ArXiV ML/AI/CV papers summary

Theme 1: Agentic Self-Evolution and Recursive Improvement

We are witnessing a fundamental shift from static, “deploy-and-forget” models to dynamic systems capable of recursive self-improvement (RSI). By moving beyond simple token prediction, these agents are beginning to manage their own cognitive architectures and safety protocols.

• Recursive Self-Improvement: Research into multi-agent topologies, such as Recursive Self-Improvement through Multi-Agent Self-Supervision and ReSI: Recursive Safety Improvement toward Resistant and Resilient AI, demonstrates how models can bootstrap their own capabilities by critiquing their reasoning and alignment. This is complemented by Thought-Like-Pro: Enhancing Reasoning of Large Language Models through Self-Bootstrapped Prolog-based Chain-of-Thought and DENSE: Distilling Agent Trajectories into Evidence-Grounded Shortcut Trees for Self-Refinement, which distill successful reasoning traces into reusable feedback.
• Harness and Architecture Evolution: Rather than retraining massive weights, we are seeing a focus on evolving the “harness”—the runtime environment and logic. Papers like The Harness as the Only Mutable Surface: Compliance-Bounded Self-Evolution of LLM Agents in Credit Pipelines, with a Measured Admission Gate, AgentEvolver: System-Wide Self-Evolution Through Task Execution, and Neural Architecture Discovery via Autonomous Evolution suggest that confining evolution to the runtime allows for auditability while maintaining adaptability.
• Co-Evolutionary Loops: SynCo: Data Synthesis Co-Training for Self-Evolving LLMs via Multi-Agent Reinforcement Learning and The Red Queen G"odel Machine: Co-Evolving Agents and Their Evaluators highlight that as agents improve, their training data and evaluation rubrics must evolve in tandem to prevent stagnation. Similarly, Agentic Critical Training and Self-Retrospection Distillation: Turning Post-hoc Experiences into Prior Foresight enable agents to turn past mistakes into future foresight.

Theme 2: Verification, Safety, and “Epistemic Humility”

As agents gain autonomy, the focus shifts from reactive guardrails to proactive assurance. We are moving toward systems that are “verifiable by construction,” capable of acknowledging their own limitations—a trait we call “epistemic humility.”

• Verification-Centric AI: To solve the “verifier’s dilemma” (preventing reward hacking), researchers are evolving the verifiers themselves. Verification and Self-Improvement in Agentic AI: Foundations and Limits and Who Verifies the Verifier? Co-Evolving Inspectable Graders with Self-Improving Agents propose using inspectable, deterministic detectors. This is bolstered by formal methods in Decidable By Construction: Design-Time Verification for Truly Fearless Systems and NOMOS: Compiling Written Policies into Statically Verified Tool-Call Gates for LLM Agents.
• Safety and Deception Detection: Safety now encompasses “obligations”—actions an agent must perform—as explored in Safe Actions Alone Do Not Ensure Safe Agents: Identifying Unfulfilled Obligations with Guard Models. Furthermore, Caught in the Act: Probes Effectively Detect Sabotage and Catch Unverbalized Deception and SAFETY SENTRY: Context-Aware Human Intervention via EXECUTE-ASK-REFUSE Routing provide mechanisms to catch subtle, non-obvious deception.
• Governance and Humility: Accurate but Not Humble: Evaluating Epistemic Humility in LLM Agents under Knowledge Conflict argues that safe agents must communicate uncertainty. This is supported by frameworks like TRACE: A Governance Framework for Measuring Explainability Debt in Production AI Systems and From Reactive Containment to Proactive Assurance: Lessons from OpenAI, Anthropic, and Google Agent Security Incidents, which provide lifecycle-oriented security standards.

Theme 3: Efficient Reasoning and Memory Management

To operate in complex, long-horizon environments, agents must manage their “cognitive load.” This involves optimizing memory, compute budgets, and the decision-making process itself.

• Context and Memory Management: Agents are learning to curate their own memory to save costs and improve focus. Agent-Controlled Forgetting for Tool-Using Agents: Reversible Context Curation in Practice and Curating Always-Loaded Context for LLM Agents: A Capacitated Assortment Model with Censored Feedback address active forgetting, while RaReCache: Bridging the Gap in Cross-Model KV Cache Reuse via Rank disagreement-based Selective Recomputation and MemoWM: How World Models Change What Agents Need to Remember focus on efficient memory reuse via world models.
• Adaptive Reasoning and Scaling: We are moving away from the assumption that agents must “think” at every step. When Should Agents Think? Adaptive Reasoning via Cross-Turn Estimation and From Chain-of-Thought to Loops: Non-Autoregressive Latent Reasoning via Looped Transformers propose mechanisms for selective reasoning. These are complemented by Efficient Reasoning via Constrained Optimization in Latent Space and LoRi: Low-Rank Distillation for Implicit Reasoning, which compress reasoning chains into compact representations.
• Efficient Architectures: MoEMB: Scaling Universal Multimodal Embeddings with Efficient Mixture-of-Experts Models and S4oP: Operator-level Pruning of Structured State Space Models for Resource-Constrained Devices demonstrate how to scale capacity while maintaining performance on edge devices.

Theme 4: Embodied Intelligence and Scientific Discovery

The final frontier is the application of agentic principles to the physical world and scientific inquiry, where agents must understand causal dynamics rather than just processing language.

• World Modeling and Robotics: Multi-Agent Egocentric World Model with Fine-Grained Embodied Interaction and LeWAM: A JEPA World Action Model with Diffusion-Steering-Based MPC focus on causal dynamics. For robotics, ARC: A Reasoning Recipe for Robot Foundation Models, ExecVLA: Following Fine-Grained Execution Constraints in Vision-Language-Action Models with Bi-Level Action Representation, NavGPT-3: Harnessing Context in a Hierarchical Navigation Runtime, and REACT: Rolling Denoising and Dual Decoupling for Reactive Robot Control with VLA Models bridge the gap between high-level reasoning and low-level physical control.
• Digital and Scientific Frontiers: MobileWorldBench: Towards Semantic World Modeling For Mobile Agents and LiteGUI: Lightweight GUI Agents via Multi-Solution Guided Distillation and Dual-Level Reinforcement Learning apply these principles to digital interfaces. In science, LLM-IDEA: Identifiability-Driven Experimental Agent for Autonomous Discovery of Mechanistic World Models and EvoSim: Learning to Model, Modeling to Learn showcase agents that autonomously design experiments and refine mechanistic models.
• Data Efficiency: Data-Free On-Policy Distillation: How Far Can We Go Without External Data? challenges the need for massive external datasets, showing that agents can generate their own high-quality training data through self-play.