Enterprise AI deployments often route each LLM call independently to balance cost and quality, but agentic workflows only get evaluated by delayed, task-level outcomes, misaligning per-call routing with actual performance signals. TRACE-Router fixes this by assigning an entire task to one model at the start via a contextual bandit, then updating its policy using the task's terminal reward balancing accuracy and latency. Across agentic benchmarks like tau2-Bench and Terminal-Bench, it achieved notable accuracy and latency improvements over single-model baselines. This has clear applications in enterprise LLM infrastructure, optimizing cost-quality tradeoffs for long-horizon, multi-step agentic applications.

Authors: Ritik Raj, Souvik Kundu, Sarbartha Banerjee, Dheemanth Joshi, Ishita Vohra, Tushar Krishna

Paper: https://arxiv.org/abs/2607.22465v1

Quantum chemistry simulations require efficient state-preparation circuits, but iterative methods like ADAPT-VQE become computationally prohibitive for large, pharmaceutically relevant molecules. ADAPT-GQE addresses this with a generative AI framework trained on ADAPT-VQE-generated reference circuits, then improved further via reinforcement learning to exceed the training data's accuracy. Demonstrated on imipramine, a tricyclic antidepressant, and executed on Quantinuum's Helios-1 quantum hardware, the approach achieves order-of-magnitude speedups in circuit generation. Applications span drug discovery, materials science, and utility-scale quantum computational chemistry, marking a step toward automated, AI-driven quantum circuit synthesis for real-world molecular modeling.

Authors: Alex Koziell-Pipe, Jasmine Brewer, Jem Guhit, Marwa H. Farag, Kripa Panchagnula, Gabriel Laude, Fabian Finger, Carlo Gaggioli, Ludmila Szulakowska, Oliver J. Backhouse, Christos Papalitsas, Jason G. Mustakis, Thomas Soini, David Munoz Ramo, Stephen Clark, Elica Kyoseva, Enrico Rinaldi

Paper: https://arxiv.org/abs/2607.22468v1

LLM-based test-driven code generation struggles when only natural-language requirements exist, since automatically generated tests can be faulty or inconsistent, misleading optimization. MineValiCoder tackles this with a closed-loop framework: filtering unreliable tests via self-validation, iteratively refining diverse code candidates, and using a bipartite graph model to jointly validate code-test interactions for stable final selection. Tested across four LLMs and major benchmarks, it achieved strong Pass@1 scores including 96.34% on HumanEval. Applications include automated software engineering pipelines, reducing reliance on human-crafted tests, and improving reliability of AI code generation tools in production development workflows.

Authors: Zhen Zhao, Qihang Yang, Feifei Dai, Xiangfang Li, Bo Li

Paper: https://arxiv.org/abs/2607.22471v1

Low-Rank Adaptation (LoRA) fine-tunes large models efficiently but conventionally updates all matrices uniformly, wasting compute on matrices that contribute little. κ-LoRA shows that matrices with higher condition numbers hold underdeveloped directions driving most adaptation gains, while low-condition-number matrices are already balanced and add little value. By restricting updates to the top 50% of matrices by condition number, the method halves trainable parameters, cuts fine-tuning time by 16.2%, and reduces memory use, while matching standard LoRA accuracy. This is directly applicable to efficient large-model fine-tuning, particularly for edge deployment and resource-constrained on-device adaptation scenarios.

Authors: Jianghui Wang, Silong Yong, Francesco Orabona, Marco Canini, Katia P. Sycara, Yaqi Xie

Paper: https://arxiv.org/abs/2607.22489v1

LLM-driven automated research is limited by unreliable evaluation, since LLM reviewers can accept fabricated results near chance levels. CausalForge addresses this by grounding causal inference research in the Lean proof assistant, pairing a machine-checked causal inference library (Causalean) with an agentic pipeline (CausalSmith) that proposes, formalizes, and proves results, then audits whether formal statements faithfully represent the intended informal claims. This offers a template for trustworthy automated scientific discovery, with applications in mathematics and causal inference research acceleration, verifiable AI-assisted theorem generation, and building reviewer-independent quality assurance into autonomous research systems.

