Agentic AI

This research paper addresses inefficiencies in multi-turn reinforcement learning for LLM agents, where uniform rollouts waste budget on dead ends. The authors propose a Process-Scorer Guided Adaptive Tree Rollout method that structures trajectories as trees, using intermediate rewards to guide exploration. This approach prioritizes promising intermediate states over complete trajectory sampling, improving sample efficiency in long-horizon tasks.

• Replaces uniform trajectory sampling with tree-based exploration for better budget allocation.
• Uses process rewards to identify and branch from promising intermediate agent states.
• Reduces wasted computation on uninformative dead-end attempts in long-horizon agentic tasks.
• Reframes RL exploration as a decision process at each turn rather than at episode end.

Autonomous data science agents currently suffer from expensive trial-and-error loops. DSWorld introduces a Data Science World Model that predicts environment state transitions before executing operations. The framework combines structured state construction, cost-aware routing, and an LLM-based simulator to anticipate outcomes efficiently.

• Predicts operation effects before execution to reduce costly trial-and-error.
• Uses a Data Science World Model to simulate environment state transitions.
• Integrates cost-aware routing and lightweight real execution for efficiency.
• Leverages LLM-based simulation for handling expensive computational operations.

The SkillCorpus framework addresses the fragmentation and redundancy in the open-source SKILL.md ecosystem by aggregating approximately 821,000 crawled skills. It applies a multi-stage filtering pipeline to distill these into a high-quality corpus of 96,401 skills organized by a 16-class taxonomy. The system also evaluates the utility of these skills on real-world agent tasks to establish performance bounds.

• Filters 821k raw SKILL.md artifacts down to 96k curated, high-quality skills
• Organizes skills using a standardized 16-class taxonomy for better discoverability
• Provides empirical evaluation of skill value on real-world LLM agent tasks
• Addresses critical quality and redundancy issues in the open skill ecosystem

This study isolates the contributions of executable world modeling, scheduled simplification, and exact replay verification in coding agents solving ARC-AGI-3. Using four nested Codex-based agents evaluated with GPT-5.4, GPT-5.5, and GPT-5.6-sol, the authors compare a textual baseline against variants with flexible or fixed interface controls. The goal is to determine which specific architectural components drive the agent's problem-solving success.

• Uses ablation studies to attribute performance gains in ARC-AGI-3 solving agents.
• Compares textual baselines against executable world models with and without verification.
• Evaluates four agent variants using GPT-5.4, GPT-5.5, and GPT-5.6-sol models.
• Tests impact of scheduled simplification and exact replay verification on reasoning effort.
• Provides clear evidence on whether world modeling or verification is critical for success.

ToolVerse addresses the fragility of LLM agents in complex, dynamic environments by introducing a framework that scales agentic reinforcement learning. The system automatically constructs massive training environments by integrating approximately 4,500 tools from nearly 400 real-world Model Context Protocols (MCPs). It employs a task design strategy based on tool dependency graphs to enable robust long-horizon reasoning in Tool-Integrated Reasoning tasks.

• Automates environment creation using ~400 MCPs with ~4,500 tools for agentic RL training.
• Enables agents to handle long-horizon tasks in large-scale, diverse real-world scenarios.
• Uses tool dependency graphs to structure tasks and improve reasoning robustness.
• Shifts focus from compact, well-defined scenarios to complex, dynamic tool integration.

jcode is a new coding agent harness designed to improve developer productivity through multi-session workflows and extensive customization. The tool emphasizes extreme performance and resource efficiency, claiming optimized RAM usage and boot times compared to alternatives. It supports installation on macOS, Linux, and Windows via script or package managers.

• Optimized for low RAM and fast boot times to support scaling multi-session workflows.
• Designed for infinite customizability to raise the skill ceiling for coding agents.
• Supports automated setup via provider integration or manual configuration.
• Cross-platform support includes macOS, Linux, and Windows 11+.