EGL-SCA: Structural Credit Assignment for Co-Evolving Instructions and Tools in Graph Reasoning Agents
Researchers have introduced EGL-SCA, a novel verifier-centric dual-space framework designed to enhance graph reasoning agents operating on natural-language inputs. Addressing the limitations of existing approaches that isolate instruction or tool improvements, EGL-SCA models agents using two collaborative components: an instruction-side policy space for reasoning strategies and a tool-side program space for executable algorithms. The framework's core mechanism, structural credit assignment, maps trajectory evidence to conditional updates, precisely directing failures toward either prompt optimization or tool synthesis and repair. To ensure robust learning, the method employs a training distribution stratified by task family and a Pareto-style retention strategy to balance success, generality, and parsimony. Experimental results across four graph reasoning benchmarks demonstrate that EGL-SCA achieves a state-of-the-art average success rate of 92.0%. This performance significantly surpasses both pure-prompting and fixed-toolbox baselines, highlighting the effectiveness of co-evolving instructions and tools for complex structured reasoning tasks in artificial intelligence.
Wire timeline
EGL-SCA: Structural Credit Assignment for Co-Evolving Instructions and Tools in Graph Reasoning Agents
Researchers have introduced EGL-SCA, a novel verifier-centric dual-space framework designed to enhance graph reasoning agents operating on natural-language inputs. Addressing the limitations of existing approaches that isolate instruction or tool improvements, EGL-SCA models agents using two collaborative components: an instruction-side policy space for reasoning strategies and a tool-side program space for executable algorithms. The framework's core mechanism, structural credit assignment, maps trajectory evidence to conditional updates, precisely directing failures toward either prompt optimization or tool synthesis and repair. To ensure robust learning, the method employs a training distribution stratified by task family and a Pareto-style retention strategy to balance success, generality, and parsimony. Experimental results across four graph reasoning benchmarks demonstrate that EGL-SCA achieves a state-of-the-art average success rate of 92.0%. This performance significantly surpasses both pure-prompting and fixed-toolbox baselines, highlighting the effectiveness of co-evolving instructions and tools for complex structured reasoning tasks in artificial intelligence.
cs.AI updates on arXiv.org