AI Agents for Finance and Security, Medical Vision-Language Models, and Kernel Regression Asymptotics

Today’s selection of 8 noteworthy AI/ML papers from arXiv spans the evolving capabilities of LLM agents—from financial trading and UAV navigation to security auditing—alongside breakthroughs in medical imaging, geometry formalization, and the mathematical foundations of kernel learning.

1. RetailAgent: Structured Adverse Timing in Self-Conditioned Multimodal LLM Trading Agents

Authors: Yupeng Zhang, Liuyuan Jiang, Hongyi Huang, Bingheng Li, Lisha Chen | Categories: cs.AI, q-fin.TR Link: arxiv.org/abs/2608.28399v1

This paper introduces RetailAgent, a framework to test whether LLM trading agents exhibit predictable directional structure in their decisions. Using exposure-matched analysis on intraday equity prices, the authors find persistent negative timing across various model families and conditions, revealing that LLM policies are sequentially predictable. The study suggests these behavioral patterns could be exploited by other market participants.

Takeaway: Demonstrates a measurable and stable behavioral bias in LLM-driven financial decisions, a critical consideration for anyone deploying autonomous agents in high-stakes, adversarial environments.

2. D-TAIA: Domain-Aware LLM Adaptation for Multi-Task Predictive Process Monitoring

Authors: Sjoerd van Straten, Christine Jacob, Marwan Hassani | Categories: cs.LG Link: arxiv.org/abs/2608.28236v1

D-TAIA presents a framework for joint next-activity and remaining-time prediction in business processes by fine-tuning a foundation model backbone. It combines domain-aware triplet loss pre-training with FAISS-based nearest-neighbor retrieval and a specialized inference strategy to handle distributional shift and data scarcity. Evaluated on four real-world logs, the approach achieves state-of-the-art performance with a compact 10M-parameter model.

Takeaway: Proves that parameter-efficient fine-tuning of small foundation models can outperform larger, task-specific architectures in process mining, and effectively transfers techniques from NLP and CV to a new domain.

3. NL2AGBench: Benchmarking LLM Auto-Formalization for AlphaGeometry

Authors: Samuel Xiao, Judy Song, Rory Hu, Ziliang Zong | Categories: cs.CL, cs.AI Link: arxiv.org/abs/2608.28481v1

This work introduces NL2AGBench, a benchmark for evaluating LLMs on translating natural-language geometry problems into the formal DSL required by the AlphaGeometry system. Using execution-based verification, the authors show a significant performance gap: leading closed-source models achieve over 80% executable translation, while open-source models struggle with constraint preservation. They propose an error taxonomy and test mitigation strategies like few-shot and fine-tuning.

Takeaway: Highlights that auto-formalization is a critical usability bottleneck for neuro-symbolic systems, and provides a standardized benchmark and analysis to drive progress in this underexplored area.

4. ARC-CT: Anatomy-Routed Contrastive Vision-Language Learning for 3D Chest CT

Authors: Huseyin Umut Isik, Mehmet Alp Ozaydin, Sila Kurugol, Şeyda Ertekin | Categories: cs.CV, cs.AI Link: arxiv.org/abs/2608.28455v1

ARC-CT is a region-aware contrastive learning framework for 3D chest CT scans that keeps anatomical locality while addressing the false-negative problem in standard contrastive objectives. It integrates an AnatomyQFormer, a soft InfoNCE loss based on label-set overlap, and organ-level alignment to learn abnormality classifiers without manual annotations. The method achieves a 0.86 mask-free macro AUC on 18 abnormalities with a compact ResNet-18 backbone.

Takeaway: An encouraging approach that shows how combining LLM-extracted labels with anatomical priors can push efficient medical-imaging models past much larger transformers.

5. LLM-Based Agents for Software and Systems Security: Approaches, Applications, and Assessment

Authors: Jingjing Nie, Jiawei Guo, Krishna Meda, Haipeng Cai | Categories: cs.CR, cs.AI Link: arxiv.org/abs/2608.28490v1

This systematic literature review synthesizes 2023-2026 peer-reviewed research on LLM agents in security, covering technical approaches, applications, and assessment methodologies. The authors find a field that has successfully built agents able to act autonomously, yet lacking in bounded authority and auditable behavior. They outline major limitations and challenges in current designs, pointing toward future research directions.

Takeaway: A critical, sobering analysis for the security community—it clarifies that while autonomous agents are feasible, the field is far from ensuring they are safe and accountable.

6. Spatial-Semantic Reasoning using Large Language Models for Efficient UAV Search Operations

Authors: Marin Maletic, Marijana Peti, Tamara Petrovic, Stjepan Bogdan | Categories: cs.RO, cs.AI Link: arxiv.org/abs/2608.28270v1

The authors present a real-time framework for UAV Object Goal Navigation (ObjectNav) that uses an LLM to interpret natural language instructions and reason over spatial and semantic context. The system dynamically updates search priorities based on new observations, enabling smooth, feasible trajectories via polynomial spline interpolation. Simulations and real-world tests show significant reductions in mission duration with maintained accuracy.

Takeaway: Pushes LLM-based navigation beyond simulator constraints into real-time deployment, proving that semantic reasoning can effectively guide physical systems for time-critical tasks.

7. Learning between the peaks: sharp asymptotics for kernel ridge regression under power-law anisotropy

Authors: Lorenzo Rizzi, Arie Wortsman Zurich, Bruno Loureiro | Categories: stat.ML, cs.LG Link: arxiv.org/abs/2608.28564v1

This paper derives exact asymptotic expressions for kernel ridge regression under anisotropic data with power-law decay and polynomial kernels. The analysis reveals how anisotropy reshapes learning curves, damping variance peaks at integer sample complexities and decoupling bias transitions under strong anisotropy. The results unify and extend classical learning theories by connecting source and capacity rates to the data geometry.

Takeaway: Offers a deep mathematical understanding of how input data geometry fundamentally impacts kernel learning, resolving subtle behaviors in interpolation peaks and bias dynamics.

8. On the Maintenance and Co-evolution of Agent Plugins: An Empirical Study of Claude Code Plugin Marketplaces

Authors: Ahmed Hereiz, Yingzhe Lyu, Hao Li, Bram Adams, Ahmed E. Hassan | Categories: cs.SE, cs.AI Link: arxiv.org/abs/2608.28497v1

This empirical study analyzes 8,351 plugins from 1,926 Claude Code plugin marketplace repositories, revealing rapid expansion since its October 2025 launch. The study finds feature-driven development dominates, with Claude co-authoring nearly 35% of commits, and that instruction files and scripts co-evolve in a functionally coupled way, a new maintenance dependency. This contrasts sharply with conventional OSS development patterns.

Takeaway: The first deep look at the emerging ecosystem of AI-agent plugins, identifying novel co-evolution dynamics that will define the next generation of software maintenance.

This content was generated with AI assistance. Paper information sourced from arXiv.