Kimi-Dev: Agentless Training as Skill Prior for SWE-Agents
🤖 AI Summary
This work addresses the limited adaptability and low efficiency of SWE-Agents in software engineering tasks. We propose leveraging the Agentless training paradigm as a skill prior construction mechanism: a single-round supervised fine-tuning (SFT) is performed on 5K publicly available code trajectories to model structured capabilities—including code localization, editing, and self-reflection—yielding transferable skill representations. To our knowledge, this is the first approach to bridge the gap between agent-free training and multi-turn agent frameworks, significantly enhancing both generalization and reasoning efficiency of SWE-Agents. On the SWE-bench Verified benchmark, the base model Kimi-Dev achieves 60.4% accuracy, enabling the resulting SWE-Agent to attain 48.6% pass@1—comparable to the closed-source Claude 3.5 Sonnet. Our method establishes a reusable, open-weight paradigm for skill transfer in code intelligence agents.
Application Category
📝 Abstract
Large Language Models (LLMs) are increasingly applied to software engineering (SWE), with SWE-bench as a key benchmark. Solutions are split into SWE-Agent frameworks with multi-turn interactions and workflow-based Agentless methods with single-turn verifiable steps. We argue these paradigms are not mutually exclusive: reasoning-intensive Agentless training induces skill priors, including localization, code edit, and self-reflection that enable efficient and effective SWE-Agent adaptation. In this work, we first curate the Agentless training recipe and present Kimi-Dev, an open-source SWE LLM achieving 60.4% on SWE-bench Verified, the best among workflow approaches. With additional SFT adaptation on 5k publicly-available trajectories, Kimi-Dev powers SWE-Agents to 48.6% pass@1, on par with that of Claude 3.5 Sonnet (241022 version). These results show that structured skill priors from Agentless training can bridge workflow and agentic frameworks for transferable coding agents.
Problem
Research questions and friction points this paper is trying to address.
Bridging agentless training with SWE-agent frameworks
Developing transferable skill priors for coding agents
Enhancing software engineering LLM performance on benchmarks
Innovation
Methods, ideas, or system contributions that make the work stand out.
Agentless training creates skill priors for agents
Kimi-Dev achieves state-of-the-art workflow performance
Skill transfer bridges workflow and agentic frameworks
🔎 Similar Papers
No similar papers found.
