How2Everything : Mining the Web for How-To Procedures to Evaluate and Improve LLMs

2026

Read How2Everything : Mining the Web for How-To Procedures to Evaluate and Improve LLMs (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-08808.txt.

Description

Step-by-step how-to procedures are a core language-model skill, yet measuring procedural validity at scale on real tasks remains understudied. Yapei Chang and three co-authors introduce How2Everything, a framework for evaluating and improving goal-conditioned procedure generation. Their arXiv paper mines 351 thousand procedures from 980 thousand web pages across fourteen topics.

From that pool, a seven-thousand-example benchmark balances topics for evaluation, while a scoring protocol uses a language-model judge to flag critical failures that would block the goal. Distilling a frontier judge into an open eight-billion-parameter model reaches 80.5 percent agreement with humans. Reinforcement learning with the scorer as reward improves benchmark scores by over ten points across three models without regressions elsewhere. Readers get a closed loop from web mining to evaluated improvement.

Catalog record

YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE142 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2602-08808

Browse the full library · About the Tower of Babel