In-the-Flow Agentic System Optimization for Effective Planning and Tool Use (Oct 2025) скачать в хорошем качестве

In-the-Flow Agentic System Optimization for Effective Planning and Tool Use (Oct 2025)

Title: In-the-Flow Agentic System Optimization for Effective Planning and Tool Use (Oct 2025) Link: http://arxiv.org/abs/2510.05592v1 Date: October 2025 Summary: This paper introduces AGENTFLOW, a trainable agentic framework for coordinating specialized modules (planner, executor, verifier, generator) through an evolving memory, and optimizes the planner inside the multi-turn loop. To train on-policy in live environments, the paper proposes Flow-based Group Refined Policy Optimization (Flow-GRPO), converting multi-turn optimization into tractable single-turn policy updates. Results show significant performance gains over baselines across ten benchmarks. Key Topics: Agentic systems Reinforcement Learning Tool-augmented reasoning Policy Optimization Long-horizon training Multi-turn interaction Large Language Models (LLMs) Flow-GRPO Chapters: 00:00 - Intro to Agent Flow 00:15 - Core Problem 00:37 - Specialized Parts 00:53 - FlowBase Group Refined Policy Optimizer 01:16 - Headline Result 01:38 - Agents Falling Over 02:17 - Instability 02:36 - Agentic Systems 02:55 - Adaptability 03:19 - Errors 03:42 - Four Modules 04:06 - Action Planner 04:23 - Tool Executor 04:48 - Execution Verifier 05:19 - Solution Generator 05:33 - Evolving Memory 06:11 - Relevant Pieces 06:23 - Reinforcement Learning Challenge 06:47 - FlowGRPO 07:14 - Final Outcome 07:40 - Every Step 08:10 - Group Normalization 08:43 - On Policy 09:06 - Imitation 09:28 - Results 09:41 - Efficiency 10:11 - Task Gains 10:30 - Complex Agent Tasks 10:53 - Significant Gains 11:16 - Architecture 11:35 - Scaling Task Complexity 11:52 - Used Them Effectively 12:17 - Peek Inside 12:25 - Cool Analysis 12:37 - Tool Calling Reliability 13:06 - Adaptive Tool Selection 13:32 - Medical Question 13:58 - Self Correction 14:17 - Failure Case 14:51 - Real Resilience 15:25 - Core Value Proposition 15:37 - Key Takeaways 16:25 - Modular 16:43 - Provocative Thought 17:23 - Interesting Question