1
Solver-Hard Is Not Model-Hard: A Hardness-Controlled Diagnostic for LLM Constraint Reasoning
LLM约束推理的性能评测常现误区，这篇论文提出“求解难≠模型难”的诊断方法，揭示评估陷阱。
arXiv:2607.17047v1 Announce Type: cross Abstract: LLM constraint reasoners are often evaluated near the random-SAT phase transition, confounding densi…