Trending storyDeveloping

Carnegie Mellon paper proposes one likelihood-based rule for all LLM training stages

1 article1 sourcesince Oct 30Last article 1h ago · Oct 30, 5:52 PM ET

Overview

AISummary of 1 article

Carnegie Mellon researchers publish a paper recasting MaxRL, a reinforcement learning method, as likelihood training on the model's own correct answers.

The authors say this lets one trainer mix pretraining text, curated data, and RL samples in any proportion over time.

In one reported result, a 0.76B model trained by blending 5% midtraining data into math RL scored 32.69% on MBPP+, compared with 29.81% for staged training.

Written by AI from the articles below · updated Oct 30, 6:19 PM ET

Check the sources:

Latest developmentOct 30, 5:52 PM ET

CMU paper unifies LLM training stages under one likelihood-based rule

Article timeline

The articles in this story. Times are ET.

Oct 30

5:52 PM
Rohan Paul@rohanpaul_aiX
CMU paper unifies LLM training stages under one likelihood-based rule
AICarnegie Mellon researchers recast MaxRL, a reinforcement learning method, as likelihood training on the model's own correct answers. This lets one trainer mix pretraining text, curated data, and RL samples in any proportion over time. On a 0.76B model, blending 5% midtraining data into math RL reached 32.69% on MBPP+, versus 29.81% for staged training.

Heat Trend

Not enough continuous observations to show a trend yet.