ArXiV ML/AI/CV papers summary

Theme 1: The Rise of Agentic Intelligence and Process-Awareness

The field is undergoing a fundamental shift from “black-box” prediction to “glass-box” reasoning. We are moving away from models that simply output a token toward agentic systems that actively refine their strategies, diagnose their own failures, and operate within structured, auditable frameworks.

• Self-Evolving Agents: Models are increasingly using world models to simulate outcomes before taking action, as seen in WMLLM: Self-Evolving Optimization Agents via Predict-Then-Act World Modeling. This “Test-Time Intelligence” (TTI) paradigm, explored in A Survey on Self-Improving Test-Time Intelligence: Feedback-Driven Adapting, Learning, and Scaling at Inference, allows models to adapt and scale their reasoning on the fly.
• Credit Assignment & Reliability: To solve long-horizon tasks, we must move beyond simple outcome-based rewards. Cliff: Learning Process Rewards from the First Mistake and Coverage, Not Targeting: A Structural Regime in Multi-Turn Agent Credit Assignment emphasize fine-grained supervision. Furthermore, Beyond Outcome Gaps: Process-Aware Fairness Diagnosis for LLM-based Multi-Agent Decision Systems and Diagnosing with Insights: Structured Analysis of Agent Failures via Behavioral Abstractions treat decision trajectories as structured data to uncover hidden biases and reasoning drifts.
• Production-Grade Deployment: The focus is shifting toward the “reliability budget” of AI. READY or Not: Reliable Enterprise Agent Deployment and How Fast Do Agents Rot? An Empirical Study of Long-Horizon Degradation in LLM Agents for Production Decision-Making highlight that current benchmarks often mask performance decay, necessitating new architectures for long-term stability.

Theme 2: Procedural Memory and Skill Distillation

Just as humans rely on muscle memory, AI systems are moving toward distilling reusable, procedural knowledge rather than relying on monolithic context windows.

• Skill Consolidation: SkillGLoW: Procedural-Family Skill Consolidation for Self-Improving Agents on Long-Horizon Task Streams demonstrates that aggregating skills into “procedural families” allows for better transfer of methods.
• Operational Knowledge: By distilling complex repositories into verified skills, models can perform research tasks with higher precision. Key research includes Repo-To-Skill: Distilling GitHub Repositories Into AI4AI Skills and APEx: Distillation of Agent Procedural Experience for Adaptive Deep Research Question Answering.

Theme 3: Physics-Awareness and Geometric Grounding

Intelligence is not just statistical; it is physically situated. Modern models are increasingly incorporating geometric priors and physical laws to navigate the world and interpret data more accurately.

• Navigation and Perception: KSG-Net: Key-Sparse and Global-Context Learning for Maritime 3D Ship Detection and If It Moves, Radar Knows: A Physics-Aware Radar Transformer for Class-Agnostic Moving-Object Detection show that radar and LiDAR data are best processed when the model understands the underlying physics.
• Geometric Consistency: Self-Geometry: GT-Free and Plug-and-Play Test-Time Adaptation for Geometrically Consistent 3D Vision Foundation Models and Geometric Distillation from Rectified Stereo: Leveraging Epipolar Cues for Monocular Depth teach models to respect multi-view constraints without needing expensive ground-truth labels.
• World Simulators: Generative models are evolving into “World Action Models” (WAMs). World-Coherent Decoding: Self-Verifying Test-Time Planning for World Action Models, SolarWM: Open Data and Scalable Training for Long-Horizon Video World Models, and Spatially Aware World Action Model via Geometric Latent Diffusion allow robots to simulate the physical consequences of their actions in 3D space.

Theme 4: Efficient Inference and Architectural Innovation

As models grow, we are finding ways to make them leaner and more interpretable through structural compression and fundamental architectural shifts.

• Compression & Efficiency: CRISP: Cliff-awaRe Input-adaptive Sparse Prefilling with Structural-Mass-Motivated Routing and Emergence of Fibrations, Compression, and Symmetry Breaking in Artificial Neural Networks provide theoretical and practical paths to reducing model size and computational overhead. hLLM: Single Pass Decoding for Generative Reranking further optimizes inference by treating ranking as a combinatorial problem.
• New Foundations: Research into RecKAN: Kolmogorov-Arnold Networks with a Learnable Recursive Polynomial Basis, FlashKAN: B-Spline KANs via Truncated Power Form, and On the Expressive Power and Limitations of Multi-Layer SSMs explores alternatives to the standard Transformer, aiming for better interpretability and performance.

Theme 5: Safety, Alignment, and Forensic Integrity

Safety is transitioning from reactive filtering to “isolation-by-construction,” where architectural boundaries prevent harm.

• Architectural Safety: Isolation as a First-Class Principle for LLM-Agent System Safety: Concepts, Taxonomy, Challenges and Future Directions and PRISM: An Agentic Multi-Model Architecture for Proactive Safety in Autonomous Transportation Systems define the security perimeters necessary for high-stakes environments.
• Adversarial Resilience: As models become more capable, so do attackers. ASCII Attack: Recontextualising Harmful Requests as Artistic Critique in Large Language Models and Jailbreaking Text-to-Image Models Through Cracks: Navigating Heterogeneous Safety Filters via Multi-Agent Debate highlight the need for robust, multi-agent stress testing.
• Forensics: From Detection to Localization: A Unified Forensics Framework for Fully Synthetic and Tampered Images, Evidence-Guided Detection, Localization and Explanation for Text-Centric Image Forensics, and Multi-Tool Image Editing Attribution in Facial Forgery provide the tools necessary to maintain digital provenance in an era of synthetic media.

Theme 6: Domain-Specific Reasoning and Scientific Discovery

AI is increasingly grounded in specialized domains, where interpretability and data quality are paramount.

• Scientific Discovery: Mol-JEPA: A multimodal Joint Embedding Predictive Architecture for Molecules and HiPoly: a hierarchical polymer-native AI framework for property prediction and generative design integrate physical context into latent representations. A Computational Comparison of Fourier Spectral Differentiation and Spatial Automatic Differentiation in Periodic Physics-Informed Neural Networks optimizes training for physics-informed models, while When Literature Data Mislead Artificial Intelligence in Materials Discovery warns of the dangers of structured noise in scientific data.
• Computational Pathology: The Diagnosis a Reporter Leaves Unspoken: Surfacing Frozen Tumor Features for Brain-Tumor MRI Reporting, Synergistic Information Disentanglement for Omni-modal Slide Representation Learning in Computational Pathology, and AtlasPatch: Scalable Foundation Model-based Tissue Detection and Patch Extraction for Computational Pathology focus on evidence-based, transparent medical reporting.
• Cultural Competence: VakyArth: Evaluating Pragmatic Competence in LLMs across Indic Languages and MemeCULT-1K: Benchmarking South Asian Cultural Context and Humor Understanding of Multimodal Models remind us that true intelligence must be culturally situated.