Coordinating multiple AI models and tools within a single reasoning pipeline is challenging when systems rely on crude task-matching rather than accounting for cost and performance differences. This paper proposes an auction-based framework where reasoning steps are treated as tradeable tasks, and expert models "bid" based on calibrated competence, routing work to the most capable rather than most confident solver. Evaluated across five benchmarks, it outperforms standard routing and cascade baselines while allowing a tunable cost-quality trade-off. Applications include more efficient orchestration of AI agent ecosystems, especially in enterprise settings needing controllable cost versus quality trade-offs.

Authors: Kaiji Zhou, Ales Leonardis, Yue Feng

Paper: https://arxiv.org/abs/2607.09600v1

Quizbowl-style question answering, where answers must be given as clues are incrementally revealed, tests both confidence calibration and reasoning under uncertainty. This paper describes a submission using two specialized agents: one deciding when to answer tossup questions using confidence calibration and numeric reasoning safeguards, and another handling bonus questions with structured, multimodal reasoning. Achieving the top leaderboard score without retrieval pipelines or ensembles, the system shows lightweight task-specific strategies can be highly effective. Applications include efficient, resource-constrained multimodal QA systems for trivia, education, and other domains requiring calibrated confidence under partial information.

Authors: Nirjhar Das, Md. Al-Mamun Provath

Paper: https://arxiv.org/abs/2607.09623v1

Autonomous driving requires perceiving both objects and the surrounding environment, but existing radar-camera fusion methods focus mainly on detection rather than joint scene understanding. This paper proposes a framework treating semantic occupancy as an evolving internal state rather than a final output, using specialized modules to strengthen spatial and temporal feature fusion from 4D radar and camera data. The authors also extend existing driving datasets with new occupancy labels. Applications include more robust perception systems for self-driving cars, particularly in adverse conditions where radar's reliability complements camera limitations, improving both object detection and scene-layout understanding simultaneously.

Authors: Xiaokai Bai, Lianqing Zheng, Runwei Guan, Songkai Wang, Siyuan Cao, Hui-liang Shen

Paper: https://arxiv.org/abs/2607.09629v1

Quantum information theory underpins quantum computing and communication, but formalizing its theorems in machine-checkable form has lacked reusable infrastructure. This paper presents a Lean 4 library providing composable, verified building blocks for quantum states, channels, codes, and rate constructions, used to formally prove major theorems like Schumacher's source-coding theorem and the Holevo-Schumacher-Westmoreland capacity theorem. Applications include rigorous, bug-free verification of quantum communication protocols, a foundation for future automated theorem proving in quantum information science, and a knowledge base enabling AI-assisted formalization and agentic reasoning in quantum computing research.

Authors: Chengkai Zhu, Ziao Tang, Guocheng Zhen, Yimeng Cao, Yusheng Zhao, Ranyiliu Chen, Xuanqiang Zhao, Lei Zhang, Xin Wang

Paper: https://arxiv.org/abs/2607.09632v1

Financial fraud detection suffers from extreme class imbalance, often causing models to simply predict "no fraud" and miss rare cases. This paper proposes a multi-objective reinforcement learning approach that converts transaction data into natural-language narratives encoded by LLMs, then optimizes a reward balancing fraud detection, customer friction, and semantic insight - without relying on distortive resampling techniques. Tested on e-commerce and credit datasets, it avoids the "zero-recall trap" common to imbalanced classifiers. Applications include real-time fraud detection systems that better balance catching fraud against inconveniencing legitimate customers, offering banks and payment platforms a tunable trade-off frontier.

Authors: Claudio Lucio do Val Lopes, Lucca Machado da Silva

Paper: https://arxiv.org/abs/2607.09641v1

Concept-based explainable AI aims to make model reasoning human-readable, but the concepts themselves can be unreliable. This paper introduces an auditing framework that perturbs input regions, measures how concept outputs shift, and fits a surrogate model to assess whether concept explanations are trustworthy - evaluating attribution accuracy, faithfulness, and stability. Tested on retinal fundus images comparing segmentation-based versus vision-language-derived concepts, it reveals that reliability varies by concept type and pathway. Applications include validating explainable AI systems before clinical deployment, particularly in medical imaging, where trustworthy explanations are critical for physician confidence and regulatory approval.

Authors: Mohadeseh Mollapour, Koorosh Aslansefat, Zeinab Dehghani, Bhupesh Kumar Mishra, Tejal Shah, Zhibao Mian

Paper: https://arxiv.org/abs/2607.09649v1

IoT devices are notoriously vulnerable due to weak defaults and outdated firmware, yet automated security testing tailored to IoT is underdeveloped. This paper presents a multi-agent LLM framework that autonomously discovers and exploits IoT vulnerabilities, pairing a detection agent with an attack-execution agent to plan and carry out exploits. Tested across ten attack scenarios in IoTGoat and Metasploitable environments, it achieved up to 100% success with low computational overhead and fast execution. Applications include automated penetration testing, continuous security auditing of IoT deployments, and reducing manual effort in vulnerability assessment - though such tools also raise dual-use security considerations.

Authors: Katherine Swinea, Kshitiz Aryal, Lopamudra Praharaj, Maanak Gupta

Paper: https://arxiv.org/abs/2607.09653v1

Vision-language model benchmarks have mostly used simple scenes and small human-description samples, leaving model errors on complex social scenes poorly understood. This study introduces a new dataset of images depicting complex social interactions and compares a decade of vision-language models (2017-2025) against human describers, analyzing five error types including hallucination and spatial reasoning. Results show multimodal LLMs now match top human performance and have closed the gap between simple and complex scenes. Applications include informing benchmark design, guiding development priorities for social scene understanding, and identifying remaining weaknesses (spatial dependence) for assistive and robotic vision systems.

Authors: Shravan Murlidaran, Miguel P. Eckstein

Paper: https://arxiv.org/abs/2607.09654v1

Language models are typically trained on plain text extracted from documents, discarding layout, figures, and typography that carry meaningful information. This paper investigates training foundation models directly on visual renderings of documents instead of extracted text, testing whether "seeing" a page beats "reading" it. Across multiple model backbones and benchmarks, visual pretraining outperforms text-only pretraining on the same source material. Applications include more efficient and capable foundation models for tasks involving richly formatted documents, web pages, scientific papers, and forms - domains where converting to plain text currently discards valuable structural and visual information.

Authors: Yiming Zhang, Zhonghan Zhao, Wenwei Zhang, Haiteng Zhao, Tianyang Lin, Yunhua Zhou, Demin Song, Kuikun Liu, Haochen Ye, Haian Huang, Yuzhe Gu, Haijun Lv, Qipeng Guo, Bin Liu, Gaoang Wang, Kai Chen

Paper: https://arxiv.org/abs/2607.09657v1

Dream detection from EEG has long relied on spectral power features, capping performance around 0.70 AUC. This paper reframes the problem geometrically: by embedding EEG signals in phase space and tracking topological features (Betti curves) across sliding windows, it captures the shape of brain dynamics rather than just energy content. Combined with a topology-conditioned generative model for synthesizing dream-state EEG, the approach targets substantially higher classification accuracy (0.82-0.90 AUC). Potential applications include wearable brain-computer interfaces for sleep and dream monitoring, clinical sleep diagnostics, and generating synthetic EEG data to augment scarce dream-state datasets for future research.

Authors: Ren Takahashi, Emre Yusuf, Jayabrata Bhaduri

Paper: https://arxiv.org/abs/2607.09662v1