Research map / Multi-agent systems and games / LLM agents

Kimi-Dev: Agentless Training as Skill Prior for SWE-agents

ICLR 2026 · Multi-agent systems and games › LLM agents · approach: Robust and efficient agent execution · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces Kimi-Dev, a 72B SWE model trained via an Agentless recipe (mid-training, cold-start SFT, outcome-reward RL, test-time self-play) that reaches 60.4% on SWE-bench Verified and, after light SFT on 5k agent trajectories, 48.6% pass@1 as a SWE-Agent.

Problem

Agentless workflow methods and multi-turn SWE-Agent frameworks are viewed as mutually exclusive, and SWE-Agents are hard to train end-to-end due to sparse long-horizon rewards and sensitivity to initialization on SWE-bench-style tasks.

Approach

Treat Agentless training (localization, code edit, self-reflection skills learned via mid-training, cold-start SFT, and RLVR) as inducing skill priors that enable efficient SWE-Agent adaptation with few agentic trajectories; plus a BugFixer/TestWriter duo with execution-based test-time self-play for patch selection.

Key results

• Kimi-Dev (72B) reaches 60.4% resolve rate on SWE-bench Verified under Agentless, the best among workflow-based open-source approaches.
• After SWE-Agent SFT on 5,016 trajectories, Kimi-Dev reaches 48.6% pass@1 on SWE-bench Verified, comparable to Claude 3.5 Sonnet (241022) at 49.0%.
• Test-time self-play scales from 48.0% (1x1) to 60.4% (40x40 patch-test pairs).

Surprising findings

• The RL-trained prior needs only 2^23 SWE-Agent SFT tokens to match the top pass@1 of the Base prior, which required 1.5x2^28 tokens.
• SFT prior shows significant degradation with 200 SWE-Agent trajectories, attributed to memorization of a single data mode.

Limitations

• TestWriter still produces false positives, and self-play remains below pass@N with ground-truth tests.
• Reflection skill analysis is coarse-grained and conflates TestWriter activity in SWE-Agent trajectories.
• End-to-end SWE-Agent RL from MT prior collapsed quickly due to few solvable problems.

Experimental setup

Datasets: SWE-bench Verified SWE-Gym SWE-bench-extra R2E-Gym-Lite SWE-smith trajectories GitHub PR data

Benchmarks: SWE-bench Verified SWE-bench-live SWE-bench Multilingual

Metrics: pass@1 pass@3 pass@10 resolve rate

Baselines: Llama3-SWE-RL Seed1.5-Thinking OpenAI-o1 DeepSeek-R1-0120 OpenAI-o3-mini-high Claude 3.5 Sonnet (241022) MiniMax-M1 DeepSeek-R1-0528 SWE-SWISS SWE-agent-LM

Builds on: Agentless SWE-Agent Kimi k1.5 policy optimization GRPO SWE-Gym

Closest papers by meaning

• daVinci-Dev: Agent-native Mid-training for Software Engineering (ICML 2026 · similarity 0.90)
• SWE-rebench: An Automated Pipeline for Task Collection and Decontaminated Evaluation of Software Engineering Agents (NeurIPS 2025 · similarity 0.89)
• Training Software Engineering Agents and Verifiers with SWE-Gym (ICML 2025 · similarity 0.89)
• Toward Training Superintelligent Software Agents through Self-Play SWE-RL (ICML 2026 · similarity 0.88)
• Generalizable End-to-End Tool-Use RL with Synthetic CodeGym (ICLR 2026 · similarity 0.88)
• Hybrid-Gym: Training Coding Agents to Generalize Across Tasks (ICML 2026 · similarity 0.88)