Title: RL Post-Training Builds Compositional Reasoning Strategies

Source: http://arxiv.org/abs/2607.07646v1

Summary:

This paper provides foundational insights into how reinforcement learning post-training enables models to transition from simple skills to complex, multi-step compositional reasoning strategies. By demonstrating how RL systematically organizes and compresses primitive actions into stable, higher-level reduction procedures, it advances our theoretical understanding of reasoning emergence in advanced generative models.

Title: From Atomic Actions to Standard Operating Procedures: Iterative Tool Optimization for Self-Evolving LLM Agents

Source: http://arxiv.org/abs/2607.07321v1

Summary:

This work introduces a novel framework for self-evolving agents by enabling them to autonomously synthesize low-level, atomic actions into reusable, higher-order Standard Operating Procedures (SOPs). This dynamic tool-optimization loop represents a major paradigm shift in reducing reasoning overhead and improving long-horizon task execution for autonomous systems.

Title: Information Gain-based Rollout Policy Optimization: An Adaptive Tree-Structured Rollout Approach for Multi-Turn LLM Agents

Source: http://arxiv.org/abs/2607.06223v1

Summary:

This paper introduces a novel policy optimization framework (IGRPO) that dynamically allocates rollout budgets based on node-level information gain during tree-structured exploration. By unifying adaptive search-tree exploration with a principled reinforcement learning target, it provides a foundational methodology for scaling and training multi-turn reasoning agents.

Title: Doomed from the Start: Early Abort of LLM Agent Episodes via a Recall-Controlled Probe Cascade

Source: http://arxiv.org/abs/2607.06503v1

Summary:

This paper introduces a novel framework that uses internal activation probes to detect and early-abort doomed agent trajectories, saving up to 47% of inference compute. This represents a significant efficiency breakthrough for agentic loops, addressing the critical challenge of compute wastage in multi-step agent environments.

Title: Weak-to-Strong Generalization via Direct On-Policy Distillation

Source: http://arxiv.org/abs/2607.05394v1

Summary:

This work introduces a novel post-training paradigm that transfers the reinforcement learning policy shift of a smaller, cheaper weak model as an implicit reward signal to a stronger target model. By bypassing the need for explicit reward modeling or expensive on-policy rollouts on scaled architectures, it represents a major efficiency breakthrough for training high-capability reasoning models.

Title: Multi-Turn On-Policy Distillation with Prefix Replay

Source: http://arxiv.org/abs/2607.04763v1

Summary:

This paper addresses a critical training bottleneck for LLM agents by introducing an off-environment distillation method that resolves the 'prefix trap' in multi-turn interactions. By enabling scalable on-policy distillation with a 4x training speedup and zero tool calls, it establishes a highly efficient framework for training LLM agents across diverse environments.

Title: LLM-as-a-Verifier: A General-Purpose Verification Framework

Source: http://arxiv.org/abs/2607.05391v1

Summary:

This paper formalizes solution verification as a major new scaling axis for language models, introducing a framework that computes continuous scores over logit distributions rather than discrete judgments to evaluate complex reasoning. It establishes a highly versatile, training-free mechanism that significantly boosts performance on agentic benchmarks while offering a scalable source of dense feedback for reinforcement learning.

Title: Biological Motifs for Agentic Control

Source: http://arxiv.org/abs/2607.04240v1

Summary:

This work establishes a formal mathematical framework that models multi-agent systems using control motifs from systems biology and category theory. By introducing the Agentic Operad, it provides a typed syntax for agent composition with provable error-suppression bounds and scaling laws for complex agentic systems.

Title: Auto: The AGI Compiler

Source: http://arxiv.org/abs/2607.04542v1

Summary:

This paper introduces Auto, a novel compiler framework that records live AI agent behavior and compiles deterministic sequences into verified WebAssembly binaries. This approach dramatically reduces the inference cost and latency of complex agentic workflows while maintaining reliability through calibrated fallback guards.

Title: Agent Reinforcement Learning via Pivotal-Aware Self-Feedback Retry

Source: http://arxiv.org/abs/2607.03702v1

Summary:

This paper introduces PivoARL, a novel agentic reinforcement learning framework that optimizes agent trajectories by identifying and retrying only from pivotal erroneous states. By isolating correct prefixes and addressing credit assignment near error boundaries, it significantly reduces interaction costs and improves agent reasoning efficiency.