- Shanghai Jiao Tong University researchers introduced SkillForge, a self-distillation framework that learns project-specific knowledge by synthesizing and resolving artificial issues within a codebase before any real bugs appear.
- On SWE-bench Pro, SkillForge improved issue resolution over Mini-SWE-Agent by up to 5.8 percentage points across both open-source and closed-source models, without requiring historical issue data or costly per-issue exploration.
- SkillForge generates entity-grounded skills tied to specific files, functions and modules, letting agents retrieve targeted knowledge at resolution time rather than exploring the codebase from scratch on every new issue.
A team at Shanghai Jiao Tong University has built a way to give AI agents codebase-specific knowledge before they ever touch a real bug. The framework, called SkillForge, synthesizes its own training material directly from a repository’s test-covered code, sidestepping the two most expensive requirements of existing approaches: a backlog of historical issues and heavy per-issue exploration at runtime.
How SkillForge Builds Skills
Most SWE agents learn reactively. They either accumulate knowledge from past issue resolutions or spend significant compute exploring a codebase each time a new ticket arrives. Both paths are slow, and both fail on fresh or fast-moving projects where there is no issue history to draw from.
SkillForge flips that. The framework starts by identifying test-covered core functionalities in the repository, then traces their execution to find the code regions that work together to produce a given outcome. It rewrites those regions under constrained contexts, effectively manufacturing project-specific issues. Because the synthetic issues are grounded in the repository’s own tests, they stay behaviorally accurate and immediately relevant.
A SWE agent then resolves each synthetic issue, generating resolution trajectories. SkillForge distills those trajectories into what the paper calls “entity-grounded skills”: reusable, codebase-specific problem-solving patterns tied to particular files, functions or modules. When a real issue lands, the agent retrieves the skills linked to the relevant entities rather than exploring cold. The published paper, posted to arXiv on August 19, 2026, includes code and data at github.com/cslsolow/SkillForge.
Benchmark Results
On SWE-bench Pro SkillForge improved issue resolution over Mini-SWE-Agent by +5.8 and +4.1 percentage points across open-source and closed-source model variants respectively. The gains held consistently across baselines tested by the research team. For context on how SWE-bench Pro compares to other agent coding benchmarks, recent SWE-bench Pro evaluations have become a standard reference point for ranking production-grade SWE agents.
The numbers matter less than what they demonstrate: that proactively forging project knowledge, before encountering real issues, consistently outperforms reactive skill acquisition. The cold-start problem that hobbles most SWE agents on new repositories is a solvable engineering problem, not an inherent limitation of LLM-based code agents.
What This Changes for Builders
For teams deploying agents on active codebases, the practical implication is significant. An agent that arrives pre-loaded with entity-grounded skills specific to your repository can contribute meaningfully from day one, without needing weeks of accumulated issue data or expensive per-ticket exploration budgets. That matters most in fast-moving projects where the codebase shifts faster than any historical dataset can track.
The entity-grounding mechanism is the detail worth watching. By tying distilled skills to specific files and functions rather than abstract patterns, SkillForge makes retrieval precise. The agent is not searching for vaguely relevant past experience; it is pulling knowledge anchored to the exact module the incoming issue touches. That architecture is closer to how a senior engineer navigates a familiar codebase than how most current agents operate. Research on what agents must learn from production environments has pushed in the same direction, the gap between sandbox performance and real-world contribution is a retrieval and grounding problem as much as a model capability problem.
What Comes Next
SkillForge’s synthetic issue generation approach also opens a path toward continuous learning. As a codebase evolves, the framework can regenerate skills from updated test-covered functionality, keeping the agent’s knowledge current without waiting for real bugs to accumulate. Whether that continuous loop holds up in practice across large monorepos or rapidly shifting dependency graphs is the open question the paper does not fully answer.
