Replacing thinking with tool usage enables reasoning in small language models

TL;DR

This paper introduces a method where small language models use tool interactions instead of internal reasoning to perform tasks, enabling faster learning and reasoning capabilities.

Contribution

It proposes formatting reasoning as multi-turn tool interactions with a custom DSL, improving training efficiency and reasoning in small models.

Findings

Small models can learn to repair Python code using tool interactions.

The approach enables faster sampling and denser reward signals.

Models up to 3B parameters achieve proficient task performance.

Abstract

Recent advances have established a new machine learning paradigm based on scaling up compute at inference time as well as at training time. In that line of work, a combination of Supervised Fine-Tuning (SFT) on synthetic demonstrations and Reinforcement Learning with Verifiable Rewards (RLVR) is used for training Large Language Models to expend extra compute during inference in the form of "thoughts" expressed in natural language. In this paper, we propose to instead format these tokens as a multi-turn interaction trace with a stateful tool. At each turn, the new state of the tool is appended to the context of the model, whose job is to generate the tokens necessary to control the tool via a custom DSL. We benchmark this approach on the problem of repairing malfunctioning Python code, and show that this constrained setup allows for faster sampling of experience and a denser reward signal, allowing even models of size up to 3B parameters to learn how to proficiently expend additional compute on the task.