Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Agnostics: Learning to Code in Any Programming Language via Reinforcement with a Universal Learning Environment

Aleksander Boruch-Gruszecki, Yangtian Zi, Zixuan Wu, Tejas Oberoi, Carolyn Jane Anderson, Joydeep Biswas, Arjun Guha

TL;DR

Agnostics introduces Agnostics, a language-agnostic post-training pipeline that eliminates per-language engineering, and releases the language-agnostic training datasets, making RL post-training in any programming language as simple as editing a short YAML file.

Abstract

Large language models (LLMs) already excel at writing code in high-resource languages such as Python and JavaScript, yet stumble on low-resource languages that remain essential to science and engineering. Besides the obvious shortage of pre-training data, post-training itself is a bottleneck: every new language seems to require new datasets, test harnesses, and reinforcement-learning (RL) infrastructure. We introduce Agnostics, a language-agnostic post-training pipeline that eliminates this per-language engineering. The key idea is to judge code solely by its externally observable behavior, so a single verifier can test solutions written in any language. Concretely, we (i) use an LLM to rewrite existing unit-test datasets into an I/O format, (ii) supply a short configuration that tells the verifier how to compile and run a target language, and (iii) apply reinforcement learning with verifiable rewards (RLVR) in a robust code execution environment. Applied to five low-resource languages--Lua, Julia, R, OCaml, and Fortran--Agnostics (1) improves Qwen-3 4B to performance that rivals other 16B-70B open-weight models; (2) scales cleanly to larger and diverse model families (Qwen-3 8B, DeepSeek Coder 6.7B Instruct, Phi 4 Mini); and (3) for ${\le} 16$B parameter models, sets new state-of-the-art pass@1 results on MultiPL-E and a new multi-language version of LiveCodeBench that we introduce. We release the language-agnostic training datasets (Ag-MBPP-X, Ag-Codeforces-X, Ag-LiveCodeBench-X), training code, and ready-to-use configurations, making RL post-training in any programming language as simple as editing a short YAML file.

Agnostics: Learning to Code in Any Programming Language via Reinforcement with a Universal Learning Environment

TL;DR

Agnostics introduces Agnostics, a language-agnostic post-training pipeline that eliminates per-language engineering, and releases the language-agnostic training datasets, making RL post-training in any programming language as simple as editing a short YAML file.

Abstract

Large language models (LLMs) already excel at writing code in high-resource languages such as Python and JavaScript, yet stumble on low-resource languages that remain essential to science and engineering. Besides the obvious shortage of pre-training data, post-training itself is a bottleneck: every new language seems to require new datasets, test harnesses, and reinforcement-learning (RL) infrastructure. We introduce Agnostics, a language-agnostic post-training pipeline that eliminates this per-language engineering. The key idea is to judge code solely by its externally observable behavior, so a single verifier can test solutions written in any language. Concretely, we (i) use an LLM to rewrite existing unit-test datasets into an I/O format, (ii) supply a short configuration that tells the verifier how to compile and run a target language, and (iii) apply reinforcement learning with verifiable rewards (RLVR) in a robust code execution environment. Applied to five low-resource languages--Lua, Julia, R, OCaml, and Fortran--Agnostics (1) improves Qwen-3 4B to performance that rivals other 16B-70B open-weight models; (2) scales cleanly to larger and diverse model families (Qwen-3 8B, DeepSeek Coder 6.7B Instruct, Phi 4 Mini); and (3) for B parameter models, sets new state-of-the-art pass@1 results on MultiPL-E and a new multi-language version of LiveCodeBench that we introduce. We release the language-agnostic training datasets (Ag-MBPP-X, Ag-Codeforces-X, Ag-LiveCodeBench-X), training code, and ready-to-use configurations, making RL post-training in any programming language as simple as editing a short YAML file.

Paper Structure

This paper contains 41 sections, 2 equations, 19 figures, 8 tables.

Table of Contents

  1. Introduction
  2. Contributions
  3. Background and Related Work
  4. Data Scarcity for Low-Resource Languages
  5. Synthetic Data for Low-Resource Languages
  6. Reinforcement Learning on Coding Tasks
  7. The Agnostics Approach
  8. Dataset Preparation
  9. Programming Language Preparation
  10. Trainer and Code Execution
  11. Trainer
  12. Code Execution
  13. Implementation
  14. Hyperparameters
  15. Evaluation
  16. ...and 26 more sections

Figures (19)

  • Figure 1: Overview: Agnostics Data Preparation and Training.(1) We reformulate existing coding datasets to our format. (2) We adapt the language-agnostic datasets to a particular programming language. (3, 4) We reinforce coding via Group Relative Policy Optimization shao2024deepseek-ai2025, verifying the programs in our code execution sandbox.
  • Figure 2: High-resource (left) and low-resource (right) languages in the Stack v2.
  • Figure 3: For dataset preparation, we use an LLM to reformulate fine-tuning datasets with language-specific prompts and tests (above) into equivalent language-agnostic programming tasks.
  • Figure 4: An Agnostics configuration snippet for R (slightly rephrased for presentation).
  • Figure 5: Group batch pass@1, Qwen3-4B-CF-X.
  • ...and 14 more figures