Search papers, labs, and topics across Lattice.
This paper introduces STRATAGEM, a novel self-play framework that enhances the transferability of reasoning in language models by addressing domain specificity and contextual stasis. By employing a Reasoning Transferability Coefficient and a Reasoning Evolution Reward, STRATAGEM selectively reinforces trajectories that demonstrate abstract reasoning while promoting adaptive decision-making. Experimental results show significant improvements in reasoning tasks, particularly in competition-level mathematics, highlighting the effectiveness of the proposed approach in fostering generalizable reasoning capabilities.
STRATAGEM reveals that selectively reinforcing reasoning trajectories can dramatically enhance a model's ability to transfer reasoning skills across diverse tasks, especially in complex mathematical scenarios.
Games offer a compelling paradigm for developing general reasoning capabilities in language models, as they naturally demand strategic planning, probabilistic inference, and adaptive decision-making. However, existing self-play approaches rely solely on terminal game outcomes, providing no mechanism to distinguish transferable reasoning patterns from game-specific heuristics. We present STRATAGEM, which addresses two fundamental barriers to reasoning transfer: domain specificity, where learned patterns remain anchored in game semantics, and contextual stasis, where static game contexts fail to cultivate progressive reasoning. STRATAGEM selectively reinforces trajectories exhibiting abstract, domain-agnostic reasoning through a Reasoning Transferability Coefficient, while incentivizing adaptive reasoning development via a Reasoning Evolution Reward. Experiments across mathematical reasoning, general reasoning, and code generation benchmarks demonstrate substantial improvements, with particularly strong gains on competition-level mathematics where multi-step reasoning is critical. Ablation studies and human evaluation confirm that both components contribute to transferable reasoning.
