Search papers, labs, and topics across Lattice.
This paper introduces Readout Feedback (RoFB), a novel test-time intervention that enhances the performance of recurrent models by utilizing their own readout probabilities to influence latent dynamics without the need for retraining. The authors demonstrate that RoFB effectively improves reasoning tasks in three recurrent models (AKOrN, ItrSA++, TRM) applied to Sudoku and Maze challenges, outperforming traditional methods that rely on increased computation or trajectory sampling. Key findings reveal that RoFB achieves significant performance gains in four out of six model-task pairs while maintaining comparable or lower computational costs, highlighting the potential of closed-loop steering in recurrent reasoning.
Steering latent dynamics with readout feedback can unlock performance improvements in recurrent models that traditional inference methods fail to achieve.
Recurrent models, which repeatedly update latent states with shared computation blocks, have emerged as powerful architectures for solving complex reasoning tasks. Existing inference-time methods scale computation by running more steps or sampling more trajectories, but ignore information revealed within each trajectory. Here we show that recurrent models can be improved at inference time by using their own readout probabilities to steer latent dynamics without retraining. We introduce Readout Feedback (RoFB), a test-time intervention that converts intermediate predictions into token-wise pairwise coupling forces injected into the latent dynamics. Across three recurrent models (AKOrN, ItrSA++, TRM) on Sudoku and Maze, RoFB yields clear gains in four of six model-task pairs, achieving performance unattainable by merely running more steps or selecting from multiple trajectories, at comparable or lower computational cost. These results suggest that closed-loop steering of latent dynamics can serve as a complementary inference-time control mechanism for recurrent reasoning models.
