Recursive Language Models Meet Uncertainty: The Surprising Effectiveness of Self-Reflective Program Search for Long Context

Keivan Alizadeh Parshin Shojaee Minsik Cho Mehrdad Farajtabar[0.5cm] Apple · 2026

Read Recursive Language Models Meet Uncertainty: The Surprising Effectiveness of Self-Reflective Program Search for Long Context by Keivan Alizadeh Parshin Shojaee Minsik Cho Mehrdad Farajtabar[0.5cm] Apple (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-2603-15653.txt.

Description

Alizadeh and co-authors add self-reflection to recursive language model search. Recursive models split long texts into program-guided sub-calls at run time. Which interaction program gets picked matters a lot, yet it drew little study. Their SRLM scores candidate programs with three inner signals. The signals are self-consistency, reasoning length, and stated confidence. Tests show gains up to 22 percent over plain RLM at equal time budgets. Plain recursion alone can even hurt short-context work.

Catalog record

CREATORKeivan Alizadeh Parshin Shojaee Minsik Cho Mehrdad Farajtabar[0.5cm] Apple
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE60110
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2603.15653

Browse the full library · About the Tower of Babel