Action-Driven and Self-Evolving Agentic Reasoning скачать в хорошем качестве

Action-Driven and Self-Evolving Agentic Reasoning

Reasoning is the core cognitive process that underpins inference, problem solving, and decision-making. Although large language models (LLMs) have demonstrated strong reasoning abilities in closed environments—such as standardized benchmark tasks in mathematics and code—they still face significant challenges in open and dynamic environments. The emergence of *agentic reasoning* marks a paradigm shift. By reframing LLMs as autonomous agents capable of planning, acting, and learning through continuous interaction, it bridges the gap between “thinking” and “acting.” In this survey, we construct a systematic research roadmap through three complementary dimensions. First, from the perspective of environmental dynamics, we categorize agentic reasoning into three levels. *Foundational agentic reasoning* focuses on the core capabilities of a single agent, including planning, tool use, and search, which typically operate in relatively stable environments. *Self-evolving agentic reasoning* investigates how agents continuously improve their capabilities in changing environments through feedback, memory, and adaptive mechanisms. *Multi-agent collaborative reasoning* further extends this paradigm to cooperative settings, where multiple agents coordinate through role specialization, knowledge sharing, and joint problem solving to achieve shared goals. Across all levels, we also analyze reasoning from the perspective of system constraints and optimization strategies, distinguishing between two paradigms. The first is **in-context reasoning**, which expands test-time interaction capabilities through structured orchestration and adaptive workflow design. The second is **post-training reasoning**, which optimizes model behavior through approaches such as reinforcement learning and supervised fine-tuning. In addition, we systematically review and analyze the application of agentic reasoning frameworks across a variety of real-world tasks and benchmarks, including scientific research, robotics, healthcare, automated scientific discovery, and mathematical problem solving. These examples illustrate how different reasoning mechanisms are implemented and evaluated across diverse domains. This survey integrates diverse approaches to agentic reasoning into a unified research roadmap aimed at connecting “thinking” with “acting,” and provides actionable guidance for building agent systems under varying levels of environmental dynamics, optimization strategies, and agent interaction settings. Finally, we discuss key open challenges and future research directions, including personalization, long-horizon interaction, world models, scalable multi-agent training, and governance frameworks for real-world deployment.