New arXiv study: Transfer learning fails when policy regimes shift abruptly

A new study from Roberto Garrone, posted on arXiv (2607.09685), tackles a critical blind spot in adaptive multi-agent systems: what happens when the very rules of the game change? The paper, *Transfer Learning Across Policy Regimes in Adaptive Multi-Agent Systems*, reframes regulatory shifts as a transfer-learning problem. In this setting, a 'policy regime' is a learning problem defined by an observable input distribution and a target function that maps policy variables to outcomes. The author compares two approaches: a blank-slate learner that searches a flexible hypothesis class in the new regime, and a transfer learner that restricts its hypothesis class using structural knowledge from the old regime.

The key insight is that transfer can be both beneficial and harmful. Using a stylized emissions-regulation environment and an agent-based model robustness experiment, Garrone demonstrates that when the target regime preserves an affine monotone relationship between tax and emissions, the transfer learner achieves better small-sample performance. However, when the regime introduces a threshold break—a sudden change in how taxes affect emissions—the same transferred structure causes negative transfer: held-out error remains high, online predictions generate more mistakes, and cumulative error grows over repeated streams. The paper is a methodological contribution, arguing that previous regulatory experience should be reused only when it captures stable structural invariants, and treated with caution when policy changes alter the core policy-outcome relationship.