Subgraph-Decomposed Curriculum RL for Hard Repository-Level Software Engineering
Apply curriculum-based reinforcement learning over decomposed dependency subgraphs of failed repository-generation tasks so coding agents learn multi-file software construction instead of discarding unsolved problems.
Time
Compute level
Resources
What the paper used
Prerequisites
Scope
The Gap
In Section 3.1.3 (Agentic Coding), the pipeline constructs repository-level generation and patch-level SWE tasks using GitNexus code knowledge graphs and mines multiple rollouts per task, filtering out tasks for which no successful solution is obtained. In Table 2, Sage S2.1 Flash trails leading baselines substantially on repository-level and SWE benchmarks (57.2 vs. 79.9 on NL2Repo-Bench; 60.6 vs. 76.5 on SWE-bench Pro; 67.9 vs. 73.9 on DeepSWE).
“Empirical pass rates inform task filtering, removing some overly easy tasks and tasks for which no successful solution is obtained under the current sampling configuration.”
Limitation
Discarding tasks with zero successful rollouts under current sampling budgets starves the model of training signal on precisely the hardest multi-file dependency, interface-evolution, and architectural synthesis problems reflected in NL2Repo-Bench and SWE-bench Pro.
Proposed Solution
Instead of dropping zero-pass repository tasks, use the GitNexus code-structure graph to decompose each failed repository task into an ordered curriculum of verifiable sub-tasks (e.g., core data structures, module interfaces, integration tests). Run multi-turn RL (GRPO) with dense intermediate rewards combining unit-test pass fraction, static interface/type-check compatibility, and regression-free patch execution, gradually merging sub-tasks back into full repository-level synthesis.
The paper’s authors list this direction as future work, so they may already be pursuing it.
Why This Matters
Recovering learning signal from previously unsolved tasks is expected to narrow the 15-22 point gap on repository-level coding benchmarks while improving multi-file compatibility.
How to Get Started
How to test it: Hypothesis: Curriculum RL over subgraph-decomposed zero-pass repository tasks outperforms standard rejection-sampled SFT on long-horizon coding benchmarks. Baseline: Current Sage S2.1 Flash checkpoint evaluated on NL2Repo-Bench (57.2%) and SWE-bench Pro (60.6%); stop if pass rates on held-out NL2Repo-Bench tasks fail to improve by at least 4 percentage points after curriculum RL.
- Collect the zero-pass repository-level and SWE tasks filtered out during the Section 3.1.3 rollout collection stage (or replicate zero-pass tasks from NL2Repo-Bench/OpenSWE training splits).
- Use GitNexus dependency graphs and AST analysis to split each failed task into staged interface and module-level sub-tasks with runnable unit-test subsets.
- Implement partial unit-test and interface-compatibility reward shaping in XTuner over Harbor/OpenHands sandboxes.
- Train with GRPO from stage-1 sub-tasks to full-repository tasks and benchmark on NL2Repo-Bench and SWE-bench Pro.
Suggested by AI. Adapt it to your setup.
Has This Been Done?
Works like WEAVE discuss repository-level benchmarking and task decomposition, but none combine GitNexus-based code-structure graph decomposition into ordered verifiable sub-tasks with multi-turn GRPO and dense intermediate rewards for code agents.
- Retrieval-Augmented Code Generation: A Survey with Focus on Repository-Level Approaches · 2025 · 49 citations
- WEAVE: Weaving Real-World Code Evolution into Dependency-Structured Tasks for Benchmarking and Training Code Agents · 2026 · 3 citations
We searched OpenAlex and arXiv for “repository level code generation curriculum reinforcement learning dependency graph” and “Subgraph-Decomposed Curriculum RL for Hard Repository-Level Software Engineering” and an AI compared the idea with the top 10 results. A small search can miss relevant work.
Source Paper
Sage S2.1 Flash Technical ReportKun Cai, Haodong Chen, Guangran Cheng +203 more · 2026
Sage S2.1 Flash is a 320-billion-parameter mixture-of-experts (MoE) model designed to integrate native scientific perception with general-purpose multimodal reasoning and long-horizon agentic execution. It combines dedicated scientific encoders and modality-specific projectors (Sage Sci-Perception) with post-training across large-scale verifiable environments spanning tool use, search, coding, ML engineering, workspace deliverables, and computer use, achieving top-two rankings on many of 33 evaluated benchmarks.
Related Ideas
- Hierarchical Specification-to-Patch Synthesis for Large-Scale Repository ChangesSWE-CoWork: Benchmarking Coding Agents as Digital Employees in Organizations🟠Medium
- Cross-Repository Micro-Skill Generalization via Project Manifest EmbeddingsTokenEconomizer: Experience-Driven Black-Box Cost Control for Coding Agents🟡Fairly easy
- Evaluation on Repository-Scale Benchmarks (SWE-bench)AutoDev: Automated AI-Driven Development🟡Fairly easy
- Backtracking Search for Multi-Step Coding ChainsGLM-5: from Vibe Coding to Agentic Engineering🟡Fairly easy
- Utility-Tracked Consolidation and Failure-Aware Retrieval for Harness Improvement LibrariesHarness^2: Recursive Agent Harnessing for an Open World🟡Fairly easy