Technical Debt in In-Context Learning: Diminishing Efficiency in Long Context

Taejong Joo, Diego Klabjan · 2025

Read Technical Debt in In-Context Learning: Diminishing Efficiency in Long Context by Taejong Joo, Diego Klabjan (2025) 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-2502-04580.txt.

Description

Taejong Joo and Diego Klabjan wrote this 2025 paper on how well large neural networks learn from examples placed in the prompt. Such networks can pick up a new task from a few shown cases, with no retraining. Some hoped this skill could beat models built for one task.

The authors set up a test world where each prompt defines a fresh fitting task. The hidden rule behind each task must be guessed at two levels. They compared the network's learning speed with the best possible method and with other sound methods. Early on, the network matched the best method's pace. The paper's title hints that this pace fades as prompts grow long.

Catalog record

CREATORTaejong Joo, Diego Klabjan
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE62 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2502-04580

Browse the full library · About the Tower of Babel