Title: How Should LLMs Consume High-Quality Data? Optimal Data Scheduling via Quality-Aware Functional Scaling Laws

Source: http://arxiv.org/abs/2605.25698v1

Summary:

This paper establishes foundational quality-aware functional scaling laws that provide the first theoretical closed-form solution for scheduling high-quality data during LLM training. The introduced 'Drop-Stable-Rampup' schedule optimizes training dynamics across noise-limited and signal-limited regimes, yielding significant breakthroughs in mathematical reasoning performance.

Title: Efficient Agentic Reasoning Through Self-Regulated Simulative Planning

Source: http://arxiv.org/abs/2605.22138v1

Summary:

This paper introduces a foundational three-system reasoning framework—comprising reactive, simulative, and self-regulated components—that enables agents to autonomously manage their planning depth and horizon. By treating the LLM as a world model for future-state prediction, it demonstrates that structured deliberation can allow smaller models to match the performance of systems orders of magnitude larger.

Title: Compiling Agentic Workflows into LLM Weights: Near-Frontier Quality at Two Orders of Magnitude Less Cost

Source: http://arxiv.org/abs/2605.22502v1

Summary:

This work proposes the 'subterranean agent' paradigm, which replaces external orchestration frameworks by compiling agentic workflows directly into the model's weights via fine-tuning. This foundational shift addresses the cost and latency bottlenecks of frontier-model prompting while providing a more efficient and private alternative for procedural task execution.

Title: Mem-π: Adaptive Memory through Learning When and What to Generate

Source: http://arxiv.org/abs/2605.21463v1

Summary:

Mem-π presents a foundational shift in agent memory architectures by replacing static similarity-based retrieval with a dedicated generative model that produces context-specific guidance. This framework enables agents to dynamically adapt their memory usage, leading to substantial improvements in complex reasoning and long-horizon task execution.

Title: Agent JIT Compilation for Latency-Optimizing Web Agent Planning and Scheduling

Source: http://arxiv.org/abs/2605.21470v1

Summary:

This paper introduces Agent Just-In-Time (JIT) compilation, a novel architectural primitive that transforms natural language task descriptions into optimized, executable code plans. It represents a significant breakthrough in agentic efficiency by replacing traditional sequential loops with a compiled, parallelized execution framework that drastically reduces latency.

Title: PEEK: Context Map as an Orientation Cache for Long-Context LLM Agents

Source: http://arxiv.org/abs/2605.19932v1

Summary:

This work introduces 'context maps' as a novel architectural primitive for long-context agents, enabling them to cache and maintain structured orientation knowledge about recurring external datasets. By implementing a programmable cache policy for distilling and translating inference-time signals, it significantly improves efficiency and accuracy across multi-turn reasoning workloads.

Title: Formal Skill: Programmable Runtime Skills for Efficient and Accurate LLM Agents

Source: http://arxiv.org/abs/2605.19604v1

Summary:

This work introduces a foundational architectural primitive for agents that replaces informal natural-language instructions with programmable, stateful runtime skills governed by hook policies and action schemas. This shift from prompting to executable state machines provides a more enforceable and token-efficient control surface for reliable agentic workflows in real-world environments.

Title: Formal Skill: Programmable Runtime Skills for Efficient and Accurate LLM Agents

Source: http://arxiv.org/abs/2605.19604v1

Summary:This work introduces a foundational architectural primitive for agents that replaces informal natural-language instructions with programmable, stateful runtime skills governed by hook policies and action schemas. This shift from prompting to executable state machines provides a more enforceable and token-efficient control surface for reliable agentic workflows in real-world environments.

Title: A Measure-Theoretic Analysis of Reasoning: Structural Generalization and Approximation Limits

Source: http://arxiv.org/abs/2605.19944v1

Summary:

This paper establishes fundamental theoretical bounds for LLM reasoning, proving that scaling physical layer depth is a non-negotiable requirement for out-of-distribution generalization that cannot be bypassed by scaling width. It also formalizes why specific architectural choices, such as shift-invariant embeddings, are mathematically necessary to maintain reasoning equivariance across domain shifts.

Title: Learning to Hand Off: Provably Convergent Workflow Learning under Interface Constraints

Source: http://arxiv.org/abs/2605.19140v1Summary:

This research provides the first finite-sample guarantee for neural Q-learning in decentralized multi-agent settings, a foundational breakthrough for reliable agentic workflow learning. By formalizing handoffs as interface-constrained SMDPs, it enables provably convergent learning in complex LLM pipelines where agents have restricted observability.