Skip to content

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

🕒 Published (v1): 2026-01-01 · Source: ICLR · Venue: ICLR 2026 · link

Why this paper was selected

Kimi team; agentless skill prior lifts SWE-agent training, state-of-the-art SWE-bench

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

Kimi-Dev proposes treating Agentless workflow training as a skill-prior stage rather than an end goal, using it to inject localization, code editing, and self-reflection capabilities before adapting to full multi-turn SWE-agent frameworks. A 72B model trained with mid-training, cold-start, RLVR, and test-time self-play achieves 60.4% on SWE-bench Verified under the Agentless framework; with only 5k SFT trajectories added for agentic fine-tuning, it reaches 48.6% pass@1 under SWE-Agent—matching Claude 3.5 Sonnet (20241022).

Problem

Training SWE-agents end-to-end from scratch is brittle: sparse outcome rewards over long trajectories cause RL instability, generic initialization leads to degenerate tool-use behaviors (infinite loops, tool-call failures), and high-quality agentic trajectory collection is expensive. Agentless methods are easier to train but sacrifice exploration flexibility. The field treats the two paradigms as mutually exclusive, leaving the structured inductive bias of Agentless training underutilized.

Method

Agentless training recipe (BugFixer + TestWriter duo): 1. Mid-training on ~150B tokens from GitHub issues/PRs: ~50B tokens of Agentless-formatted natural diff patches, ~20B curated PR commit packs, ~20B synthetic reasoning + agentic data (4Ă— upsampled). Base: Qwen 2.5-72B-Base. 2. Cold-start SFT on reasoning trajectories generated by DeepSeek-R1 acting as BugFixer/TestWriter over SWE-Gym and SWE-bench-extra, activating long chain-of-thought. 3. RL (code-edit stage only): REINFORCE-style policy gradient (from Kimi k1.5) with execution-based binary rewards. Outcome-only rewards (no format signals). Adaptive curriculum: start with pass@16 > 0 prompts (~1,200), reintroduce harder prompts every 100 steps when success rate crosses threshold. Positive-example replay in late-stage training. 4. Test-time self-play: Generate 40 BugFixer patches \(B\) and 40 TestWriter tests \(T\); cross-execute all \((b_i, t_j)\) pairs. Score each patch \(b_i\) as: $\(S_i = \frac{\sum_j \text{FP}(i,j)}{\sum_j F(j)} + \frac{\sum_j \text{PP}(i,j)}{\sum_j P(j)}\)$ where FP/PP are fail-to-pass and pass-to-pass counts; select \(\arg\max_i S_i\).

SWE-Agent adaptation: SFT Kimi-Dev on 5,016 publicly available SWE-Agent trajectories (SWE-smith, collected with Claude 3.7 Sonnet), max 64K context training / 128K + 100 turns inference, in the SWE-Agent five-stage framework.

Key Contributions

  • Reframes Agentless training as a skill-prior induction stage rather than a final system, challenging the agentless/agentic dichotomy.
  • Kimi-Dev (72B): open-source SWE LLM reaching 60.4% on SWE-bench Verified—SoTA among workflow-based/agentless systems at time of publication.
  • Demonstrates that each training stage (mid-training → cold-start → RL) progressively improves the efficiency of downstream SWE-agent SFT adaptation (measured by fast adaptation with small data subsets).
  • Test-time self-play with 3Ă—3 patch-test pairs already outperforms majority-voting over 40 BugFixer patches, showing execution-based test feedback is highly informative.
  • Skills generalize to out-of-distribution benchmarks: SWE-bench-live and SWE-bench Multilingual.

Results

  • Agentless framework: 60.4% on SWE-bench Verified (40 patches Ă— 40 tests); surpasses SWE-SWISS (58.2%), DeepSeek-R1-0528 (57.6%), MiniMax-M1 (56.0%), Claude 3.5 Sonnet agentless (50.8%).
  • SWE-Agent framework (after 5k-trajectory SFT): 48.6% pass@1; comparable to Claude 3.5 Sonnet SWE-Agent (49.0%), surpasses SWE-agent-LM (40.2%) using identical SFT data.
  • Higher ceiling of agentic mode: pass@10 = 74.0% (SWE-Agent SFTed) vs. pass@30 = 73.8% (Agentless), confirming agentic framework has higher resolution potential.
  • Self-play scaling: performance rises from 48.0% (1Ă—1) to 60.4% (40Ă—40); 3Ă—3 already beats 40-patch majority vote.
  • Mid-training scaling: performance increases monotonically from 50B → 100B → 150B training tokens.
  • RL training: both BugFixer pass@1 and TestWriter reproduction rate increase steadily with training steps, correlated with increasing response length (in-depth self-reflection).
  • Skill prior ablation: RL-trained Kimi-Dev prior requires fewer SWE-Agent SFT trajectories to reach the same performance versus Base / MT / SFT-cold-start priors.

Limitations

  • TestWriter false positives occur during RL training due to insufficient reproduction coverage; acknowledged as future work.
  • Self-play performance remains below pass@N with ground-truth tests as oracle, indicating TestWriter quality is a bottleneck.
  • SWE-Agent SFT uses only 5k trajectories from a synthetic environment (Claude 3.7 Sonnet); no additional multi-turn agentic RL was performed—leaving likely headroom.
  • Evaluation under SWE-Agent uses a specific five-stage framework (Anthropic-style); generalization to other agentic scaffolds (e.g., OpenHands) is not fully characterized for Kimi-Dev.
  • Sandbox infrastructure (Kubernetes, >10k concurrent instances) represents a resource cost that may limit reproducibility for academic labs.

Relevance to Agentic AI / LLM Agents

Kimi-Dev directly addresses the core bottleneck in training capable LLM agents: cold-start brittleness and sparse credit assignment over long horizons. By establishing that structured Agentless training induces transferable atomic skill priors—localization, code editing, self-reflection—the work provides a practical curriculum recipe for bootstrapping agentic LLMs without expensive end-to-end trajectory collection. The finding that RL-trained long-CoT reasoning (originally single-turn) transfers to multi-turn horizon extension is significant: it implies that reasoning-intensive single-turn training is not wasted for agent scenarios and may serve as a general pre-adaptation technique across domains beyond SWE. This complements the broader line of work on modular vs. end-to-end agent training (SWE-Agent, OpenHands, DeepSWE) by offering a principled intermediate stage rather than forcing a binary choice.