Authors: Jiyuan Tan, Vasilis Syrgkanis

Paper: https://arxiv.org/abs/2607.22511v1

This study probes how commercial LLMs (Claude, Grok, GPT, Gemini) evaluate ethnonationalist pseudo-scientific claims across time and interfaces, finding Grok's default versions rated such claims far more credible than competitors, with behavior shifting via undocumented silent patches and diverging between API and web access for the same model. The findings reveal that an LLM's epistemic stance isn't fixed but shaped by system prompts, safety layers, and routing decisions invisible to users. Applications include AI governance, auditing frameworks for deployed models, journalism and policy research on AI accountability, and informing regulatory approaches to LLM content moderation transparency.

Authors: Davide Scarso, Hugo Noronha de Almeida, Joaquim Pina

Paper: https://arxiv.org/abs/2607.22513v1

Quantum machine learning models typically encode matrix-valued data using generic rotation gates that ignore matrix-level spectral structure. Quantum Spectral Models (QSMs) instead build the data-encoding unitary's generator directly from each input's spectral properties, testing symmetric, global-block, and patch-local Hamiltonian variants. Evaluated on Pendigits and synthetic spectral-statistics tasks, QSM variants achieved leading accuracy, with different variants excelling on different task types. This research has applications in advancing quantum machine learning architecture design, informing how structured, input-aware encodings can improve inductive bias, and offering broader principles for structure-aware model design applicable beyond quantum computing.

Authors: Peiyong Wang, Udaya Parampalli, Casey R. Myers

Paper: https://arxiv.org/abs/2607.22516v1

Adding "skills" (procedural guidance) to LLM agents is usually judged by average performance gains, but this masks cases where skills actively cause failures on previously solvable tasks. Analyzing nearly 6,000 runs across office automation benchmarks, the authors identify three regression mechanisms: skill presence alone altering behavior, skills overriding correct input interpretation, and skills suppressing self-verification. They find top-performing skills win mainly by regressing less, not gaining more. This has direct applications for designing and evaluating agent tooling, suggesting skill libraries should emphasize grounding and verification support rather than pure procedural instructions to improve real-world agent reliability.

Authors: Darshan Tank, Baran Nama

Paper: https://arxiv.org/abs/2607.22520v1

As AI moves into safety-critical domains like aviation, healthcare, and autonomous driving, trust hinges on explainability. This work applies explainability techniques to a reinforcement learning agent trained in a simplified air traffic control environment, where the agent chooses alternative flight routes to avoid no-fly zones. Using saliency maps, the authors expose which input features most influence the agent's routing decisions. The approach offers a preliminary but concrete path toward human-interpretable AI decision support for controllers, with potential application in building operator trust, certifying AI-assisted aviation tools, and extending explainability methods to other high-stakes automated decision systems.

Authors: Anduel Mehmeti, Gabriella Gigante, Salvatore Venticinque

Paper: https://arxiv.org/abs/2607.22525v1

Most monocular motion-perception systems treat every pixel's movement as an independent displacement, ignoring that real objects move as rigid bodies. SM4RT challenges this by encoding scene dynamics as a small set of shared "motion bases," each a temporal sequence of 6D twists in SE(3), so points belonging to the same object inherit a common rigid trajectory. A parallel encoder-decoder recovers 3D geometry and world-coordinate motion from a single RGB video in one pass. Potential applications include robotics perception, autonomous driving scene understanding, AR/VR content creation, and any pipeline needing physically consistent 4D reconstructions from ordinary video.

Authors: Shing Ho J. Lin, Wenzhao Zheng, Dong Zhuo, Yuqi Wu, Jie Zhou, Jiwen Lu

Paper: https://arxiv.org/abs/2607.22534v1