ArXiV ML/AI/CV papers summary

Theme 1: Physics-Informed and Geometry-Aware Intelligence

We are witnessing a departure from the “black box” era of neural networks toward a future where physical laws and geometric constraints are foundational. By embedding these priors into our models, we move beyond mere pattern matching to true scientific reasoning.

• Physics-Informed Operators: Physics-Informed Conformal Prediction: Embedding PDE Consistency into Distribution-Free Uncertainty Quantification for Neural Operators and PDE-constrained inverse problems at the $\sqrt{n}$ rate via debiased physics-informed neural networks ensure that uncertainty estimates and inverse problem solutions are physically consistent and statistically efficient.
• Geometric Priors: Space as an Interventional Invariant: Cross-Modal Predictive Geometry for Stratified Cities and Em-Spaced Intelligence, $\text{GSF-}\chi$: Global Stereochemical Fields for Chiral Graph Transformers, and Predicting Collision Cross Sections with GRACE: Geometric Residual Adduct Conditioning via Early-fusion demonstrate that encoding the geometry of the world—from molecular chirality to urban structure—dramatically improves model reliability.
• Physical Reasoning: Teaching Vision-Language Models to Use the Scale They Are Given: Label-Free Equivariance Training for Metric Physical Reasoning and Physics-Aware Video Generation via Agentic Planning and Graph-Guided Optimization highlight the necessity of planning that respects kinematic trajectories rather than relying on stochastic interpolation.

Theme 2: The Agentic Turn and Procedural Reasoning

The field is shifting from passive text generation to “agentic” systems capable of executing complex, multi-step workflows. This transition requires a new focus on verification, error recovery, and long-horizon planning.

• Agentic Workflows: ParaRecover: A Process-Level Benchmark for Error Localization and Recovery in Parallel Tool-Use Agents, Look Before You Leap: Pre-Action Verification for LLM Agents, and What Drives Recovery in Agentic Text-to-Cypher? LAST-CQ: An LLM Agent Self-Refinement Framework emphasize that the ability to detect failure and self-correct is the “secret sauce” of agentic performance.
• Long-Horizon Challenges: Tasks over Application Manuals: Revealing Gaps in Long-Horizon Procedural Reasoning for Language Models and Mr. LHDR: A Benchmark for Multimodal Real-World Long-Horizon Deep Research Agents reveal that current models struggle with interdependent rules over extended sequences.
• Autonomous Research: SAEScientist-Bench: Can AI Agents Conduct Autonomous SAE Interpretability Research? and Autonomous Research for Open-Ended Problems: A Case Study on Telecom Ticket Retrieval push toward a future where AI conducts its own scientific inquiry.
• Safety and Assurance: Reality Is the Final Verifier: On Two Key Gaps in Agentic Software Engineering and K-Bench: A Benchmark for LLM Unlearning in Agentic Deployments address the critical need for assurance-revision loops and secure unlearning in deployed agents.

Theme 3: Embodied AI and Robotic Perception

As AI gains a “body,” the challenge shifts to bridging the gap between digital simulation and the messy, unpredictable physical world.

• Sim-to-Real Transfer: EVIS: Real-Time Event Camera Simulation with Multimodal Supervision in NVIDIA Isaac Sim and PhysCodeBench: Benchmarking Physics-Aware Symbolic Simulation of 3D Scenes via Self-Corrective Multi-Agent Refinement provide the high-fidelity environments necessary to train robots without the cost of real-world data.
• Robotic Control: Agent as Policy for Robotic Manipulation, ASTRIL-MPC: Autonomous Traversal Framework of Articulated Tracked Robots with Language-Guided Neural-Kinematic MPC, and PAVXploreRL: Physical-Action-Visual World Model Reinforcement Learning with Action Exploration demonstrate how language-guided reasoning can be married to low-level control for complex tasks like agricultural manipulation.

Theme 4: Reliable Evaluation and the “Validity Gap”

We are currently facing a crisis of trust in automated evaluation. As we rely on “LLM-as-a-Judge,” we must ensure these judges are not merely echoing biases or decorrelating “satisfaction” from “success.”

• Critique of Metrics: Can We Trust LLM Judges: A Study of Capability-Dependent Biases and Multi-Judge Ensemble for Bias Calibration, GAUGE: When Not to Trust LLM-as-a-Judge in User-Simulated Evaluation of Task-Oriented Agents, Measurement Without Validity: The Compounding Reliability Problem in Agentic AI Evaluation, and Auditing Frame-Level AUC in Weakly Supervised Video Anomaly Detection: Granularity, Resolution, and Scene Bias expose the systematic failures in current evaluation pipelines.
• Calibration and Expert-in-the-Loop: Decomposing LLM-Judge Uncertainty to Target Expert Labels and domain-specific efforts like When Rubrics Fail: Hallucinations Reveal Blind Spots in Medical AI Evaluation and Scaling Clinical Judgment to Evaluate Medical AI suggest that we must direct human expert attention to the most ambiguous cases to maintain validity.

Theme 5: Infrastructure, Efficiency, and Sustainability

To move AI from the cloud to the edge, we must rethink our architectures to be more sustainable and computationally efficient.

• Architectural Innovation: Efficient AI Model Deployment Using Quantization Analysis Tool, Attention Quantization for Tabular Foundation Models, Fixed State, Long Reach: What a Constant-Size Cache Buys Block Diffusion at Scale, and RunningTensor: Generalizing Linear Attention to Higher-Order Recurrent States focus on shrinking models and constant-memory inference.
• System-Level Engineering: Rethinking Heterogeneous System Disaggregation for Subquadratic Attention and RoofLang: Enabling AI-Driven Architecting of LLM Inference Systems argue for hardware-aware design, while The Battery Price of edge AI: A study of the Environmental Impact of LLM Inference on Mobile Devices and Computing at Sea: Floating and Offshore Data Centres as a Pathway to Sustainable AI Infrastructure confront the environmental reality of our compute-hungry future.

Theme 6: Domain-Specific Intelligence (Medical and Geospatial)

AI is increasingly applied to high-stakes, multi-scale systems, requiring specialized models that respect the unique constraints of medicine and the environment.

• Medical Imaging: MCSeg: Pre-training and Fine-tuning Volumetric Pyramid Transformer for Multi-modal Cardiac Image Segmentation, Uni-Light: An Ultra-Lightweight Framework via Uncertainty-Aware Knowledge Distillation for Brain Tumour Segmentation, GNM Head: A Generative aNthropometric Model of the human head, and Self-Supervised Cardiac Phase Detection via Single-Parameter Latent Orbits demonstrate how inductive biases can overcome data scarcity in clinical settings.
• Geospatial Intelligence: Genesis: A Generative Engine for Hierarchical Satellite Image Synthesis and High-resolution Calibrated Probabilistic Hourly Precipitation from a Deterministic Forecast show how context-aware generation can maintain consistency across vast spatial and temporal scales.
• Privacy and Human Alignment: BodhiPromptShield: Pre-Inference Prompt Mediation for Surface-Form Privacy Propagation in LLM Agent Pipelines and Video2Reaction: Training Foundation Video Models to Predict Audience Reaction ensure that as AI becomes more integrated into our lives, it remains private, transparent, and aligned with human perception.