Papers.Data.Code
Paper

MindForge: Teaching Small Language Models Whole-Life-Cycle Software Engineering via Source-Free Program Synthesis

Yihao Chen, Shi Chang, Khaled Chawa, Feng Lin, Boyuan Chen, Shaowei Wang, Ahmed E. Hassan

Source-free environments replace source-visible repos, and teacher trajectories supervise full program construction from docs plus executable behavior.

arXiv ↗Jul 29, 2026

  • nlp
  • llm
  • code generation
  • agent distillation
  • software engineering
Headline benchmark comparison showing MindForge vs base gains across multiple tasks
Headline benchmark comparison showing MindForge vs base gains across multiple tasks

What it is

Builds source-free software-engineering environments from open-source CLIs and distills long program-synthesis trajectories into a 27B model. Instead of training on source-visible bug-fix tasks, it turns open-source CLI repos into source-free executable environments and distills whole-life-cycle synthesis trajectories from a strong teacher.

Key results

  • 49.51% on ProgramBench vs 37.98% for Qwen3.6-27B base
  • Beats DeepSeek V4 Pro on ProgramBench: 49.51% vs 47.80%
  • +31.00 pts on RepoZero-C2Rust; +14.16 on DeepSWE vs base

Whole-life-cycle, source-free synthesis trajectories transfer beyond from-scratch coding and improve broad software-engineering agent performance.

One useful ML email a week.

More from Jul 27 – Aug 02

The whole week →

Picked Jul 31

Picked Jul 30

Picked Jul 28

  • Repo
    awesome-4dgs/awesome-4dgs

    Collects 79 papers into task-specific sections with contribution summaries, code links, and representation tags for faster literature triage.

    GitHubJul 23, 2026
    →

Worth your time, not all of your time.

We track new ML work across papers, code, and data, then keep only the pieces worth a closer look. Selected, not collected.

Papers.Data.Code

The Weekly ML Signal

Selected, not collected.

One useful ML email a week: the papers, repos and datasets worth your attention.

Worth your time, not all of your time.