ArXiV ML/AI/CV papers summary

Theme 1: Foundations of Explainability, Trust, and Verification

The machine learning community is shifting away from treating “explainability” as a post-hoc visualization task, moving instead toward structural, foundational requirements for system reliability. This transition is critical as AI enters high-stakes domains like medicine and law, where “black box” models pose significant liabilities.

• Position: Explainability Research Must Prioritize Foundations over Ad-hoc Methods argues for a pivot toward structural challenges, such as human-in-the-loop integration.
• Local Additive Feature Attribution: A Mathematical Taxonomy and Reporting Checklist provides the necessary rigor to organize disparate methods into a common framework.
• Towards a Unified Multidimensional Explainability Metric: Evaluating Trustworthiness in AI Models proposes a standardized score based on fidelity, simplicity, and stability.
• Faithful Autoformalization of Natural Language Assertions and Verifying formulas for interventional distributions bridge the gap between informal intent and formal, verifiable specifications.
• Anatomically Faithful but Temporally Blind: Auditing Attribution for Left-Ventricular Ejection-Fraction Estimation from Echocardiography warns that spatial faithfulness in XAI does not guarantee temporal accuracy.
• The Refusal Residue: When Probes Catch Alignment Faking and When They Don’t explores the phenomenon of “alignment faking,” where models maintain hidden, non-compliant behaviors despite appearing safe.

Theme 2: Physics-Informed Modeling and World Models

To move beyond passive text assimilation, researchers are grounding AI in the laws of physics and mechanistic structures. This approach enables models to understand 3D space and dynamical systems, often with greater data efficiency than massive, unconstrained models.

• LIGO-PINN: Learned Initialization via Gated Optimization to Alleviate Convergence Failures in Physics Informed Neural Networks addresses training instability in PINNs.
• Operator-Informed Gaussian Processes for Complex Helmholtz Wavefields: From Synthetic Benchmarks to In Vivo Brain Elastography provides uncertainty quantification for complex-valued wavefields.
• A Minimal Interpretable Architecture for Zero-Shot Reconstruction of Dynamical Systems demonstrates that minimal architectures (DynaBase) can outperform massive foundation models in specific dynamical tasks.
• Grounded world models in biological organisms and future embodied AI and From Observation to Insight: Mechanistic World Models and the Quest for Autonomous Discovery advocate for reusable, explanatory structures.
• SeeSE3: Emergence of 3D Space in Vision Features shows that frozen vision models inherently contain latent representations of 3D Euclidean space.
• OPINE-World: Programmatic World Modeling with Ontology-error-Prioritized Interactive Exploration for ARC-AGI-3 proves that programmatic models are more data-efficient than deep-network-only approaches for object-centric world modeling.

Theme 3: Agentic Workflows and Embodied Intelligence

The field is transitioning from “AI as a chatbot” to “AI as an agent” capable of long-horizon planning, tool use, and physical interaction. This requires “harness engineering”—wrapping deterministic scaffolding around LLM cores to ensure reliability.

• Branching Policy Optimization: Sandbox-Native Language Agent Reinforcement Learning exploits sandbox environments to reduce training variance.
• NexForge: Scaling Executable Agent Tasks via Requirement-First Synthesis shifts from substrate-first to requirement-first data synthesis.
• AgentWorm: Self-Propagating Attacks Across LLM Agent Ecosystems highlights the security risks of autonomous, interconnected agents.
• Harnessing LLMs for Reliable Academic Supervision: A Comparative Study, MemoHarness: Agent Harnesses That Learn from Experience, and AgentCheck: A Reproduce-Intervene-Mitigate Workbench for LLM Agents over MCP provide frameworks for fault injection, memory management, and learning from failure.
• Action QFormer: Structured Representation Shaping under Action Supervision in Vision-Language-Action Models and FoMoVLA: Bridging Visual Foresight and Motion Guidance for Vision-Language-Action Models focus on injecting action supervision into VLA models without degrading semantic grounding.
• Native Video-Action Pretraining for Generalizable Robot Control and Open-AoE: An Open Egocentric Manipulation Dataset and Toolchain for Embodied Learning provide the infrastructure for training robust embodied agents.

Theme 4: Healthcare and Scientific Discovery

AI in high-stakes scientific domains is moving toward robust, domain-specific deployment, where the “foundation” is defined by the structure of the data rather than just the scale of the model.

• TEDDY: A Pediatric Foundation Model for Risk Forewarning from ICD-Coded Diagnostic Histories shows that compact, specialized models outperform general-purpose ones in clinical settings.
• A Temporal Machine Learning-Based Time-to-Event Model for Predicting ALS Progression and Healthcare Utilization utilizes digital-twin-inspired frameworks for personalized medicine.
• LATTICE: Graph Self-Supervised Learning for Multimodal Spatial Omics Integration integrates disparate biological modalities.
• Causal-Adversarial Probing of Clinical Covariates for Prostate MRI Grading and CRISP: Constrained Refinement via Iterative Squeezing Process for Robust Medical Image Segmentation under Domain Shift use causal reasoning and structural priors to mitigate “shortcut” correlations and domain shift.

Theme 5: Efficiency, Architecture, and 3D Scene Representation

The “Green AI” movement is driving innovations in hardware-software co-design, efficient attention mechanisms, and high-fidelity 3D reconstruction.

• PolyQ: Codesigning End-to-End Quantization Framework for Scalable Edge CPU LLM Inference, ExaGEMM: Exploration Framework for CPU-Driven ML Inference, and ExTernD: Expanded-Rank Ternary Decomposition Ternary LLM PTQ with Accuracy Approaching Any Quantization Level focus on reducing the memory footprint and compute requirements of large models.
• FlashDiff: Efficient Regional Execution and Scheduling for Diffusion Model Serving, VideoSEMA: a scalable and efficient Mamba-like attention for video understanding, and FlashDecoder: Real-Time Latent-to-Pixel Streaming Decoder with Transformers optimize inference for generative models.
• STKAN: Kolmogorov-Arnold Networks for Spatio-Temporal Forecasting and A Hybrid Mamba for Audio-Visual Navigation explore alternatives to standard Transformers.
• Instant NuRec: Feed-Forward 3D Gaussian Reconstruction for Driving Scene Simulation, RoGS: Adaptive Meshgrid Gaussian for Large-Scale Road Surface Mapping, and MVFusion-GS: Motion-Variance Guided Temporal Attention for High-Quality Dynamic Gaussian Splatting advance 3D Gaussian Splatting for real-time digital twins.