GraphEvoResearcher
A training-free self-reflective reasoning protocol for multi-hop question answering. It organizes the reasoning process as an explicit reasoning graph in which nodes represent intermediate conclusions or evidence and edges encode support, contradiction, and dependency relations. A graph-aware meta-reward function combining correctness, efficiency, reflection quality, diversity, and structural signals decides whether to revise, backtrack, or stop, thereby improving accuracy and reducing reasoning cost.
Timeline (last 90 days)
GraphEvoResearcher proposes a training-free self-reflection protocol that lifts self-reflection meta-rewards onto an explicit reasoning graph, where nodes are intermediate conclusions/evidence and edges encode support, contradiction, and dependency relations.
GraphEvoResearcher designs a graph-aware meta-reward function that combines five types of signals to decide the next action: correctness, efficiency, reflection quality, diversity, and structural signals (connectivity, acyclicity, branch coverage, and path convergence).
On a real LLM (DeepSeek), the researchers evaluated 50 samples each on the GSM8K, MATH, and BBH benchmarks, and GraphEvoResearcher achieved accuracies of 98%, 92%, and 82% respectively, the highest among all baselines.
On the MATH dataset, GraphEvoResearcher improves by 24 percentage points over single-pass generation and by 12 percentage points over flat reflection.
The construction process of GraphEvoResearcher stays at about 2 steps and 4.5–5 nodes, far below the consumption of exhaustive reflection.
The experimental scale is 50 samples per dataset, and the statistical robustness of the conclusions remains to be verified at larger scale.
Graph construction itself relies on the LLM correctly extracting nodes and edge relations, and extraction errors may propagate to subsequent decisions.
All Facts (7)
GraphEvoResearcher proposes a training-free self-reflection protocol that lifts self-reflection meta-rewards onto an explicit reasoning graph, where nodes are intermediate conclusions/evidence and edges encode support, contradiction, and dependency relations.
50%UnverifiedGraphEvoResearcher designs a graph-aware meta-reward function that combines five types of signals to decide the next action: correctness, efficiency, reflection quality, diversity, and structural signals (connectivity, acyclicity, branch coverage, and path convergence).
50%UnverifiedOn a real LLM (DeepSeek), the researchers evaluated 50 samples each on the GSM8K, MATH, and BBH benchmarks, and GraphEvoResearcher achieved accuracies of 98%, 92%, and 82% respectively, the highest among all baselines.
50%UnverifiedOn the MATH dataset, GraphEvoResearcher improves by 24 percentage points over single-pass generation and by 12 percentage points over flat reflection.
50%UnverifiedThe construction process of GraphEvoResearcher stays at about 2 steps and 4.5–5 nodes, far below the consumption of exhaustive reflection.
50%UnverifiedThe experimental scale is 50 samples per dataset, and the statistical robustness of the conclusions remains to be verified at larger scale.
50%UnverifiedGraph construction itself relies on the LLM correctly extracting nodes and edge relations, and extraction errors may propagate to subsequent decisions.
50%