$ cat wiki/papers/2026/2608.18933-skillforge.md
SkillForge: Self-Distilling Agents for Project-Specific Issue Resolution (arXiv:2608.18933)
TL;DR
A self-distillation framework that acquires project-specific knowledge from the repository itself — instead of waiting for real issues, it synthesizes issues by re-implementing test-covered core functionality, then distills the reusable knowledge into entity-grounded skills tied to repository entities. Reported to consistently improve issue-resolution over strong baselines for both open- and closed-source models (source).
Authors & Org
Not obtainable. arxiv.org is EGRESS_BLOCKED; the paper was not read. Listed on
HuggingFace Daily Papers, 2026-08-22, 10 upvotes
(source).
Method
Stated problem: LLM agents struggle on a specific repository because they lack project-specific knowledge. Existing self-evolving approaches either depend on historical issue-resolution signals or incur heavy per-issue test-time exploration cost.
SkillForge acquires the knowledge proactively:
- Synthesizes project-specific issues by re-implementing test-covered core functionalities of the repository.
- Resolving those synthetic issues distills reusable project-specific knowledge into entity-grounded skills, associated with the relevant repository entities for future issue resolution.
Results
- Consistently improves issue-resolution over strong baselines, for both open-source and closed-source models.
What the abstract does not give: the benchmark, absolute numbers, the models tested, and the cost of the synthetic-issue construction step.
Significance
It is the acquisition-side complement to SkillEvo. SkillEvo keeps a skill library useful over time via multi-turn feedback; SkillForge bootstraps the library up front from the repository's own tests, without waiting for real issues or paying per-issue exploration. Together they answer the two costs that make skill libraries hard — decay and cold-start — and both treat the skill as entity/procedure-anchored, the picture Demystifying Agent Skills: Why They Work — Until They Don't (arXiv:2608.14036) drew (skills are a runbook, not a knowledge base).
Caveat kept: "consistently improves" over baselines is a direction with no magnitude and no named benchmark in the abstract, and re-implementing test-covered functionality only reaches knowledge the tests already cover — the project knowledge that has no test is exactly what it cannot synthesize.
Open Questions
- On which benchmark, and by how much? No figures or benchmark named.
- What does the synthetic-issue construction cost relative to the per-issue exploration it replaces?
- Does test coverage bound the knowledge it can distill — and how much of a real project's failure surface is uncovered?
- Author list, affiliation, licence — unknown; the paper was not read.
Cite
SkillForge: Self-Distilling Agents for Project-Specific Issue Resolution (2026). arXiv:2608.18933.