Saved in:
Bibliographic Details
Main Authors: Morris, John X., Mireshghallah, Niloofar, Ibrahim, Mark, Mahloujifar, Saeed
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2602.04118
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866908811577524224
author Morris, John X.
Mireshghallah, Niloofar
Ibrahim, Mark
Mahloujifar, Saeed
author_facet Morris, John X.
Mireshghallah, Niloofar
Ibrahim, Mark
Mahloujifar, Saeed
contents Recent research has shown that language models can learn to \textit{reason}, often via reinforcement learning. Some work even trains low-rank parameterizations for reasoning, but conventional LoRA cannot scale below the model dimension. We question whether even rank=1 LoRA is necessary for learning to reason and propose TinyLoRA, a method for scaling low-rank adapters to sizes as small as one parameter. Within our new parameterization, we are able to train the 8B parameter size of Qwen2.5 to 91\% accuracy on GSM8K with only 13 trained parameters in bf16 (26 total bytes). We find this trend holds in general: we are able to recover 90\% of performance improvements while training $1000x$ fewer parameters across a suite of more difficult learning-to-reason benchmarks such as AIME, AMC, and MATH500. Notably, we are only able to achieve such strong performance with RL: models trained using SFT require $100-1000x$ larger updates to reach the same performance.
format Preprint
id arxiv_https___arxiv_org_abs_2602_04118
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Learning to Reason in 13 Parameters
Morris, John X.
Mireshghallah, Niloofar
Ibrahim, Mark
Mahloujifar, Saeed
Machine Learning
Recent research has shown that language models can learn to \textit{reason}, often via reinforcement learning. Some work even trains low-rank parameterizations for reasoning, but conventional LoRA cannot scale below the model dimension. We question whether even rank=1 LoRA is necessary for learning to reason and propose TinyLoRA, a method for scaling low-rank adapters to sizes as small as one parameter. Within our new parameterization, we are able to train the 8B parameter size of Qwen2.5 to 91\% accuracy on GSM8K with only 13 trained parameters in bf16 (26 total bytes). We find this trend holds in general: we are able to recover 90\% of performance improvements while training $1000x$ fewer parameters across a suite of more difficult learning-to-reason benchmarks such as AIME, AMC, and MATH500. Notably, we are only able to achieve such strong performance with RL: models trained using SFT require $100-1000x$ larger updates to reach the same performance.
title Learning to Reason in 13 Parameters
topic Machine Learning
url https://arxiv.org/abs/2602.04118