ArXiV ML/AI/CV papers summary

Theme 1: Agentic Reasoning, Orchestration, and Tool-Use

The frontier of AI has shifted from static, monolithic models to agentic systems—dynamic architectures that interleave reasoning, tool use, and environment interaction. To manage the “decision-space explosion” inherent in complex tasks, research is moving toward hierarchical, stack-based execution where agents manage memory and capabilities through structured, verifiable paths.

• Architectural Evolution: A Formal Hierarchical Architecture for Agentic Orchestration with Stack-Based Execution and Lazy Discovery and GRADE: Gated Routing and Adaptive Depth for Efficient Reasoning propose tree-based and gated routing to scale reasoning depth. StructAgent: Harness Long-horizon Digital Agents with Unified Causal Structure emphasizes that long-horizon tasks require a unified causal representation to ensure progress is verifiable.
• Reasoning Dynamics: Interpreting Latent CoT Reasoning as Dynamical Systems models reasoning traces as trajectories in representation space, while Track, Rank, Crack: Epistemic Working Memory Scales Multi-Hop Reasoning in Language Agents introduces explicit epistemic memory to prevent context dilution.
• Verification & Grounding: Evidence-Grounded Verified Agentic Reasoning: A Path Toward Eliminating LLM Hallucination in Empirical Inference via Tool-Attested Kernel Proofs and A Neurosymbolic Approach to Natural Language Formalization and Verification utilize formal kernels to ensure outputs are structurally derived from evidence. STEC: Evidence Compression for Deep Search in Open-domain Multi-Hop QA and STAMP: Provenance-Guided Credit Assignment for Deep Search Agents focus on the provenance of information to solve reward-credit mismatches.
• Interaction Scaling: Interaction Scaling: Grounding the Third Axis of Test-Time Compute argues that true breakthroughs require models to propose artifacts and revise based on real-world feedback, a concept supported by Self-Verified Reasoner (SVR-R1) and ToMap: Efficient Test-Time Optimization for Multi-Agent Proof Autoformalization.

Theme 2: Embodied Intelligence and Spatial Reasoning

As AI enters the physical world, it must move beyond token prediction to master gravity, geometry, and social norms. This requires “spatial-aware” models that treat the physical world as a first-class citizen.

• Spatial & World Modeling: World Narrative Model for Highly Controllable Video Generation: A Paradigm Shift from Pixel Sampling to Physical World Orchestration, Authoring for Living Worlds: Tool-Constrained LLM Agents for Executable Multi-Actor Scenarios, and ABot-3DWorld 0: A Universal World Model to Explore Any 3D Space treat generation as a controllable, structured process. FlowWAM: Optical Flow as a Unified Action Representation for World Action Models bridges video generation and control.
• Geometric Priors: 2.5-D Decomposition for LLM-Based Spatial Construction offloads vertical placement to deterministic executors. FoundationGeo: Learning Spatial Pixel-Wise Fields for Monocular Metric Geometry, Spherical-GOF: Geometry-Aware Panoramic Gaussian Opacity Fields for 3D Scene Reconstruction, You Only Gaussian Once: Controllable 3D Gaussian Splatting for Ultra-Densely Sampled Scenes, and TRIG: Trajectory-Rig Decoupled Metric Geometry Learning embed physical constraints to ensure metric accuracy.
• Biological & Social Integration: TouchThinker: Scaling Tactile Commonsense Reasoning to the Open World with Large-scale Data and Action-aware Representation introduces tactile reasoning, while NormAct: A Benchmark for Hidden Social Norm Compliance in Embodied Planning and Look, Focus, Act: Efficient and Robust Robot Learning via Human Gaze and Foveated Vision Transformers incorporate social norms and human-inspired perception. Seeing Through Uncertainty: Free-Energy-Inspired Real-Time Adaptation for Robust Visual Navigation applies the Free Energy Principle for real-time adaptation.

Theme 3: Governance, Safety, and Reliability

With increased autonomy comes the risk of “silent failures” and policy violations. The field is shifting toward “fail-closed” systems where human authority and verifiable evidence govern agentic behavior.

