GRPO matches POMO solution quality within 2% on TSP/CVRP while avoiding REINFORCE training collapse on TSP-100 without needing a rollout baseline.
Preference optimization for combinatorial optimization problems,
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Baseline-Free Policy Optimization for Neural Combinatorial Optimization
GRPO matches POMO solution quality within 2% on TSP/CVRP while avoiding REINFORCE training collapse on TSP-100 without needing a rollout baseline.