multi-agent setting requiring fluid, context-sensitive coordination.

Authors: Qinhong Zhou, Chuang Gan, Anoop Cherian

Paper: https://arxiv.org/abs/2606.28182v1

Predicting how hard an exam question will be for human test-takers — without running expensive human trials — would transform educational assessment. This paper proposes using the reasoning traces of large language models as a proxy for human cognitive effort. Rather than treating these traces as raw text, Epi2Diff structures them into meaningful "cognitive episodes" — functional states like planning, implementing, and verifying — and uses the dynamics between these states to predict difficulty. Tested on four real-world human difficulty datasets including SAT-derived benchmarks, it consistently outperforms strong baselines. Applications include automated test construction, adaptive learning platforms, and AI-assisted item difficulty calibration for standardized assessments.

Authors: Chenguang Wang, Ming Li, Xinyue Zeng, Zhuochun Li, Hong Jiao, Tianyi Zhou, Dawei Zhou

Paper: https://arxiv.org/abs/2606.28186v1

Remittances — money sent home by migrant workers — are a lifeline for many developing economies, yet surprisingly hard to forecast reliably. This study applies rigorous time-series and machine learning methods to 32 years of Sri Lankan migration and remittance data, finding that external factors like exchange rates and global oil prices drive inflows far more than domestic indicators. A multivariate Ridge Regression model outperforms traditional approaches by 73.8% in accuracy, projecting 2026 remittances at approximately USD 9 billion. These findings can inform central bank policy, foreign exchange management, diaspora engagement strategies, and international development planning in remittance-dependent economies across South and Southeast Asia.

Authors: Dhinanjaya Fernando, Dinura Ginige, Kalana Lakshan, Chanupa Gurusinghe, Lasana Pahanga, Subavarshana Arumugam, Sandeepa Weerasekara, Sandareka Wickramanayake, Nisansa de Silva

Paper: https://arxiv.org/abs/2606.28190v1

Building robots that can understand and interact with the physical world requires massive amounts of 3D training data — but capturing that data with multi-camera rigs is expensive and impractical at scale. HAT-4D proposes using ordinary monocular video as a data source, reconstructing the 3D geometry and temporal dynamics of multiple interacting objects with the help of vision-language models and targeted human feedback. It introduces a new benchmark, MVOIK-4D, for evaluating such reconstructions on physical plausibility and temporal consistency. Applications include scalable embodied AI training data pipelines, robotic manipulation learning, AR/VR scene reconstruction, and virtual environment generation from real-world video footage.

Authors: Jiaxin Li, Yuxiang Wu, Zhenkai Zhang, Xinrui Shi, Haoyuan Wang, Yichen Zhao, Su Linxiang, Chenyang Yu, Mingyu Zhang, Yifan Ding, Boran Wen, Li Zhang, Ruiyang Liu, Yong-Lu Li

Paper: https://arxiv.org/abs/2606.28215v1

As AI systems increasingly act as proxies for human stakeholders in shared learning environments, a thorny question arises: how do you fairly reward each participant's contribution when different contributors have different values — and when some contributions might violate those values? This paper proposes a framework that filters gradient updates by each principal's value profile before computing credit, grounded in a "traversal learning" substrate that preserves richer attribution paths than standard federated averaging. Relevant to decentralized AI training cooperatives, privacy-preserving machine learning, pluralistic AI alignment efforts, and emerging contexts like data DAOs where multiple parties co-own and co-train models under heterogeneous ethical constraints.

Authors: Young Yoon, Jimin Kim, Soyeon Park

Paper: https://arxiv.org/abs/2606.28217v1