When to Re-Plan: Subgoal Persistence in Hierarchical Latent Reasoning
The paper titled 'When to Re-Plan: Subgoal Persistence in Hierarchical Latent Reasoning' explores the balance between stability and adaptability in long-horizon reasoning systems. It introduces a feudal-style manager-worker interface to enhance subgoal persistence, which is crucial for effective multi-step computation. The findings suggest that moderate periods of subgoal persistence lead to better performance compared to very frequent or very long planning horizons.
- ▪The study focuses on the stability-adaptivity tradeoff in latent reasoning systems.
- ▪A feudal-style manager-worker interface is proposed to improve subgoal persistence.
- ▪Moderate periods of subgoal persistence consistently outperform both very frequent and very long planning horizons.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Artificial Intelligence arXiv:2606.03741 (cs) [Submitted on 2 Jun 2026] Title:When to Re-Plan: Subgoal Persistence in Hierarchical Latent Reasoning Authors:Ayushi Chadha View a PDF of the paper titled When to Re-Plan: Subgoal Persistence in Hierarchical Latent Reasoning, by Ayushi Chadha View PDF HTML (experimental) Abstract:Long-horizon reasoning requires a system to commit to medium-horizon intent without becoming rigid: re-plan too often and computation never coheres into multi-step structure; commit too long and the plan goes stale. We study this stability-adaptivity tradeoff in the latent reasoning setting, where multi-step computation occurs inside hidden state rather than externalized token traces.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
