#recursiveSearch

"Self-Steering Language Models"

This paper introduces DisCIPL, a method for "self-steering" LMs (Language Model) where a Planner model generates a task-specific inference program that is executed by a population of Follower models. Our approach equips LMs with the ability to write recursive search procedures that guide LM inference

arxiv.org/abs/2504.07081

#LLM #recursiveSearch #Planner

In our method (DISCIPL, bottom), a Planner LM writes an inference program that defines step-by-step computations to steer a population of Follower LMs.Examples of various tasks that DISCIPL-SMC was asked to do.

Client Info

Server: https://mastodon.social
Version: 2025.04
Repository: https://github.com/cyevgeniy/lmst