ArXiV ML/AI/CV papers summary

Theme 1: Agentic Workflows and Autonomous Discovery

The frontier of AI is shifting from passive chatbots to autonomous agents capable of complex, multi-step reasoning. This transition necessitates a move toward “agentic reliability,” where systems must manage long-horizon tasks while maintaining operational security (OPSEC) and safety.

• Autonomous Discovery: EvoPINN: Agentic Discovery of Executable Algorithms for Physics-Informed Neural Networks showcases agents discovering new scientific computing mechanisms, while Can AI agents conduct open-ended AI research? Early evidence from two case studies offers a sobering reality check on current limitations in judgment. Crucially, One Run Is Not an Idea: The Implementation Lottery in Automated Research warns that experimental outcomes are often sensitive to specific sampled implementations, necessitating multi-implementation validation. FirstResearch: Auditable Question Formation for LLM Scientific Discovery Agents proposes “Research Question Certificates” to make the discovery process transparent and inspectable.
• Workflow and Security: Agent-UCT: Upper Confidence Bounds Applied to Trees for Agentic Workflow Optimization with Cost-Awareness and IDP AutoOpt: Agent-Driven Optimization of Document Processing Pipeline Configurations optimize RAG and document pipelines. On the security front, GoGoTB: Agentic RTL Verification with Specification-Grounded Coverage Closure aids hardware verification, while AgentSnare: Learning to Delay, Divert, and Defuse Autonomous Penetration Agents and StealthBench: Measuring Operational Stealth in Autonomous Offensive-Security Agents address the defensive and operational risks of agentic autonomy.
• Governance: Intent-Governed Tool Authorization for AI Agents and Explanation-Bound Tool Execution for AI Agents: Server-Verified Action Claims Without Trusting Model Rationales advocate for moving away from static permissions toward intent-based verification.

Theme 2: Memory, Context, and Long-Horizon Reasoning

As models tackle longer workflows, memory management has become a primary bottleneck. Research is moving from external retrieval modules to native, persistent memory architectures.

• Architectural Innovations: Metis: Memory Foundation Model proposes a native, persistent memory state, while MemAgent: Reshaping Long-Context LLM with Multi-Conv RL-based Memory Agent optimizes long-text processing. Filesystem-Based Memory for LLM Agents: Organization, Evolution, and Sustainability and MemSecBench: Tracking Agent Memory Poisoning from Persistence to Consequence and Repair highlight the risks of memory poisoning and the need for better organizational structures.
• Serving and Efficiency: InferScale: GPU-Native KV Injection for Personalized LLM Serving and FinCacheServe: Dependency-Consistent Answer Reuse for Cost-Efficient RAG Serving over Mutable Enterprise Documents optimize serving through state reuse. RedKnot: Efficient Long-Context LLM Serving with Head-Aware KV Reuse and SegPagedAttention and Keyless Attention: Value-Space Routing and Value-Only Caching for Efficient Transformers further refine KV cache management.
• Stability: Dense Local Dependencies Induce Attention-Logit Explosion and Training Instability During Long-Sequence Transformer Training identifies the root causes of instability in long-sequence training.

Theme 3: Embodied Intelligence and Physical AI

The transition from digital to physical environments requires “embodiment awareness”—the understanding that physical constraints fundamentally alter the meaning of visual and sensory input.

• Navigation and Control: EA-Nav: Learning Safe Visual Navigation Policies with Embodiment Awareness injects morphology as a conditional token, while Walk through Paintings: Egocentric World Models from Internet Priors repurposes video models for motor-conditioned world modeling. Reinforcement Learning on Cost-Constrained Quadrupedal Hardware and Balancing Centralized Learning and Distributed Self-Organization: A Hybrid Model for Embodied Morphogenesis explore biological strategies for overcoming latency and noise.
• Infrastructure: PUDA: An AI-Native Hardware Harness for Self-Driving Laboratories and BioVLN: A Simulation Platform for Visual Language Navigation in Biomedical Laboratories provide the necessary infrastructure for physical lab interaction. LabRobFail: A Benchmark for Robotic Failure Analysis in Chemical Self-driving Laboratory and LLM-Grounded Dynamic Task Planning with Hierarchical Temporal Logic for Human-Aware Multi-Robot Handover bridge the gap between high-level instructions and low-level execution.

Theme 4: Alignment, Safety, and Behavioral Modeling

Alignment is evolving from a post-training fix into a continuous, system-level requirement, often informed by psychometric frameworks.

• Alignment Frameworks: Constitutional Midtraining: Content Presence Drives Alignment Gains and Towards Verifiable Transformers: Solver-Checkable Circuit Explanations push for more durable, verifiable alignment. ToxScreen: Detecting Whether an LLM Has Been Poisoned and Lilith: Backdoor Generalization under Training-Inference Trigger Shift explore adversarial defenses, while Adaptively Robust LLM Monitoring via Activation Watermarking provides a method for detecting misuse.
• Psychometrics: Misalignment Has a Personality: A Big Five Account of Emergent Misalignment and From Representations to Behaviors: Exploring the Person-Situation-Behavior Triad in LLMs treat misalignment as a personality shift. When Synthetic Users Fail: A Cross-Domain Benchmark of LLM-Simulated Human Survey Responses and Aligning LLM-Simulated and Human Examinees for Psychometric Calibration: A Cognitive Diagnostic Profiling Approach evaluate the efficacy of synthetic users.

Theme 5: Multimodal Perception, Generative Synthesis, and Domain Expertise

As generative models become more precise, the focus shifts toward domain-specific adaptation and fine-grained multimodal reasoning.

• Perception and Reasoning: LOCUS: Local Visual Cue Search for Enhancing Fine-Grained Perception in Multimodal Large Language Models and SMSP: A Plug-and-Play Strategy of Multi-Scale Perception for MLLMs to Perceive Visual Illusions improve fine-grained perception. VidNum: Diagnosing VLM Failure Modes in Video-Grounded Numerical Reasoning and POINTS-Seeker: An Open Recipe for Multimodal Search Agents with Visual Memory Management address reasoning across time and modalities.
• Generative Control: SceneExpander: Text-Guided 3D Scene Expansion via Free-Form View Insertion and Explicit Layer Modeling for Video Object Insertion and Layer Decomposition enable surgical editing. AptAvatar: Fast and Vivid Long-Form Audio-Driven Video Generation for Production-Ready Avatars and Calibri: Enhancing Diffusion Transformers via Parameter-Efficient Calibration focus on production-grade efficiency.
• Domain Adaptation: SPROUT: A Scalable Diffusion Foundation Model for Agricultural Vision and SpectraDINO: Modality-Conditioned Adaptation of RGB Vision Foundation Models Across Infrared Bands demonstrate domain-specific foundation models. PIKS: Universal Physics-Informed Kernel Methods and Equivariant Eikonal Neural Networks integrate physical laws into learning, while ClinLens: Towards Long-Horizon Coding Agents for Longitudinal Multimodal Clinical Data Science and BrainG3N: A Dual-Purpose Tokenizer for Controllable 3D Brain MRI Generation apply these techniques to medicine.