This is the official implementation of the paper Jointly Reinforcing Diversity and Quality of Language Model Generations.

DARLING uses the verl (Volcano Engine Reinforcement Learning) framework to jointly optimize for both diversity and quality in language model generations through reinforcement learning.

Create conda environment and install dependencies:

conda create -n verlenv python=3.10
conda activate verlenv

Install PyTorch (tested on CUDA 12.8):

pip install torch==2.7.1 torchvision==0.22.1 torchaudio==2.7.1 --index-url https://download.pytorch.org/whl/cu128

Install verl and other dependencies:

cd verl
pip install -e ./
# This code only uses FSDP. If you need to use Megatron, remove USE_MEGATRON=0
USE_MEGATRON=0 bash scripts/install_vllm_sglang_mcore.sh
pip install vllm==0.11.0
pip install flash-attn --no-build-isolation

To use Wandb for experiment tracking:

export WANDB_API_KEY=<your_api_key>

• Verifiable tasks (math): verl/math_scripts/
• Non-verifiable tasks (creative writing): verl/wildchat_scripts/

Each directory contains:

• darling.batch: DARLING training with diversity rewards
• grpo_baseline.batch: GRPO baseline for comparison

First, serve the partition classifier used for diversity rewards:

bash verl/serve_classifier.sh <PATH_TO_CLASSIFIER_HF>

This will serve 8 instances of the classifier on ports 8000-8007.

Set the hostname where the classifier is running:

export VLLM_SERVER_HOSTNAME=<your_hostname>

Alternatively, you can manually edit verl/verl/utils/reward_score/partition_reward_vllm_serve.py.

For math tasks (Qwen-4B-Base):

# Edit verl/math_scripts/darling.batch to configure your cluster settings
sbatch verl/math_scripts/darling.batch

For creative writing tasks (Llama-3.1-8B-Instruct):

# Edit verl/wildchat_scripts/darling.batch to configure your cluster settings
sbatch verl/wildchat_scripts/darling.batch

Hyperparameters can be configured by editing the script variables or passing command-line arguments:

Key Parameters:

• B: Training batch size (e.g., 256 for math, 64 for wildchat)
• N: Number of samples per prompt (default: 8)
• L: Maximum response length (e.g., 8192 for math, 1024 for wildchat)
• actor_rollout_ref.actor.optim.lr: Learning rate (default: 1e-6)
• actor_rollout_ref.rollout.temperature: Sampling temperature
• trainer.total_epochs: Total training epochs

For the full list of available hyperparameters, see the training scripts in verl/math_scripts/ and verl/wildchat_scripts/.

The evals/ directory contains benchmarks for evaluating model outputs:

Evaluates mathematical reasoning on standard benchmarks. See evals/math_evaluation/README.md for setup and usage.

Evaluates the diversity and novelty of model generations. This benchmark:

• Generates multiple responses from models
• Groups semantically similar responses using partitioning
• Scores response quality
• Provides a diversity-quality tradeoff analysis

See evals/novelty-bench/README.md for details and project webpage.

Evaluates creative writing capabilities using the EQ-Bench v3 system with hybrid rubric and Elo scoring. See evals/creative-writing-bench/README.md for details.

This project is licensed under the MIT License - see the LICENSE file for details.

If you find DARLING useful, please consider citing:

@article{tianjian2025jointlyreinforcingdiversityquality,
	title        = {Jointly Reinforcing Diversity and Quality in Language Model Generations},
	author       = {Tianjian Li and Yiming Zhang and Ping Yu and Swarnadeep Saha and Daniel Khashabi and Jason Weston and Jack Lanchantin and Tianlu Wang},
	year         = 2025,
	journal      = {arXiv preprint arXiv:2509.02534},
	url          = {https://arxiv.org/abs/2509.02534},
	eprint       = {2509.02534},
	archiveprefix = {arXiv},
	primaryclass = {cs.CL},
}