RNA-IFlow and RNA-IFlow-RL

Paper: RNA Design via Conditioned Flow Matching and Finite-Policy Reinforcement Learning
Authors: Zefeng Lin, Xianyong Fang, Tianfan Fu, Xiaohua Xu
Code: https://github.com/John-Lin98/RNA-IFlow
arXiv: arXiv:2609.36885 (v1).

Both exports were independently downloaded from this repository at revision fcf6669251e95aca4e3351719579c1f1d8d2bf76; their SHA256 values matched the upload sources, and each downloaded export passed a CPU single-target inference smoke test. The code used for this validation is GitHub commit 85f5bd4e7adf96ccabd67a54cf7fe6df57dde46b.

RNA-IFlow generates RNA sequences conditioned on a target secondary structure with Dirichlet flow matching. RNA-IFlow-RL maps the learned flow to a pairing-preserving finite policy and refines it with thermodynamic feedback.

Checkpoints

Directory	Paper role	Complete inference weight SHA256	Source checkpoint SHA256
RNA-IFlow/	Supervised flow model	4f33b51d28aa0fa7fab7a27cd6b0422c2eed9654f204bcedb80a771851c5a151	8e221cc4c4382a5421890724c0548cf64492fe1be4556d498878e86e83056bc5
RNA-IFlow-RL/	C3+D5 U2442 final model	8a8dcf74014be2e3ab9571a19facebaad639ca4fd6574bc97d7bdf5d9ffe9f9d	198fd79e7680f4b01e758f063aadab00c0d4e6709ac4d2a220249a267a70ebe8

Each directory contains model.safetensors, config.json, and export_manifest.json. The supervised model also includes inference_config.json. These are complete inference exports with the backbone weights embedded. They do not contain optimizer or RNG state and cannot resume training exactly. No tokenizer, vocabulary, or normalization file is needed by the repository's structure-conditioned inference path.

The HF arxiv-v1 revision is frozen to the GitHub arxiv-v1 tag and the paper's arXiv v1.

Inference

Install the dependencies from the code repository, download this model repository, and run:

python scripts/infer_flow.py --model /path/to/RNA-IFlow --structure '(((...)))...' --candidates 8
python scripts/infer.py --model /path/to/RNA-IFlow-RL --structure '(((...)))...' --candidates 8

The first command uses the supervised 50-step flow. The second uses the refined finite-policy sampler. These are single-target usage examples, not the paper's full benchmark protocol. Both exports are verified against their original tensors and fixed CPU outputs; both were freshly downloaded from this repository and used for CPU inference smoke tests.

Paper benchmark

The following RNA-IFlow-RL values are the displayed Pass@1 and Pass@8 results from Table 1 of the submitted paper, not a benchmark rerun during this model upload.

Dataset	Pass@1	Pass@8
Eterna100-v2	0.5400	0.6500
Eterna100	0.5067	0.6167
Rfam-27	0.8519	0.8765

See the paper and the evaluation documentation for conditions, metrics, and limitations.

Provenance, citation, and license

Both exports include weights from LLM-EDA/RNAErnie, whose model card identifies Apache-2.0 as its license. See the included third-party notice and upstream license copy. The code repository is Apache-2.0. The authors release their original contributions and the combined inference exports under Apache-2.0. The upstream backbone retains its attribution and license notice; see THIRD_PARTY_NOTICES.md and third_party/RNAErnie_Apache-2.0.txt.

Cite: Zefeng Lin, Xianyong Fang, Tianfan Fu, and Xiaohua Xu, RNA Design via Conditioned Flow Matching and Finite-Policy Reinforcement Learning, arXiv:2609.36885 (2026), https://doi.org/10.48550/arXiv.2609.36885.