ArXiV ML/AI/CV papers summary

The evolution of machine learning is currently undergoing a profound transformation. We are witnessing a departure from the “bigger is better” era of monolithic scaling toward a more sophisticated paradigm: the rise of System 2 AI. Much like the transition from simple observation to the rigorous, evidence-based methods of modern cosmology, our field is moving from models that merely guess to agents that know, verify, and explain.

Here is the synthesis of the current research landscape.

Theme 1: Agentic Reasoning, Reliability, and Self-Evolution

The frontier of AI is shifting toward autonomous agents capable of recursive self-improvement and complex, multi-step reasoning. The focus is no longer on static performance, but on the ability of an agent to verify its own logic and evolve its skill set.

• Recursive Self-Improvement: Agents are increasingly bootstrapping their own capabilities. Memento 3: Model-Based Recursive Self-Improvement through Reflective Rulebooks, Ghost tasking for parametrized Gaussian Processes solving linear differential equations, The Red Queen G"odel Machine: Co-Evolving Agents and Their Evaluators, and AgentEvolver: System-Wide Self-Evolution Through Task Execution demonstrate how agents can maintain persistent knowledge and accumulate reusable capabilities without constant retraining.
• Verification and Credit Assignment: To ensure reliability, we are moving toward smarter credit assignment and process verification. Learning from Own Solutions: Self-Conditioned Credit Assignment for Reinforcement Learning with Verifiable Rewards, MetaOPD: Meta-Learned Token Weighting for On-Policy Distillation, TestJack: Should you trust the results in coding benchmarks? Agentic Coding Benchmarks Auditing via Evaluator Evolution, and Beyond Type-checking: Towards Holistic Evaluation of Formal Specification Generation provide frameworks to audit reasoning and prevent “benchmark hacking.”
• Skill Evolution: ViSkill: Reinforcing VLM Agents with Evolving Visual-Native Skills, SkillGraph: Skill-Augmented Reinforcement Learning for Agents via Evolving Skill Graphs, and Skill-V: Verifiable Self-Evolving Skill Library for Interactive Agents treat skills as falsifiable contracts, allowing agents to compose complex behaviors while maintaining reliability.
• Adaptive Reasoning: Agents are learning when to think. When Should Agents Think? Adaptive Reasoning via Cross-Turn Estimation and From Chain-of-Thought to Loops: Non-Autoregressive Latent Reasoning via Looped Transformers optimize reasoning latency, while Plan-and-Patch: Diffusion Language Models for Agentic Planning allows for efficient plan repair.

Theme 2: Grounded World Models and Physical Simulation

As AI enters the physical world, it must move beyond token prediction to “World-Action Models” (WAMs) that understand physics, geometry, and causality.

• Physics-Informed Learning: Integrating physical laws transforms AI into a scientific instrument. Gen-PINNs: Generative Adversarial Physics Informed Neural Networks for solving partial differential equations, F$^3$NO: Frequency-Decomposed Finite-Time Flow-map Neural Operators with Cross-Scale Conditioning, Zatom-2: Multitask Pretraining on Atomistic Data for Generative Modeling across Domains, and NEMORA: Neural Equivariant Multipole Operators for Long-Range Atomistic Learning push the boundaries of scientific discovery.
• World Modeling: WOVEN: Weaving Visual World Modeling into Multimodal LLMs, MA-JEPA: Joint-Embedding World Models for Multi-Agent Reinforcement Learning, VICO: Visual Environments Co-Evolving for Vision-Language Model Reasoning, and AffordDrive3D: Affordance-Aware World-Action Modeling with Spatial Understanding explore how models predict environmental transitions.
• Embodiment and Dynamics: Humanoid World Action Model With Joint State–Action Generation, DreamTrue: Action-Faithful Robot World Model with Counterfactual Post-Training, and 3D Point World Models: Point Completion Enables More Accurate Dynamics Learning bridge the gap between high-level planning and physical motor control.
• Scientific Discovery: LLM-IDEA: Identifiability-Driven Experimental Agent for Autonomous Discovery of Mechanistic World Models, EvoSim: Learning to Model, Modeling to Learn, Origins of Universal Machine Learning Force-Field Errors in Multicomponent Materials, and AtomWorld-Mirror: Macro-Step World Modeling of Critical Evolution Backbones for Materials Dynamics represent the new class of “AI Scientists.”

Theme 3: Efficiency, Compression, and Memory Management

To deploy these sophisticated agents, we must move toward smarter resource allocation, moving beyond dense attention to adaptive, query-dependent memory.

• KV-Cache and Memory Optimization: Freeze the Decoder, Heal the Encoder: Parameter-Efficient Adaptation for SVD-Based KV-Cache Compression, Read What Matters: Query-Adaptive Quantization for KV Caches, PageWeaver: KV-Guided Query Unions for Sparse Attention, Rehearse Everything, Remember Nothing: Attic-KV Rehearses What Will Be Read, and Draft-Guided Eviction for Training-Free KV-Cache Compression rethink how we store and access context.
• Compression and Pruning: Spectral Weight Decay: Inducing Low-Rank Structure in Neural Network Weights, Deflating the Hessian: Rank-4 W4A4 Quantization for Multimodal Diffusion Transformers, SparseDecoding: Decoding-Aware Pruning for Accurate and Efficient LLM Inference, 3BASiL: An Algorithmic Framework for Sparse plus Low-Rank Compression of LLMs, and S4oP: Operator-level Pruning of Structured State Space Models for Resource-Constrained Devices optimize model architecture.
• Adaptive Compute: BudgetPix: Compute-Adaptive Tokenization for Pixel-Space Image Diffusion, iCATS: Fast Video Generation via Interaction-Aware Sparse Attention and Timestep-Adaptive Sparsity, and MiX: Micro-Inverted-Scaling for End-to-End Low-Bit Vision-Language Model Acceleration demonstrate hardware-aware optimization.

Theme 4: Robustness, Safety, and Interpretability

As AI systems are deployed in high-stakes environments, the ability to govern behavior, ensure provenance, and “forget” sensitive information is paramount.

• Unlearning and Safety: Nullify: Null-Space Activation Steering for Training-Free LLM Unlearning, Enabling Preference-driven Unlearning in Few-step Distilled Text-to-Image Diffusion Models, Inference-Time Machine Unlearning via Gated Activation Redirection, and LLM Persona Unlearning provide methods for knowledge removal.
• Governance and Provenance: TRACE: A Governance Framework for Measuring Explainability Debt in Production AI Systems, BRANCH: Bypassing Multi-Scanner AI Guardrails, SAFETY SENTRY: Context-Aware Human Intervention via EXECUTE-ASK-REFUSE Routing, Traceable World State: A Provenance-Aware State Representation and Deterministic Replay Framework for Robotic Systems, and Proof-of-Use: Mitigating Tool-Call Hacking in Deep Research Agents ensure auditability.
• Detection and Robustness: RH-Detect: A Unified Benchmark for Reward Hacking Detection, Caught in the Act: Probes Effectively Detect Sabotage and Catch Unverbalized Deception, MRCert: Towards Post-deployment Patch Robustness Certification for Adversarially Patched Samples via Type-specific Masking, and Jailbreak Scaling Laws for Large Language Models: Polynomial-Exponential Crossover provide the tools to detect and defend against adversarial behavior.
• Interpretability: CiteVQA: Benchmarking Evidence Attribution for Trustworthy Document Intelligence, Quantifying Retriever-Generator Alignment in RAG with Local Explanations, and How Much Do Circuits Tell Us? Measuring the Consistency and Specificity of Language Model Circuits challenge our understanding of how models store and retrieve knowledge.