PseudoAct: Leveraging Pseudocode Synthesis for Flexible Planning and Action Control in Large Language Model Agents

Researchers Yihan Wen and Xin Chen have introduced PseudoAct, a groundbreaking framework that fundamentally shifts how large language model (LLM) agents approach complex, multi-step tasks. The paper, published on arXiv, addresses critical limitations in current reactive agent paradigms like ReAct, which often lead to unstable reasoning, redundant tool usage, and high computational costs in long-horizon scenarios involving branching or iteration. PseudoAct's core innovation is its two-stage process: first, it prompts an LLM to synthesize a structured pseudocode plan that decomposes a task into subtasks and explicitly encodes control flow logic (sequencing, conditionals, loops). Second, the agent executes actions by following this coherent global plan, making its decision logic transparent and temporally stable.

This shift from reactive, step-by-step decision-making to proactive, plan-based execution yields significant performance gains. By preventing infinite loops and avoiding uninformative exploration of alternatives, PseudoAct reduces token consumption and improves consistency. Experimental results on established benchmarks are compelling: the framework achieved a 20.93% absolute improvement in success rate on the FEVER fact-verification dataset and set a new state-of-the-art on the challenging HotpotQA question-answering benchmark. This demonstrates that explicit pseudocode synthesis can overcome the brittleness of current agent designs, paving the way for more reliable and efficient AI systems capable of handling real-world, multi-tool workflows in areas like research, customer service, and data analysis.