Learning to Reason in 13 Parameters

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.

Sentrial

Monitor metrics, track success rates, and measure ROI for your AI agents

Origin of the rule that swap size should be 2x of the physical memory

I can trace this recommendation that swap should be twice the size of the physical memory as far back as 1989, see SunOS 4 installation manual, page 37, but I wonder if there was ever any technical

blog.smartere » Floppy Disks: the best TV remote for kids

Tiny Core Linux, Micro Core Linux, 12MB Linux GUI Desktop, Live, Frugal, Extendable

Welcome - Tiny Core Linux

Lambda Calculus

One of the most minimal Turing-complete languages, lambda calculus is a tiny programming language in which the only objects are...

An official atlas of North Korea

A North Korean atlas that shows the world from the perspective of one of the most isolated countries on the planet.

Intercellular communication in the brain through a dendritic nanotubular network

Intercellular nanotubular networks mediate material exchange, but their existence in neurons remains to be explored in detail. We identified long, thin dendritic filopodia forming direct dendrite–dendrite nanotubes (DNTs) in mammalian cortex. Super-...