{"slug": "steering-recurrent-reasoners-at-inference-time-with-readout-feedback", "title": "Steering Recurrent Reasoners at Inference Time with Readout Feedback", "summary": "A new method called Readout Feedback (RoFB) improves recurrent reasoning models at inference time by using their own readout probabilities to steer latent dynamics without retraining. In tests across three recurrent models (AKOrN, ItrSA++, TRM) on Sudoku and Maze tasks, RoFB achieved clear gains in four of six model-task pairs, outperforming approaches that simply run more steps or sample more trajectories at comparable or lower computational cost.", "body_md": "arXiv:2608.24136v1 Announce Type: new\nAbstract: 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.", "url": "https://wpnews.pro/news/steering-recurrent-reasoners-at-inference-time-with-readout-feedback", "canonical_source": "https://www.machinebrief.com/news/steering-recurrent-reasoners-at-inference-time-with-readout-hd52", "published_at": "2026-08-26 04:00:00+00:00", "updated_at": "2026-08-26 06:14:05.756656+00:00", "lang": "en", "topics": ["artificial-intelligence", "machine-learning", "neural-networks", "ai-research"], "entities": ["Readout Feedback (RoFB)", "AKOrN", "ItrSA++", "TRM", "arXiv"], "alternates": {"html": "https://wpnews.pro/news/steering-recurrent-reasoners-at-inference-time-with-readout-feedback", "markdown": "https://wpnews.pro/news/steering-recurrent-reasoners-at-inference-time-with-readout-feedback.md", "text": "https://wpnews.pro/news/steering-recurrent-reasoners-at-inference-time-with-readout-feedback.txt", "jsonld": "https://wpnews.pro/news/steering-recurrent-reasoners-at-inference-time-with-readout-feedback.jsonld"}}