ArXiV ML/AI/CV papers summary

This collection of research marks a profound transition in machine learning: we are moving away from the “brute force” era of scaling parameters toward an era of structured cognition. We are no longer merely asking if a model can generate a plausible answer; we are demanding that it prove its work, manage its own resources, and remain reliable under the pressures of the physical and social world.

Theme 1: Agentic Reasoning and Self-Evolution

The field is shifting from passive chatbots to autonomous agents that plan, execute, and self-correct. The primary challenge is moving beyond “stochastic mimicry” to reliable, multi-step reasoning.

• Workflow Routing & Efficiency: Systems are now learning to optimize their own reasoning paths. Learning Compositional Meta-Routing for Agentic Workflows: An Executable Benchmark and MetaRoute-Bench: Evaluating Meta-Decision Policies for Agentic Workflow Routing show that learned routing policies significantly reduce latency. Similarly, AOSpec: Action and Observation Co-Speculation for Low-Latency Agent Serving and Global Optimization and Inference-Time Region Grafting for Agentic Workflows allow agents to speculate on outcomes and repair failed reasoning segments at runtime.
• Self-Improvement & Memory: Agents must manage their own “brain” to avoid memory hysteresis. Self-Improving Large Language Models via Progressive Experience Evolution and CrystalMem: Elastic Memory for Self-Evolving LLM Agents via Knowledge Crystallization introduce mechanisms for pruning and demoting memory, ensuring agents remain sharp under strict byte budgets.
• Robustness: As agents become more autonomous, they face new threats like “Reasoning-Level DoS” (OTora) and “intent noise.” Frameworks like IACM-RL and Agentic Self-Healing for Data and AI Pipelines provide the necessary resilience for agents to maintain performance in complex, real-world pipelines.

Theme 2: Verification, Auditing, and Scientific Rigor

We are entering an era of “provenance-by-design,” where terminal accuracy is insufficient; we must audit the topology of the reasoning process to ensure correctness.

• Formal Verification: By interfacing LLMs with symbolic solvers and formal verifiers, we can achieve mathematical certainty. Learning to Coordinate Symbolic Tools: LLM Agents for Verified Sum-of-Squares Certificates and MechGeo: Autoformalizing and Proving Euclidean Geometry in Lean 4 represent this new standard of rigor.
• Auditing & Safety: Tracing the Cascade: A Topology-Aware Evaluation Framework for Scientific Agent Hallucinations and Auditing Discovery Claims: A Two-Sided Criterion for Agentic Science provide the “scientific method” for AI. Furthermore, Coding Agents as Test-Suite Auditors and EntailLLM demonstrate how agents can audit benchmarks and vulnerability paths, ensuring consistency with domain knowledge.

Theme 3: Physics-Informed and Embodied Intelligence

AI is stepping out of the screen and into the physical world, requiring models that respect the laws of nature rather than treating them as black-box correlations.

• Scientific Discovery: A Physics-Chemistry-Informed Neural Network (PCINN) and CoSynFlow demonstrate how hard-coding physical constraints allows for real-time simulation with CFD-level accuracy. HyperODE and Convex Neural Energy Elements further bridge the gap between neural operators and traditional engineering-grade simulations.
• Embodied AI: ProtoAct and MANGO-Grasp tackle the translation gap between human protocols and robotic execution, while PhyCheck forces models to move beyond surface-level recognition to true physical reasoning.

Theme 4: Efficiency, Compression, and Hardware Co-Design

The “compute-at-all-costs” era is being replaced by a focus on “carbon-frugal” and hardware-aware optimization.

• Extreme Compression: SparseKAN and BiKAN explore parameter-efficient alternatives to MLPs, while Kilobyte Models proposes storing models as “recipes” rather than weights.
• Inference Acceleration: WaveZip and OmniScope utilize modality-specific token condensation to bypass the quadratic cost of attention. Meanwhile, DCC and HERALD show that massive speedups are possible by co-designing software kernels with hardware memory access patterns.

Theme 5: The Human-AI Interface and Social Dynamics

As AI becomes a participant in social and economic systems, we must address the behavioral challenges of human-AI collaboration.

• The Scaling Paradox: The Scaling Paradox in Human-AI Collaboration and Large language models improve physician accuracy but lead to false reliance warn that increased AI capability can lead to human performance decline due to automation bias.
• Social Simulation: No One Wins in Nuclear War and CompanionBench provide the infrastructure to study AI as a social participant, grounding evaluation in psychological theories of relational competence rather than just “surface warmth.”