When to Re-Plan: Subgoal Persistence in Hierarchical Latent Reasoning

Ayushi Chadha · 2026

Read When to Re-Plan: Subgoal Persistence in Hierarchical Latent Reasoning by Ayushi Chadha (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2606-03741.txt.

Description

Chadha studies how long a reasoning system should hold a subgoal before re-planning. Long tasks need commitment to medium-range intent without rigidity. Re-plan too often and computation never coheres. Commit too long and the plan goes stale. The setting is latent reasoning, where steps unfold in hidden state, not in token traces.

A feudal manager-worker design extends the Hierarchical Reasoning Model. A slow high-level module emits a normalized subgoal that lasts for P worker steps. It biases the worker's hidden updates and adds a cosine alignment loss. On ARC and ConceptARC, persistence itself is the key knob. Periods of three to six steps beat both constant re-planning and long holds. Loss bottoms at period three. The alignment weight has a narrow optimum near 0.05. Ablations tie interference past that point to learned directional structure. Designers get a clear rule from the sweep.

Catalog record

CREATORAyushi Chadha
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE34 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2606.03741

Browse the full library · About the Tower of Babel