ArXiV ML/AI/CV papers summary

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

The paradigm of AI agents is undergoing a profound transformation. We are moving away from simple, reactive task execution toward sophisticated, stateful systems capable of long-horizon planning, self-correction, and procedural compliance. This evolution is driven by the need for observability and auditability; as agents take on high-stakes roles, we must be able to trace their decisions back to verifiable evidence.

• Architectural Innovation: Eureka: Task-Conditioned Meta-Agent Orchestration for Scientific Discovery and PolicyGuide: From Guarding One Action to Guiding the Whole Workflow for Policy-Compliant LLM Agents demonstrate that scientific and procedural success requires architectures that mirror the cognitive structure of the task.
• Verification & Diagnostics: ComponentBench: Diagnosing Component-Level Failures in Computer-Use Agents, LEDGER: Claim-to-Evidence Trace Graphs for Auditing LLM Agents, and Adversarial Review: Structured Disagreement for Grounded Agentic Code Review highlight that reliability is best achieved through structured disagreement and component-level diagnostic bridges.
• Control & Autonomy: One Gate Is Not Enough: Composing Stateful Pre-Action Controls for Agentic AI and Task-Conditioned Least-Privilege Learning for Executable Terminal and MCP Agents address the critical balance between agent capability and safety, ensuring agents operate within the bounds of least privilege.
• Self-Evolution: Agents are beginning to master their own development through SkillForge: Self-Distilling Agents for Project-Specific Issue Resolution, SkillNet: Create, Evaluate, and Connect AI Skills, and SPADE: Self-Play in Adaptive Synthetic Executable Environments, which creates a closed-loop for open-ended skill acquisition.

Theme 2: Reliability, Safety, and Governance

As AI agents permeate finance, medicine, and infrastructure, “safety” has transcended simple output filtering. It now demands systems that can prove their reasoning, adhere to verifiable constraints, and remain robust against adversarial manipulation.

• Accountability & Governance: Verifiable abstention makes AI leak diagnosis accountable in water distribution networks and Governance Records as Supervision: Verifier-Selected Self-Training for Structured Workflow Repair propose frameworks where AI actions are grounded in machine-checkable reality.
• Market & Systemic Risk: FraudBench: Stress-Testing Policy-Grounded Banking Agents Against Adaptive Fraud and Position: Collusion Risks Among AI Reasoning Agents Justify Certification Requirements for Making Market Decisions warn that history-dependent interactions and tacit collusion necessitate rigorous behavioral certification.
• Robustness & Alignment: RepSelect: Robust LLM Unlearning via Representation Selectivity and Detecting Backdoors in Object Detection via Pre-NMS Prediction Distribution Shift provide tools for deep unlearning and backdoor detection. Furthermore, Epistemic Subordination: Generative AI and the Infrastructure of Knowledge and The Epistemic Politics of AI Anthropomorphism challenge us to consider the philosophical and political implications of how AI encodes knowledge and interacts with human users.

Theme 3: Efficient Modeling and Optimization

Efficiency is the bedrock of deployment. Whether we are running models on edge devices or training trillion-token architectures, the goal is to align optimization algorithms and hardware constraints through co-design.

• Training & Scaling: DeltaMomentum: A Key-Value based Anisotropic Momentum Update via Delta Rule and Let’s Scale Step by Step: Compute-Efficient Hyperparameter Transfer for Large-Scale Mixture-of-Experts provide novel ways to accelerate training and optimize hyperparameter scaling.
• Hardware-Aware Inference: Daedalus-150M: A Convolution-Attention Hybrid Designed for CPU Inference and HBVLA: Pushing 1-Bit Post-Training Quantization for Vision-Language-Action Models demonstrate that tailoring architectures to specific hardware—from CPUs to robotic controllers—is essential for performance.
• Memory & Context: MemFuse: Multi-Source Memory Fusion from Fragmented Observations, EgoCITE: Context-Augmented Indexing and Time-Aware Retrieval for Long-Horizon Egocentric Memory, and ReCache: Efficient KV Cache Reuse and Compression for Tool-Augmented LLM Agents optimize how models store and retrieve information over long horizons.

Theme 4: Scientific Discovery and Domain-Specific AI

AI is evolving into a powerful scientific instrument, capable of automating discovery in chemistry, physics, and medicine by incorporating physical laws and domain-specific inductive biases.

• Scientific Methods: SeisEvo: Evolution of Seismic Data Reconstruction Algorithms by Agents, Science Done on a Machine by a Machine: AI Agents in Computational Chemistry, and ProteinZero: Self-Improving Protein Generation via Online Reinforcement Learning show agents discovering new methods rather than just optimizing weights.
• Physics-Informed Modeling: Guided Diffusion by Optimized Loss Functions on Relaxed Parameters for Inverse Material Design, Physics-Unrolled Neural Operator for Wireless Field Modeling, and GCNO: Gramian Chebyshev Neural Operator for Physics-Based Compression of Wireless Channels integrate physical laws directly into neural operators.
• Clinical Intelligence: Pedagogical AI in Mental Health: A Tri-Stream Fine-Tuned LLM Framework for Automated Clinical Supervision and Risk Triage, Teaching agentic AI to learn expert reasoning for rare disease diagnosis, and MKG-CARE: Case-Aware Reasoning with Multimodal Knowledge Graphs for Explainable Medical Image Diagnosis demonstrate AI acting as a partner in high-stakes clinical environments.

Theme 5: Embodied AI, Robotics, and Multimodal Grounding

The frontier of robotics and multimodal AI lies in grounding reasoning in physical reality and visual evidence. This requires models that understand geometry, physics, and temporal consistency.

• Robotic Manipulation: What Matters for Latent Actions in Robot Learning, SoftVTBench: A Safety-Aware Visuo-Tactile Benchmark for Physically Constrained Robotic Manipulation of Deformable Objects (Early Version), and SafeBranch: Branch-Pair Safety Alignment for Embodied Agents focus on teaching robots to navigate physical constraints safely.
• Multimodal Reasoning: Video Evidence to Reasoning Efficient Video Understanding via Explicit Evidence Grounding and SportsGrounder: Proposal-Aided Interleaved Grounding for Dense Sports Video Reasoning emphasize the “Chain of Evidence” to reduce hallucinations.
• Generative Synthesis: OmniCamera: A Unified Framework for Multi-task Video Generation with Arbitrary Camera Control, 4DAnyone: Create Anyone in 4D from a Casual Monocular Video, and Manifold Drift in Flow Preference Optimization: A Root Cause of Reward Hacking push the boundaries of temporally consistent, identity-preserving generation.
• Physics & Geometry: ReynoldsFlow: Physics-Inspired Spatiotemporal Flow Representation for Video Understanding and Functionalization via Structure Completion and Motion Rectification illustrate how incorporating geometry and biomechanics allows AI to understand the functional properties of the physical world.