Model: jiangchengsong/MemDuet-S2-7B Source: Original Platform
license, library_name, pipeline_tag, language, base_model, base_model_relation, tags, model_name
| license | library_name | pipeline_tag | language | base_model | base_model_relation | tags | model_name | |||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| apache-2.0 | transformers | text-generation |
|
|
finetune |
|
MemDuet-7B |
MemDuet System 2 — 7B
MemDuet System 2 verifies evidence selections and updates memory for long-context reasoning. It checks the context chunks proposed by System 1, rewrites bounded memory in document order, and generates the final answer. Use this model with the System 1 selector and controller in the MemDuet code.
Base model: Qwen/Qwen2.5-7B-Instruct. The weights in this repository are a standalone BF16 model.
Paper · Code · Project page
Load
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "jiangchengsong/MemDuet-S2-7B"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, dtype="auto")
model.eval()
Use Transformers 5.15.0 with PyTorch 2.13.0, matching the MemDuet inference environment. For the verification and memory-update prompts and the complete controller, use the MemDuet inference implementation.
Use with MemDuet
From the MemDuet code directory, after installing its inference dependencies, download this model to system2:
hf download jiangchengsong/MemDuet-S2-7B --local-dir system2
python -m memduet serve --config configs/inference.example.json --role system2
Follow the code repository instructions to provide system1, wait for the System 2 service to be ready, start the System 1 service, and run evaluation. Retain this model's tokenizer and chat template.
License
Apache License 2.0. The upstream Qwen license is included in LICENSE; source attribution is retained in NOTICE.
Citation
@misc{song2026memduet,
title = {MemDuet: Learning Dual-Process Memory for Long-Context Reasoning},
author = {Song, Jiangcheng and Jing, Haodong and Mu, Yulin and Tao, Lesong and
Wang, Jixin and Cui, Jin and Ren, Tianjing and Xie, Runxuan and
Ma, Yongqiang and Zheng, Nanning},
year = {2026},
eprint = {2610.19928},
archivePrefix = {arXiv},
primaryClass = {cs.CL},
url = {https://arxiv.org/abs/2610.19928}
}