In this comparison, DiscoveryRSI-27B achieves the highest actionable-reference rate at 6.3%, followed by DeepSeek (5.4%), GLM (2.8%), GPT (2.5%), and Claude (2.3%).
PhAI Labs is running a technical livestream that shows a scientific agent move from an open question to a testable research plan, then learn from evidence, critique, and revision.
Click the animation to jump directly to the live website.
The arena makes the process inspectable: model comparisons, research domains, candidate hypotheses, compiled experiments, and the growing hypothesis tree update as the run progresses.
1. A new paradigm for scientific discovery. DiscoveryRSI turns research into a recursive learning process centered on hypotheses. Each proposal is a versioned research object combining an evidence-grounded claim, an investigation plan, explicit decision criteria, and alternative explanations. Multi-agent evolutionary search develops a research graph, selected computational investigations are compiled into executable Harbor tasks, and the evidence, decisions, and artifacts behind each revision are preserved as reusable experience.
2. A 27B model trained from research experience. DiscoveryRSI-27B learns through Hypothesis Credit Optimization, which links assessments of scientific quality to documented contributions such as finding evidence, adding a comparator, clarifying a measurement, or repairing a falsification rule. In matched runs, DiscoveryRSI-27B selected 16.7% more unique hypotheses and successfully compiled 50% more Harbor tasks than its base model.
3. DFM-Bench, a benchmark for cumulative research progress. DFM-Bench evaluates how research develops across a run in life science, with a battery-science extension. It scores individual hypotheses, revisions against their actual predecessors, ordered development paths, complete research runs, and deduplicated scientific contributions. This separates a good proposal from a productive research process, and an executable package from a scientifically validated result.
DiscoveryRSI couples multi-agent hypothesis development, contribution attribution, recursive model and workflow improvement, and process-level evaluation.
| Type | Name | Link |
|---|---|---|
| 📄 Paper | DiscoveryRSI | arXiv:2610.23839 |
| 🤖 Model | DiscoveryRSI-27B | 🤗 Hugging Face |
| 📊 Dataset | Compiled hypothesis task examples | 🤗 Hugging Face |
| 🧪 Benchmark | DFM-Bench | GitHub |
| 📺 Demo | DiscoveryRSI live arena | Open arena |
Clone the repository and install the public harness:
git clone https://github.com/Gen-Verse/DiscoveryRSI.git
cd DiscoveryRSI
python3 -m venv .venv
source .venv/bin/activate
python -m pip install -e .Run the deterministic local demo and tests:
discovery-rsi demo --iterations 2
pytestThe released checkpoint is Gen-Verse/DiscoveryRSI-27B. Serve it behind an OpenAI-compatible endpoint, then point the harness at that endpoint:
export MODEL_ID=Gen-Verse/DiscoveryRSI-27B
vllm serve "$MODEL_ID" --served-model-name discoveryrsi-27bIn a second terminal:
export DISCOVERY_RSI_BASE_URL=http://localhost:8000/v1
export DISCOVERY_RSI_MODEL=discoveryrsi-27b
export DISCOVERY_RSI_API_KEY=local
discovery-rsi campaign \
--adapter openai-compatible \
--manifest configs/research_manifest.json \
--roles plugins/discovery-rsi/roles \
--iterations 3 \
--replay episodes.jsonlThe harness can also run with Codex, Claude Code, another OpenAI-compatible endpoint, or any implementation of the ModelAdapter protocol.
The public harness includes the portable agentic core, provider adapters, role prompts, tool registry, replay store, campaign manifest, and domain adapters.
It supports research workflows that:
- generate falsifiable, data-aware study plans;
- add controls, measurements, alternatives, and evidence boundaries;
- preserve parent links and revision lineage;
- identify whether an idea was evolved, repaired, crossed over, or restarted;
- replay tool failures, repeated rediscovery, and lost context; and
- evaluate complete research paths instead of isolated text generations.
The system was developed across clinical and population health, therapeutic-target prioritization, disease mechanisms, and population multi-omics.
DiscoveryRSI focuses on life-science research, with public workflows and adapters for:
- clinical and population health;
- therapeutic-target prioritization;
- disease mechanisms;
- population multi-omics; and
- GEO-derived pathway and target records.
Bring only data you are authorized to use. No patient rows, HCC material, UK Biobank material, private campaign traces, credentials, or laboratory infrastructure are bundled.
DFM-Bench evaluates cumulative research progress at five levels: Hypothesis, Transition, Trajectory, Campaign, and Discovery. Its primary setting is life science, spanning clinical and population health, therapeutic-target prioritization, disease mechanisms, and population multi-omics.
A battery-science extension applies the same evaluation structure with domain-specific rubrics for electrochemistry, materials degradation, sensing, and state estimation. Completion, compilation, and readiness are reported separately from scientific quality; proposed contributions remain subject to independent verification.
See the benchmark repository for the overview and the evaluation guide for practical setup and usage.
DiscoveryRSI can produce plausible but unsupported claims, omit relevant evidence, misread sources, or propose infeasible experiments. Generated hypotheses require expert review, source checking, and independent verification. The system is not a medical device and must not be used to make clinical or other high-stakes decisions without qualified oversight.
If you use DiscoveryRSI or DiscoveryRSI-27B in your research, please cite our paper:
@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}
}Project-owned code is licensed under Apache-2.0. See LICENSE and THIRD_PARTY_NOTICES.md. The model license, base-model terms, and dataset attribution requirements are specified separately in the model card.