This is the official code of the paper "Looking Backward: Retrospective Backward Synthesis for Goal-Conditioned GFlowNets". we propose a novel method called Retrospective Backward Synthesis (RBS) to synthesize new backward trajectories in goal-conditioned GFlowNets to enrich training trajectories with enhanced quality and diversity, thereby introducing copious learnable signals for effectively tackling the sparse reward problem.

conda env create -f environment.yml
conda activate gflownet

• Run RBS GFlowNets on the Grid tasks using

cd grid
python main.py --method db_gfn --batch_size 256 --backward 1 --horizon 32 --seed 1 --exp_name <name> --device cuda:0 --wdb

horizon=32, 64, 128 denotes small, medium, and large grid maps respectively.

• Run RBS GFlowNets on the bit sequence generation tasks using

cd gfn_bit
python main.py --seq_max_len 40 --num_bits 4 --fl 1 --device cuda:0 --wdb 1 --exp_name <name> --batch_size 256

• For small bit size: seq_max_len=40, num_bits=2
• For medium bit size: seq_max_len=60, num_bits=3
• For large bit size: seq_max_len=100, num_bits=5

@inproceedings{
he2025looking,
title={Looking Backward: Retrospective Backward Synthesis for Goal-Conditioned {GF}lowNets},
author={Haoran He and Can Chang and Huazhe Xu and Ling Pan},
booktitle={The Thirteenth International Conference on Learning Representations},
year={2025},
}