Based on arXiv:2610.27028v1, this work investigates whether compact AI agents can exhibit strategic behavior under hidden information and whether preference optimization can surpass static zero‑shot policies. We first evaluate a static baseline and three frozen language models—SmallThinker, Qwen3, and Gemma‑3—and observe substantial performance gaps between exact‑game and six‑player evaluations. To address this, we introduce Strategic Preference Optimization with KL‑anchored Equilibrium (SPOKE), which combines teacher‑relative value penalties, policy anchoring, risk constraints, and structured rationale supervision, and compare it against SFT, DPO, GRPO, and GSPO. Experiments on a SmallThinker‑4B‑A0.6B policy across three runs show that SPOKE achieves the lowest mean exploitability in Kuhn ($0.213$) and Leduc ($1.081$), and the highest six‑max top‑1 agreement ($0.775$). Teacher KL remains comparable at $0.522$ versus $0.523$ for SFT. These findings provide preliminary evidence that strategic preference optimization can improve strategic quality while preserving teacher‑consistent behavior.
Review: SPOKE balances value‑driven incentives with behavioral constraints, offering a promising direction for strategic learning in games with hidden information. Future work may explore scaling to larger models and more complex game settings.