Title: Agentic Routing: The Harness-Native Data Flywheel

Source: http://arxiv.org/abs/2607.11399v1

Summary:

This paper introduces a foundational step-level routing paradigm that treats multi-model coordination within agent execution harnesses as a dynamic, state-conditioned systems problem. It establishes a harness-native data flywheel where execution traces and environmental feedback automatically train better routers and specialized models, offering a self-improving data engine for scaling agent infrastructure.

Title: A Formal Hierarchical Architecture for Agentic Orchestration with Stack-Based Execution and Lazy Discovery

Source: http://arxiv.org/abs/2607.11138v1

Summary:

This paper presents a novel hierarchical, stack-based execution architecture for agentic orchestration that models complex, nested agent states as Pushdown Automata. By substituting flat tool registries with localized stack frames and a manifest-driven lazy-loading protocol, it provides an architectural solution to context-window saturation and decision-space explosion in agent reasoning loops.

Title: LOGOS: A Living Logic for AI Agent Teams That Evolve With HumansSource: http://arxiv.org/abs/2607.10878v1

Summary:

LOGOS introduces a pluggable layer for self-evolution and governance within multi-agent frameworks, enabling verifiable human-agent loop engineering. This framework is foundational for controlling, auditing, and safely evolving AI agent teams at scale, addressing critical questions of accountability and trustworthy automation in their long-term deployment.

Title: Agentic-DPO: From Imitation to Agentic Policy Optimization on Expert Trajectories

Source: http://arxiv.org/abs/2607.10601v1

Summary:

Agentic-DPO proposes a lightweight offline policy optimization method that transforms expert trajectories into state-conditioned preference supervision for DPO-style training. This approach represents a significant breakthrough in efficiently training robust LLM agents to make better decisions, moving beyond mere imitation by directly optimizing their underlying policies.

Title: Beyond Euclidean Clipping: Overcoming Exploration Collapse in LLM RL via Riemannian Isometric Policy Optimization

Source: http://arxiv.org/abs/2607.10169v1

Summary:

This paper introduces Riemannian Isometric Policy Optimization (RIPO), a foundational mathematical advancement that corrects the geometric mismatch in standard PPO-Clip's Euclidean-based policy distance measurements. By ensuring isometric policy updates on the Riemannian manifold, RIPO effectively prevents exploration collapse and significantly boosts LLM reasoning capabilities on competitive mathematical benchmarks.

Title: ARMOR: Stabilizing On-Policy LLM RL with Off-Policy Anchor Samples

Source: http://arxiv.org/abs/2607.10481v1

Summary:

This paper presents ARMOR, a novel post-training reinforcement learning framework that addresses the persistent issue of training instability and over-optimization in reasoning LLMs. By introducing active anchor rollouts from reference policies paired with a mixed optimization objective, it offers a highly effective alternative to standard KL regularization for stabilizing complex reasoning models.

Title: Shared Selective Persistent Memory for Agentic LLM Systems

Source: http://arxiv.org/abs/2607.09493v1

Summary:

This work introduces a foundational shared selective persistent memory architecture that solves the critical context-bloat and token-inefficiency bottlenecks in multi-session LLM agents. By isolating reusable context parameters and decoupling runtime data, the framework achieves dramatic token cost reductions and performance gains in collaborative agent environments.

Title: A Sovereign, Open-Source Foundation Model for German and English

Source: http://arxiv.org/abs/2607.09424v1

Summary:

This paper presents a novel hybrid MoE Mamba-Transformer architecture at scale, representing a significant architectural milestone for foundation model design. The hybrid model maintains a near-constant inference cache during context growth, demonstrating a major throughput and efficiency breakthrough essential for long-context generative AI.

Title: Compete Then Collaborate: Frontier AI Teachers Build a Verifiable Curriculum to Improve a Coding Student Beyond Imitation

Source: http://arxiv.org/abs/2607.08255v1

Summary:

This work delivers a foundational breakthrough in LLM post-training by demonstrating that traditional supervised fine-tuning on teacher-generated data can degrade capable student models, whereas reinforcement learning with verifiable rewards (RLVR) using a collaborative teacher-built curriculum yields significant reasoning gains. It establishes a robust, reproducible paradigm for training frontier models through verification and active learning rather than simple imitation.

Title: Cognitive-structured Multimodal Agent for Multimodal Understanding, Generation, and Editing

Source: http://arxiv.org/abs/2607.08497v1

Summary:

This paper introduces a novel architectural framework for multimodal agents that replaces monolithic context-window scaling with an Episodic Visual Memory and selective reactivation system to prevent visual token explosion. By decoupling perception, cognitive retrieval, and execution, it establishes a highly efficient and scalable paradigm for long-horizon multimodal agentic reasoning and action.