jamesding0302/memgen-checkpoints

MemGen Checkpoints

This repository contains model checkpoints and artifacts for the paper How Well Does Generative Recommendation Generalize?.

Overview

A widely held hypothesis for why generative recommendation (GR) models outperform conventional item ID-based models is that they generalize better. This work provides a systematic way to verify this hypothesis by categorizing data instances into those requiring memorization (reusing item transition patterns observed during training) and those requiring generalization (composing known patterns to predict unseen item transitions). Our study shows that GR models perform better on instances that require generalization, whereas item ID-based models perform better when memorization is more important.

Folder Structure

• SASRec/: SASRec checkpoints (.pth)
• TIGER/: TIGER checkpoints (.pth)
• semantic_ids/: semantic ID files (.sem_ids) used to train the corresponding TIGER checkpoints

All checkpoints are trained with the default configs in the code repo. If you are unsure about model size / architecture, please refer to the configs in the GitHub repository.

Download (example)

You can use the huggingface_hub library to download specific checkpoints and semantic ID files:

from huggingface_hub import hf_hub_download

repo_id = "jamesding0302/memgen-checkpoints"
out_dir = "./downloads"   # change this to desired output path

ckpt_path = hf_hub_download(
    repo_id=repo_id,
    filename="TIGER/TIGER-AmazonReviews2014-category_Sports_and_Outdoors.pth",
    local_dir=out_dir,
)

sem_ids_path = hf_hub_download(
    repo_id=repo_id,
    filename="semantic_ids/AmazonReviews2014-Sports_and_Outdoors_sentence-t5-base_256,256,256,256.sem_ids",
    local_dir=out_dir,
)

Evaluation

You can run fine-grained evaluation using the saved checkpoint_path and sem_ids_path with the scripts provided in the official repository:

CUDA_VISIBLE_DEVICES=0 python mem_gen_evaluation.py \
    --model=TIGER \
    --dataset=AmazonReviews2014 \
    --category=Sports_and_Outdoors \
    --checkpoint_path=path/to/TIGER.pth \
    --sem_ids_path=path/to/semantic_ids.sem_ids \
    --eval=test \
    --save_inference

Citation

@article{ding2026generalize,
  title={How Well Does Generative Recommendation Generalize?},
  author={Ding, Yijie and Guo, Zitian and Li, Jiacheng and Peng, Letian and Shao, Shuai and Shao, Wei and Luo, Xiaoqiang and Simon, Luke and Shang, Jingbo and McAuley, Julian and Hou, Yupeng},
  journal={arXiv preprint arXiv:2603.19809},
  year={2026}
}