Title: Adaptive Latent Agentic Reasoning

Source: http://arxiv.org/abs/2606.02871v1

Summary:

This paper introduces a dual-mode reasoning framework that dynamically alternates between compact latent reasoning and explicit chain-of-thought, optimizing the accuracy-efficiency trade-off for multi-turn agents. It establishes a significant architectural primitive by enabling agents to reserve heavy deliberation for complex decisions while maintaining high efficiency for routine tasks.

Title: SkillSmith: Co-Evolving Skills and Tools for Self-Improving Agent Systems

Source: http://arxiv.org/abs/2606.01314v1

Summary:

SkillSmith presents a foundational co-evolution framework that allows agents to simultaneously evolve their skill libraries and underlying toolsets through a synergy-aware reflection process. By utilizing an ecological utility model to manage skill-tool interactions, it establishes a novel architectural loop for autonomous, self-improving agent systems capable of repairing their own functional primitives.

Title: Science Earth: Towards A Planet-Scale Operating System for AI-Native Scientific Discovery

Source: http://arxiv.org/abs/2606.01316v1

Summary:

It introduces a planet-scale scientific runtime and the EACN protocol, enabling decentralized AI capabilities to autonomously discover and negotiate task ownership. This shifts agentic orchestration from pre-designed workflows to emergent, self-correcting coordination for large-scale discovery.

Title: RLVR without Ineffective Samples: Group Prioritized Off-Policy Optimization for LLM Reasoning

Source: http://arxiv.org/abs/2606.01281v1

Summary:

This paper introduces POPO, a novel optimization framework that solves the critical zero-variance reward bottleneck in Reinforcement Learning with Verifiable Rewards (RLVR) for LLM reasoning. By implementing prioritized group replay and decoupled off-policy optimization, it provides a foundational efficiency breakthrough for training reasoning-intensive models with significantly reduced rollout overhead.

Title: Learning to Adapt: Self-Improving Web Agent via Cognitive-Aware Exploration

Source: http://arxiv.org/abs/2605.31365v1

Summary:

SCALE introduces a foundational self-improving framework that enables agents to autonomously expand their cognitive boundaries through adversarial exploration and global planning strategies. It marks a significant shift from static, handcrafted execution pipelines to truly adaptive agentic systems that learn and generalize from their own environmental interactions.

Title: EchoRL: Reinforcement Learning via Rollout Echoing

Source: http://arxiv.org/abs/2605.31228v1

Summary:

This paper introduces EchoRL, a novel reinforcement learning primitive that prevents training signal collapse in reasoning models by recovering gradients from successfully verified rollouts. It establishes a foundational method for post-training LLMs to achieve higher reasoning performance without encountering the typical diminishing returns of standard RLVR methods.

Title: GrepSeek: Training Search Agents for Direct Corpus InteractionSource: http://arxiv.org/abs/2605.29307v1

Summary:

This paper introduces Direct Corpus Interaction (DCI), a foundational paradigm shift where search agents treat text corpora as executable environments via shell commands instead of traditional ranked indices. By training agents to find and compose evidence directly from raw data using a two-stage RL pipeline, it establishes a new architectural framework for knowledge-intensive agentic reasoning.

Title: ESPO: Early-Stopping Proximal Policy Optimization

Source: http://arxiv.org/abs/2605.29860v1

Summary:

Early-Stopping Proximal Policy Optimization (ESPO) provides a significant breakthrough in efficiency and reasoning for LLM reinforcement learning by detecting and terminating failed reasoning trajectories on-the-fly. This foundational optimization reduces compute overhead by 20% while improving performance on complex math and reasoning benchmarks by concentrating negative reward signals at the exact point of logical failure.

Title: TRACER: Turn-level Regret Matching with Inner Reinforcement Credit for Cooperative Multi-LLM Reasoning

Source: http://arxiv.org/abs/2605.28699v1

Summary:

TRACER introduces a novel turn-level reinforcement framework that unifies regret matching with role-specific rewards to optimize multi-agent cooperation and reasoning. By separating the decision of when to speak from the content of the utterance, it establishes a mathematically rigorous foundation for evolving complex collaborative protocols in multi-LLM systems.

Title: Do Agents Think Deeper? A Mechanistic Investigation of Layer-Wise Dynamics in Sequential Planning

Source: http://arxiv.org/abs/2605.27935v1

Summary:

This study provides foundational mechanistic evidence that agentic reasoning requires dynamic, adaptive recruitment of model depth, distinguishing it from static inference tasks. These insights into layer-wise dynamics are critical for developing the next generation of LLM architectures optimized for long-horizon planning and iterative tool use.