This repository contains the code release accompanying the work: Multi-Task GRPO: Reliable LLM Reasoning Across Tasks.

---

Repository Structure

verl/                 Core library (adapted for this work from verl)
scripts/exp-1/        Scripts to reproduce Experiment 1
scripts/exp-2/        Scripts to reproduce Experiment 2
installation.sh       Lightweight installation (assumes CUDA available)
installation_micromamba_cuda.sh  Full installation (local CUDA + micromamba)

---

Installation

We provide two installation options.

Option A — Full install (Recommended)

This option creates an isolated environment, installing micromamba, a local CUDA 12.2 toolkit, and all Python dependencies.

bash installation_micromamba_cuda.sh

Requirements: A Linux machine with an NVIDIA GPU and driver.

Option B — Lightweight install

If you already have a working CUDA and Python environment, you can install the dependencies directly:

bash installation.sh

---

Running Experiments

All reproduction scripts are located in the scripts/ directory.

Example usage:

bash scripts/exp-1/run_mt_grpo_0.2.sh

---

Logging (Weights & Biases)

By default, the provided scripts may attempt to use Weights & Biases (wandb). To run experiments without external logging, set:

export WANDB_MODE=offline

If you prefer to log runs to your own account during the review process, set your API key and entity:

export WANDB_API_KEY=your_key_here
export WANDB_ENTITY=your_entity_here

Notes

• Base Library: This codebase builds upon the open-source VeRL framework. Files modified for this submission are included in the verl/ directory.
• Model Downloads: Pretrained models (e.g., Qwen2.5) are configured to download automatically from Hugging Face.

---