• [2026-03-05] Training code of Temporal Consistency Reinforcement Released!
• [2026-01-27] dLLM-MidTruth is accepted by ICLR 2026.
• [2025-08-21] Code of Temporal Self-Consistency Voting and Evaluation Released!
• [2025-08-13] Paper Released!

Diffusion large language models (dLLMs) generate text through iterative denoising, yet current decoding strategies discard rich intermediate predictions in favor of the final output. Our work here reveals a critical phenomenon, temporal oscillation, where correct answers often emerge in the middle process, but are overwritten in later denoising steps. To address this issue, we introduce two complementary methods that exploit temporal consistency:

• Temporal Self-Consistency Voting, a training-free, test-time decoding strategy that aggregates predictions across denoising steps to select the most consistent output;
• Temporal Consistency Reinforcement, a post-training method that uses Temporal Semantic Entropy (TSE), a measure of semantic stability across intermediate predictions, as a reward signal to encourage stable generations.

To setup the environment, run:

conda env create -f env.yml conda activate tiaf

Then install the dependencies:

bash install.sh

This will install required packages and set up the output directories. You also need to:

1. Download the base model LLaDA-8B-Instruct to pretrained_weights/GSAI-ML/LLaDA-8B-Instruct
2. Download the required datasets (they will be loaded from dataset/ directory):
   • openai/gsm8k, Jiayi-Pan/Countdown-Tasks-3to4, ankner/math-500, ChilleD/SVAMP, simplescaling/s1K

We provide a simple plug-and-play generate script to perform temporal self-consistency voting in eval/generate.py.

You can enable the voting mechanism by passing the --enable_vote flag together with both --tokenizer and --parse_answer_func. Optionally, you may also specify --vote_method (fixed, linear, or exp) and, if using the exp method, control the weighting strength with --alpha.

We provide training scripts for three stages:

Fine-tune the base dLLM model on the s1K dataset with LoRA adapters:

bash scripts/train_sft.sh

Key hyperparameters (configurable in the script):

Train with GRPO using only the Temporal Semantic Entropy (TSE) as reward signal — no ground-truth labels needed:

bash scripts/train_rft_temporal.sh

Key hyperparameters:

Train with GRPO using both TSE and ground-truth accuracy, combined via a proper scoring rule:

bash scripts/train_rft_temporal_acc.sh

Additional hyperparameters (on top of the temporal-only setting):

You can easily evaluate our method by running the following command (using GSM8K with a generation length of 128 as an example):

cd eval CUDA_VISIBLE_DEVICES=0,1,2,3 torchrun \ --nproc_per_node 4 \ --master_port 29173 \ eval.py \ --dataset gsm8k \ --batch_size 8 \ --gen_length 128 \ --diffusion_steps 64 \ --output_dir "outputs/LLaDA-8B-Instruct/gsm8k_gen_128_steps_64_temp_0_vote" \ --model_path <your_model_path> \ --enable_vote \ --vote_method exp \ --alpha 5.0 

For convenience, we also provide a wrapper script eval/run_eval.sh that allows easier evaluation management. You can modify the parameters in this script to test different datasets, and voting strategies.

To compute accuracy, run get_acc.py after the evaluation is done.

• source code of temporal self-consistency voting and evaluation
• source code of temporal consistency reinforcement (SFT + RFT training)

For academic use, this project is licensed under the 2-clause BSD License. For commercial use, please contact Chunhua Shen.

If you find this work useful, please consider citing:

@article{wang2025temporaldynamics, title={Time Is a Feature: Exploiting Temporal Dynamics in Diffusion Language Models}, author={Wen, Wang and Bozhen, Fang and Chenchen, Jing and Yongliang, Shen and Yangyi, Shen and Qiuyu, Wang and Hao, Ouyang and Hao, Chen and Chunhua, Shen}, journal={arXiv preprint arXiv:2508.09138}, year={2025} }