Hung Nguyen

Recency Forcing is proposed, which applies a non-positive, timestep-dependent bias, termed Temporal Response Bias (TRB), on pre-softmax attention logits derived directly from $R, closing the train-inference gap without modifying context length or training objectives.