NEW: Automating Agentic LLM Workflows with AFLOW.
AFLOW presents a novel framework that automates the optimization of agentic workflows for large language models (LLMs) by reframing workflow construction as a search problem over code-represented workflows. The key insight is to represent workflows as graphs where nodes correspond to LLM invocations with parameters such as the model, prompt, temperature, and output format, and edges represent the logical flow and dependencies between these nodes. By using code to represent these workflows, AFLOW can model complex interactions, including conditional logic and iterative processes, thus enabling the exploration of an infinite and highly expressive search space that generalizes prior methods constrained by manual design.
To efficiently navigate this expansive search space, AFLOW employs a specialized Monte Carlo Tree Search (MCTS) algorithm augmented with several innovations. First, it introduces a soft mixed-probability selection mechanism that balances exploration and exploitation by combining uniform probability distribution with score-based weighted selection, preventing premature convergence to suboptimal workflows. Second, LLM-driven node expansion leverages the generative capabilities of LLMs to propose new workflow modifications directly in code, effectively using the model to optimize itself. Third, execution evaluation runs candidate workflows on validation datasets using task-specific evaluation functions, providing empirical performance feedback. This is followed by experience backpropagation, where the performance metrics are propagated up the search tree to inform future decisions, refining the search process iteratively.
A significant advancement in AFLOW is the introduction of operators—predefined, reusable combinations of nodes that encapsulate common agentic operations such as Generate, Review & Revise, Ensemble, and Test. These operators serve as higher-level abstractions that reduce the complexity of the search by structuring it around effective patterns of LLM usage. Integrating these operators into the MCTS framework allows AFLOW to construct sophisticated workflows more efficiently, ensuring that the exploration leverages known successful strategies while still discovering novel configurations. Empirical evaluations across six benchmark datasets demonstrate that AFLOW outperforms state-of-the-art baselines by an average of 5.7%, significantly improving automated workflow generation's efficacy and enabling smaller models to achieve performance comparable to larger models like GPT-4 at a fraction of the computational cost.
00:00 What are Agentic Workflows?
01:48 Autonomous AI Agents vs Agentic Workflows
02:50 3 types of Agentic Workflows
04:15 Automated Workflow optimization
05:20 AFLOW Automating Agentic Workflow
06:40 Workflow representation of Nodes and Edges
10:30 New Operators for Agentic Workflow
11:37 Main task
13:42 Monte Carlo Tree Search for AFLOW
15:47 AFLOW Algorithm explained
18:20 Reduce Human time, resources and costs
All rights w/ authors:
AFLOW:
AUTOMATING AGENTIC WORKFLOW GENERATION
[ Ссылка ]
#agent
#airesearch
#programming #ai
#performance
#newtechnology
