DiscoveryRSI-27B

🎯 About This Repository

This repository contains the DiscoveryRSI-27B model release for DiscoveryRSI, a research system that recursively develops, evaluates, revises, and attributes scientific hypotheses.

DiscoveryRSI-27B is an instruction-finetuned 27B language model intended for research ideation and evidence-grounded planning. It helps turn a research question and available evidence into an explicit hypothesis, investigation plan, alternatives, decision criteria, and falsification conditions.

🌟 Introduction

DiscoveryRSI treats a hypothesis as a versioned research object rather than a finished answer. The model is designed to support a loop of:

• 🔎 Explore: identify valuable questions and candidate explanations
• 🧪 Evaluate: propose tests, comparators, evidence, and decision criteria
• 🛠️ Revise: repair weak designs and incorporate criticism
• 🧬 Attribute: record which research behaviors improved a hypothesis
• 🔁 Learn: use replay and attribution to improve future research paths

The model works with the public DiscoveryRSI harness, role prompts, tool registry, replay store, and domain adapters. The public system supports life-science research workflows including GEO-derived pathway analysis and therapeutic-target prioritization.

📦 Resources

Type	Name	Link
📄 Paper	DiscoveryRSI	arXiv:2610.23839
🤖 Model	DiscoveryRSI-27B	🤗 Hugging Face
🧰 Code	DiscoveryRSI harness	GitHub
📚 Collection	DiscoveryRSI models	🤗 Hugging Face
📺 Demo	DiscoveryRSI live stream	Open arena

📋 Model Details

Property	Value
Model name	DiscoveryRSI-27B
Model type	Instruction-finetuned causal language model
Parameter scale	27B
Base model	Qwen/Qwen3.8-27B
Weight precision	BF16
Released checkpoint	Round 2, step 150
Primary language	English
Primary use	Scientific hypothesis development and research planning
Library	Transformers

This release contains the completed round-2 checkpoint, trained for 150 steps, with BF16 weights, model configuration, tokenizer, chat template, and processor files. See release_manifest.json for per-file sizes and SHA-256 checksums. The visual and MTP components are preserved from the base model; the DiscoveryRSI training and evaluation described here concern textual scientific-research workflows. Private campaign traces, credentials, patient rows, restricted datasets, and laboratory infrastructure are not included.

🚀 Intended Use

DiscoveryRSI-27B is intended for:

• research ideation and hypothesis generation;
• critique and revision of proposed experiments;
• evidence-grounded research planning;
• explicit alternatives, comparators, and falsification criteria; and
• use through the DiscoveryRSI agentic harness.

The model is not a medical device. Do not use it to make clinical, diagnostic, treatment, or other high-stakes decisions. Generated hypotheses require expert review and independent verification.

💻 Usage

The weights can be loaded directly from this repository. Use a recent Transformers release supporting the exported Qwen3_5ForConditionalGeneration architecture. Install the loading dependencies with:

pip install -U torch transformers accelerate jinja2

from transformers import AutoTokenizer, Qwen3_5ForConditionalGeneration

model_id = "Gen-Verse/DiscoveryRSI-27B"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = Qwen3_5ForConditionalGeneration.from_pretrained(
    model_id, device_map="auto", dtype="auto"
)

messages = [{
    "role": "user",
    "content": "Develop a testable research hypothesis about a disease pathway.",
}]
inputs = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    enable_thinking=False,
    tokenize=True,
    return_dict=True,
    return_tensors="pt",
).to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(
    outputs[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True
))

For multi-step research workflows, follow the DiscoveryRSI quick start and configure a provider adapter for the checkpoint.

🧪 Training and Evaluation

DiscoveryRSI uses hypothesis development, criticism, revision, contribution attribution, harness replay, and DFM-Bench-style process evaluation. Release documentation should report the base-model identity, training-data provenance, instruction format, evaluation configuration, and known limitations. Results should be interpreted as research-process evidence rather than proof that a generated hypothesis is correct or clinically useful.

⚠️ Limitations and Responsible Use

The model can produce plausible but unsupported claims, omit relevant evidence, misread sources, or propose infeasible experiments. It may reflect biases and errors in its base model and instruction-tuning data. Users should preserve source provenance, check citations, inspect alternatives and falsification criteria, and have a qualified researcher review outputs before acting on them.

📝 Citation

If you use DiscoveryRSI-27B or the accompanying harness in your research, please cite our paper:

DiscoveryRSI: Recursively Learning to Discover Valuable Hypotheses

@article{zhong2026discoveryrsi,
  title={DiscoveryRSI: Recursively Learning to Discover Valuable Hypotheses},
  author={Zhong, Jianyuan and Wu, Chenwei and Yu, Zhaochen and Jiang, Xixian and Li, Xuewei and Xiao, Hao and Dai, Xilin and Wang, Junyang and Zou, Jiaru and Cui, Zhiqing and Zheng, Boyuan and Xiong, Jing and Zhang, Zhen and Gao, Qiang and Morley, Michael and Nakayama, Luis and Xu, Qiang and Shen, Liyue and Gui, Tao and Celi, Leo Anthony and Zou, James and Yin, Zhenfei and Wu, Yingcheng and Yang, Ling},
  journal={arXiv preprint arXiv:2610.23839},
  year={2026}
}

See the repository citation section for the maintained technical-report and model links.

📄 License

The released model files are provided under the included Apache-2.0 license, inherited from the Qwen3.8-27B base-model release. Referenced publications and external research datasets retain their own licenses and access terms.