This model is a reinforcement learning (RL) post-trained language model optimized with the ∆Belief reward signal introduced in ∆Belief-RL.

Instead of relying solely on sparse success signals, the model is rewarded for reducing its own belief uncertainty over time. This enables dense feedback and turn-level credit assignment in long-horizon, open-ended information-seeking tasks.

The model was trained in the Twenty Questions environment and learns general information-seeking strategies that generalize beyond the training horizon.