Comparative reversal learning reveals rigid adaptation in LLMs under non-stationary uncertainty

LLMs struggle to learn from negative feedback: Across all three models tested (DeepSeek-V3.2, Gemini-3, and GPT-5.2), the models showed "markedly attenuated" loss-shift behavior compared to humans, meaning they were significantly less responsive to failures than successes—a concerning asymmetry for safety-critical applications where learning from mistakes is essential.

High performance can mask inflexible decision-making: The research found that LLMs could maintain good overall scores while simultaneously showing "rigid adaptation" to changing circumstances, demonstrating that aggregate success metrics alone don't capture whether a model can actually adjust its behavior when conditions change—an important distinction for deployment in dynamic real-world environments.

Current benchmarks may miss critical vulnerabilities: Standard evaluation methods don't capture how LLMs handle non-stationary (changing) environments. The authors argue for "reversal-sensitive diagnostics," suggesting that existing tests may give false confidence about model robustness in situations where rules or contexts shift over time.

Different models fail in different ways: DeepSeek-V3.2 showed extreme stubbornness after rule changes, while Gemini-3 and GPT-5.2 adapted faster but still underweighted negative evidence—indicating that model architecture and training produce distinct failure modes that require tailored monitoring approaches.

Uncertainty amplifies problematic rigidity: When the environment became more unpredictable (random transition schedule), LLMs became even more rigid in their responses, suggesting that volatility and uncertainty could push these systems toward inflexible behavior in real-world deployment scenarios.