A pre-trained regression model can tackle reinforcement learning problems without ever being trained on RL tasks.
Researchers introduced ICR-RL, a gradient-free method that repurposes a foundation model built for regression - specifically TabPFN, trained on a wide range of tabular regression tasks - and applies it directly to RL control problems. The trick is treating RL as a regression problem, a classical theoretical reduction that has rarely been exploited at the foundation model scale. Tested against Gymnasium's classic-control benchmark, ICR-RL held its own against established RL algorithms including DQN, PPO, and TRPO. No fine-tuning, no RL-specific environment data, no additional training loop.
The result matters because building diverse RL training environments is expensive and brittle - it's one of the main reasons scaling RL foundation models has lagged behind language and vision. If a general-purpose regression model can sidestep that bottleneck, it changes the cost calculus for anyone trying to apply RL to new domains. The finding also raises a quieter question: how much of what we attribute to RL-specific training is actually just competent function approximation?
Classic-control benchmarks like CartPole and Pendulum are the "hello world" of RL - the real test will be whether this approach survives contact with sparse rewards, partial observability, and environments that punish the kind of smooth interpolation regression models love.