EGL-SCA: Structural Credit Assignment for Co-Evolving Instructions and Tools in Graph Reasoning Agents

Researchers introduce EGL-SCA, a framework for graph reasoning agents that jointly optimizes both natural language instructions and computational tools through structural credit assignment. The system achieves 92.0% success rate on graph reasoning benchmarks by precisely routing failures to either prompt optimization or tool synthesis, outperforming isolated improvement approaches.

EGL-SCA addresses a fundamental challenge in AI agent design: determining what component fails when a complex system produces incorrect outputs. Traditional approaches treat instruction quality and tool capabilities as separate problems, leaving developers uncertain whether to refine prompts or rebuild algorithms. This research proposes a verifier-centric framework that treats both dimensions as co-evolving systems, using structured credit assignment to map failure evidence back to its root cause.

The breakthrough mechanism lies in its dual-space architecture. Rather than treating instructions and tools as fixed components, EGL-SCA models them as complementary policy spaces that adapt together. The framework introduces a stratified training distribution and Pareto retention strategy to balance success rates, generality across tasks, and algorithmic simplicity—preventing the system from converging on overly complex or brittle solutions.

For the AI development community, this represents meaningful progress toward more debuggable and maintainable agent systems. Current deployed agents often fail in ways that are difficult to diagnose; this work provides a systematic approach to understanding whether improvements should target reasoning strategies or underlying computational tools. The 92% average success rate across four benchmarks demonstrates practical viability, particularly for applications requiring structured outputs like knowledge graphs or program synthesis.

The implications extend beyond academia to production systems. As organizations deploy graph reasoning agents for enterprise applications, the ability to efficiently diagnose and fix failures becomes economically valuable. Future work likely focuses on scaling this approach to more complex reasoning tasks and exploring how structural credit assignment applies to other multi-component AI systems.

• →EGL-SCA uses structural credit assignment to route failures between instruction optimization and tool synthesis in graph reasoning agents
• →The framework achieves 92.0% average success rate across four graph reasoning benchmarks, exceeding pure-prompting and fixed-toolbox baselines
• →Dual-space architecture treats natural language reasoning strategies and executable tools as co-evolving components rather than isolated systems
• →Pareto-style retention strategy balances performance, generality, and algorithmic parsimony during training
• →Verifier-centric design prioritizes structured correctness over textual plausibility for graph reconstruction tasks