DARS: Dynamic Action Re-Sampling to Enhance Coding Agent Performance by Adaptive Tree Traversal

2025-03-18

AI Summary

Current LLM-based coding agents frequently converge to suboptimal decisions in complex software tasks, necessitating either manual intervention or computationally inefficient scaling. To address this, we propose a dynamic action resampling mechanism that adaptively explores alternative actions at critical decision points, leveraging historical trajectories and execution feedback. We further introduce the first inference-time adaptive tree traversal paradigm for computational scaling—enabling robust error recovery without additional training or human involvement. Our approach integrates Claude 3.5 Sonnet V2-based trajectory modeling, feedback-driven resampling, and dynamic tree construction with pruning. Evaluated on SWE-Bench Lite, our method achieves pass@1 = 47% and pass@k = 55%, substantially outperforming existing open-source state-of-the-art methods.

Technology Category

• Search and Optimization: Sampling/Simulation-based Search
• Planning, Routing, and Scheduling: Planning with Language Models
• Natural Language Processing: Code Generation / Program Synthesis from Natural Language

Application Category

• Search and Retrieval-Augmented AI: Agentic search
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems

Abstract

Large Language Models (LLMs) have revolutionized various domains, including natural language processing, data analysis, and software development, by enabling automation. In software engineering, LLM-powered coding agents have garnered significant attention due to their potential to automate complex development tasks, assist in debugging, and enhance productivity. However, existing approaches often struggle with sub-optimal decision-making, requiring either extensive manual intervention or inefficient compute scaling strategies. To improve coding agent performance, we present Dynamic Action Re-Sampling (DARS), a novel inference time compute scaling approach for coding agents, that is faster and more effective at recovering from sub-optimal decisions compared to baselines. While traditional agents either follow linear trajectories or rely on random sampling for scaling compute, our approach DARS works by branching out a trajectory at certain key decision points by taking an alternative action given the history of the trajectory and execution feedback of the previous attempt from that point. We evaluate our approach on SWE-Bench Lite benchmark, demonstrating that this scaling strategy achieves a pass@k score of 55% with Claude 3.5 Sonnet V2. Our framework achieves a pass@1 rate of 47%, outperforming state-of-the-art (SOTA) open-source frameworks.

Problem

• Improves coding agent decision-making efficiency
• Enhances recovery from sub-optimal decisions
• Outperforms state-of-the-art frameworks in benchmarks

Innovation

• Dynamic Action Re-Sampling (DARS) enhances coding agents
• DARS uses adaptive tree traversal for decision-making
• Achieves 55% pass@k score on SWE-Bench Lite