ArXiV ML/AI/CV papers summary

This collection of research represents a vibrant, multi-disciplinary frontier in machine learning. As we push the boundaries of what AI can achieve, we are moving beyond simple pattern matching toward systems that understand the physical world, respect human privacy, and operate with verifiable reliability. We are transitioning from the “brute force” era of scaling parameters toward a paradigm of structural intelligence, where physics, geometry, and rigorous statistical frameworks build AI that is not only powerful but fundamentally aligned with the world it inhabits.

Theme 1: Physics-Informed and Embodied Intelligence

The field is shifting from “black-box” statistical correlations to models that respect the laws of nature and the constraints of 3D space. By embedding physical constraints into neural architectures, we achieve unprecedented stability and data efficiency.

• Neural Operators & Dynamics: LiNO: Lifting based multiresolution neural operator and CSympNet-ID: conformal-symplectic map learning for linearly damped Hamiltonian systems learn the solution operators of differential equations, allowing for generalization across physical regimes. A Physics-Regulated Neural Framework for Learning 3D Grain Growth Dynamics and PDEFlow: Autonomous Agentic PDE Pipelines for Neural Operator Learning and Solver-Free Inference automate scientific discovery, while LRX-PINN: A Layer-Resolving XNet Physics-Informed Neural Network with Integrated Cauchy Activations for Convection-Dominated Problems and Diffusion learning reveals viable parameter manifolds and compensation geometry in biological dynamical systems use physical equations to solve complex problems with fewer parameters.
• World Models & Embodiment: We are moving toward “World Action Models” (WAMs) that simulate the consequences of actions. WorldBagel: Uncovering the Power of Unified Multimodal Models for Vision-Language-Action-World Modeling, Multiplayer Interactive World Models with Representation Autoencoders, MoP-JEPA: Hard-Assigned Predictor Mixtures for Stochastic JEPA World Models, ABot-M0.5: Unified Mobility-and-Manipulation World Action Model, and WAM4D: Fast 4D World Action Model via Spatial Register Tokens define this frontier. Humanoid Everyday: A Comprehensive Robotic Dataset for Open-World Humanoid Manipulation, iFLYTEK-Embodied-Omni Technical Report, and R3D: Quantitative 3D Spatial Reasoning for Egocentric Wearables provide the data and spatial grounding necessary for real-world robotics. G3Splat: Geometrically Consistent Generalizable Gaussian Splatting, SharpSplat: Edge-Regularized 3D Gaussian Splatting for High Fidelity Urban Building Reconstruction from UAV images, and Argus: Metric Panoramic 3D Reconstruction for Indoor Scenes ensure that 3D reconstructions respect geometric consistency.

Theme 2: Agentic Reasoning, Verification, and Governance

We are moving from “reactive” models to autonomous agents that plan, verify, and correct themselves. This shift necessitates a “control plane” for governance and formal methods for safety.

• Agentic Workflows: AutoResearch: An Execution-Grounded Multi-Agent Framework for Reliable Research Workflow Automation, MechMath Agent Team: LLM Driven Agents for Mathematical Research, and BioProVLA-Agent: An Affordable, Protocol-Driven, Vision-Enhanced VLA-Enabled Embodied Multi-Agent System with Closed-Loop-Capable Reasoning for Biological Laboratory Manipulation demonstrate agents navigating complex scientific cycles. Forethought: Verifiable Reasoning from Neurosymbolic Primitive Programming and Heaviside Continuity of Rolling Coefficients for Eliminating Epistemic Entropy in Large Language Models introduce verification-first execution.
• Governance & Security: AGL-1: The Enterprise AI Governance Layer as a Control Plane for Trusted Enterprise Intelligence, CAGE-1: Control, Assurance, and Governance Evaluation for Enterprise Agentic AI, and AgentLTL: A Trace-Verification Framework for Measuring, Enforcing, and Training Procedural Compliance in Tool-Using LLM Agents provide the necessary guardrails. DualView: Preventing Indirect Prompt Injection in Personal AI Agents and Securing Multi-Tool AI Agent Chains With Dynamic, Real-Time Compositional Policies address security in multi-tool chains.
• Formal Verification: Why3-py: A Tool for Formal Verification of Hypothesis Testing and Meta-Analysis in Python and NormWorlds-CF: Solver-Verified Counterfactual Normative Reasoning with Metamorphic-Relation GRPO push toward mathematically verifiable logic.

Theme 3: Efficient Inference and System-Level Optimization

As models grow, the cost of inference becomes a primary constraint. Researchers are optimizing the full stack—from quantization and pruning to hardware-level execution.

• Compression & Acceleration: Variable Bit-width Quantization: Learning Per-Group Precision for “Bigger-but-Smaller” Language Models, Full-Stack FP4: Stable LLM Pretraining with Quantized Projections, Optimizers, and Attention, PuzzleMoE: Efficient Compression of Large Mixture-of-Experts Models via Sparse Expert Merging and Bit-packed inference, and ARCQuant: Boosting NVFP4 Quantization with Augmented Residual Channels for LLMs reduce memory footprints. AdaptiveSD: A Stability-Aware, Runtime-Adaptive Speculative Decoding Framework, KVpop – Key-Value Cache Compression with Predictive Online Pruning, and DSpark: Confidence-Scheduled Speculative Decoding with Semi-Autoregressive Generation optimize the KV cache and decoding process.
• System Infrastructure: BluTrain: A C++/CUDA Framework for AI Systems, PEEK: Predictive Queue-Informed KV Cache Management for LLM Serving, and Sangam: Efficiently Serving Diffusion LLMs with the AR Stack focus on the infrastructure layer to reduce latency and costs.

Theme 4: Privacy, Reliability, and Rigorous Evaluation

We are entering an era where “accuracy” is insufficient; we require systems that quantify uncertainty, protect privacy, and undergo rigorous auditing.

• Privacy & Federated Learning: Federated Learning for Object Detection: Enabling Collaborative Drone Learning Without Centralizing Data, F-ACVAE: A Federated Adaptive Conditional Variational Auto-Encoder for Privacy-Preserving Intrusion Detection in IoT Networks, FedACT: Federated Adaptive Coordinate Trust Modulation for Robust Transformer Training under Data Heterogeneity, and SpecGradFilter: A Spectral Gradient Filtering Framework for Taming Federated Heterogeneity enable decentralized intelligence.
• Uncertainty & Auditing: Trading Confidence: Comprehensive Uncertainty Estimation in Algorithmic Trading, Robustness Meets Uncertainty: Evidential Adversarial Training for Robust Selective Classification, and Evaluating LLM Uncertainty in Long-Form Generation Using Deterministic Ground Truth build systems that “know what they don’t know.” Auditing the Audit: Five Failure Modes in Benchmark-Validity Audits, QEDBENCH: Quantifying the Alignment Gap in Automated Evaluation of University-Level Mathematical Proofs, and ToolFailBench: Diagnosing Tool-Use Failures in LLM Agents demand a higher standard of evaluation.
• Medical & Forensic Trust: PRIMA: Pre-training with Risk-integrated Image-Metadata Alignment for Medical Diagnosis via LLM and AgentFoX: LLM Agent-Guided Fusion with eXplainability for AI-Generated Image Detection demonstrate the application of these rigorous standards to high-stakes medical and forensic domains.