• Governance Frameworks: LOGOS: A Living Logic for AI Agent Teams That Evolve With Humans and The Compliance Trap: Diagnosing How AI Agents Consume Conflicting Memory address the dangers of conflicting memory and the need for human-controlled policy checks. Filtering Harmful Actions Isn’t Enough: Phantom Transfer in Agentic SDF warns against diffuse misaligned dispositions.
• Deterministic Guardrails: Reason Less, Verify More: Deterministic Gates Recover a Silent Policy-Violation Failure Mode in Tool-Using LLM Agents and A Low-Latency Fraud Detection Layer for Detecting Adversarial Interaction Patterns in LLM-Powered Agents provide real-time defense mechanisms. Safety from Honesty in a Disinterested AI Predictor explores training models as disinterested predictors to mitigate agency risks.
• Robustness & Fairness: VanillaBench: The Hidden Accuracy Cost of Adversarial Robustness, Attention Misses Visual Risk: Risk-Adaptive Steering for Multimodal Safety Alignment, and Correcting Visual Blur Induced by Attention Distraction to Reduce Hallucinations: Algorithm and Theory highlight the trade-offs between robustness and performance. Understanding Sources of Demographic Predictability in Brain MRI via Disentangling Anatomy and Contrast and Controllable Generation of Diverse Dermatological Imagery for Fair and Efficient Malignancy Classification provide frameworks for mitigating bias in clinical AI.

Theme 4: Scientific Machine Learning and Domain-Specific Systems

Specialized agentic frameworks are moving beyond general-purpose chat to solve high-stakes problems in science and engineering by embedding physical laws directly into neural architectures.

• Physics-Informed Learning: MUSA-PINN: Multi-scale Weak-form Physics-Informed Neural Networks for Fluid Flow in Complex Geometries, A hybrid analytical-PINN model for subsurface simulation of geothermal heat exchangers in heterogeneous underground, Near-Optimal Learning of Gaussian Sobolev Operators, and Deep Learning-based Surrogate Modelling of the LOD Method for Multiscale Problems advance operator learning and PINNs. Exact and Calibrated Diffusion Reconstruction for Digital Breast Tomosynthesis and Real-time fall detection based on vision for low-power edge platforms apply these principles to clinical and safety-critical tasks.
• Generative Science: BattVAE-GP: Generative Modeling of Long-Horizon Battery Degradation with Uncertainty Quantification and Generating Developable 3D Molecules via Pocket-Conditioned Diffusion and Property-Aware Optimization optimize for complex physical properties.
• Domain Frameworks: NVAITC AI Scientist: A Governed End-to-End Research System – A Hypertension GWAS Case Study, QwenPaw-Data: Bridging Facts, Methodology, and Execution for Autonomous Enterprise Data Analytics, Imaging-101: Benchmarking LLM Coding Agents on Scientific Computational Imaging, and Opti-Agent-Bench: Benchmarking End-to-End Optimization R&D Agents on Real-World Business Problems demonstrate the efficacy of specialized, governed agents.

Theme 5: Memory, Unlearning, and Efficient Inference

As models grow, managing state, privacy, and computational overhead has become a critical engineering challenge.

• Memory & Unlearning: Signal-Guided Optimization for Machine Unlearning, Inference-Time Machine Unlearning via Gated Activation Redirection, and Do You Remember? Toward Memory-Centric Multimodal AI explore how to edit, forget, and store information. Reducing information dependency does not cause training data privacy. Adversarially non-robust features do and Extractable Memorization From First Principles provide rigorous frameworks for privacy and memorization.
• Optimization & Efficiency: A JoLT for the KV Cache: Near-Lossless KV Cache Compression via Joint Tucker and JL-Residual Allocation for LLMs, LiteTopK: Exploiting the Curse of Dimensionality for a Fused Indexer-TopK Kernel in Long-Context Sparse Attention, CARE-LoRA: Compressed Activation REconstruction for Memory-Efficient LoRA, and dMX: Differentiable Mixed-Precision Assignment for Low-Precision Floating-Point Formats optimize memory and inference.
• Scaling & Kernels: GrandCode: Achieving Grandmaster Level in Competitive Programming via Agentic Reinforcement Learning, Recursive Multi-Agent Systems, EmbodiSkill: Skill-Aware Reflection for Self-Evolving Embodied Agents, and FARS: A Fully Automated Research System Deployed at Scale push the boundaries of autonomous research and skill evolution. CUDA-L2: Surpassing cuBLAS Performance for Matrix Multiplication through Reinforcement Learning, Gefen: Optimized Stochastic Optimizer, and Less Experts, Faster Decoding: Cost-Aware Speculative Decoding for Mixture-of-Experts focus on hardware-level efficiency.