Bridging the Knowledge Void: Inference-time Acquisition of Unfamiliar Programming Languages for Coding Tasks
Chen Shen, Wei Cheng, Jingyue Yang, Huan Zhang, Yuhan Wu, Wei Hu · 2026
Read Bridging the Knowledge Void: Inference-time Acquisition of Unfamiliar Programming Languages for Coding Tasks by Chen Shen, Wei Cheng, Jingyue Yang, Huan Zhang, Yuhan Wu, Wei Hu (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-2602-06976.txt.
Description
Chen Shen and five co-authors ask how language models can code in a programming language they have never seen. Coding skill in models usually mirrors pretraining data, and collapses on unfamiliar tongues. Rather than costly fine-tuning, the team studies inference-time language acquisition, where a model masters a new tongue by interacting with limited outside resources during the task itself.
Their framework, ILA-agent, gives models learning moves that mimic human learning: exploring docs, applying knowledge, and verifying results by running code. To test it fully they built Cangjie-bench, a multi-task benchmark for the novel statically typed language Cangjie. Across code making, translation, and repair, ILA-agent clearly beat retrieval-augmented rivals, and a look at agent paths reveals new behavior patterns alongside lasting gaps.
Catalog record
| CREATOR | Chen Shen, Wei Cheng, Jingyue Yang, Huan Zhang, Yuhan Wu, Wei Hu |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 42 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2602-06976