Training Language Models for Grounded Scientific Reasoning
Cross-domain "polymathic" insights without hallucinated citations
BioReasoner is not a theoretical proposal—it is an active, evolving project with established momentum.
We are training a 7B language model to perform rigorous scientific reasoning across domains, generating novel hypotheses that are:
- Grounded in real literature (no hallucinated citations)
- Transparent in reasoning (explicit
<think>blocks) - Cross-domain ("polymathic" bridges between disparate fields)
Our approach uses Direct Preference Optimization (DPO) with a unique "Scientific Tribunal" evaluation framework where multiple AI judges assess outputs for logical rigor, novelty, and citation accuracy.
| Metric | Result | Notes |
|---|---|---|
| Hallucination Rate | 0% | Zero fabricated citations across 83 test samples |
| Format Adherence | 100% | All outputs include proper <think> reasoning blocks |
| Logic Score | 3.4/5 | Multi-judge consensus |
| Novelty Score | 3.9/5 | Cross-domain hypothesis generation |
| Version | Date | DPO Pairs | Key Achievement |
|---|---|---|---|
| v2.0 | 2026-01-02 | 280 | Zero hallucinations, 20% missing <think> |
| v2.1 | 2026-01-03 | 46 | 100% format adherence via force-prefixing |
- Base Model: Qwen/Qwen2.5-7B-Instruct
- Training Data: 751 paper triplets from 7 Vanderbilt faculty
- Method: LoRA (r=16, alpha=32)
- Task Types:
critique_extend,method_bridge,gap_analysis
- Chosen Responses: Claude Opus 4.5 generated critiques
- Rejected Responses: Student model outputs with lower scores
- Beta (KL penalty): 0.1
- Key Innovation: Force-prefix
<think>to eliminate format regression
A multi-judge evaluation framework where:
- Lead Judge (Claude Opus 4.5) scores logic, novelty, citation accuracy
- Second Judge (Codex) provides independent verification
- Auto-Filter catches format violations and low-quality outputs
┌─────────────────────────────────────────────────────────────┐
│ SCIENTIFIC TRIBUNAL │
├─────────────────────────────────────────────────────────────┤
│ Input: Model-generated scientific critique │
│ │
│ ┌─────────────┐ ┌─────────────┐ ┌─────────────┐ │
│ │ Auto-Filter │→ │ Lead Judge │→ │Second Judge │ │
│ │ (Format) │ │ (Claude) │ │ (Codex) │ │
│ └─────────────┘ └─────────────┘ └─────────────┘ │
│ ↓ │
│ Final Verdict │
│ (PASS/FAIL + Detailed Rubric) │
└─────────────────────────────────────────────────────────────┘
We are building an unprecedented cross-domain knowledge base spanning 35+ theoretical domains and 15,294 papers:
|
AI/ML Methods (13 domains)
|
Theoretical Foundations (6 domains)
|
Interdisciplinary Bridges (8 domains)
|
Current Status: 15,294 papers harvested | 73.3% with abstracts | 288+ PDFs downloaded
bioreasoner/
├── scripts/
│ ├── train_dpo_8k.py # DPO training with 4-bit quantization
│ ├── merge_adapter.py # LoRA → full model merging
│ ├── bioreasoner_21_pipeline.py # Best-of-N → Filter → DPO pairs
│ ├── novelty_scoring.py # Rubric-based scoring system
│ └── download_all_pdfs.py # Multi-source PDF acquisition
├── data/
│ ├── train_750.jsonl # SFT training data
│ ├── test_84.jsonl # Held-out evaluation set
│ └── dpo_v21_pairs.jsonl # DPO preference pairs
├── evaluation/
│ ├── tribunal_evaluation.py # Multi-judge evaluation framework
│ └── format_checker.py # Auto-filter for format compliance
└── docs/
├── LAB_NOTEBOOK.md # Detailed training logs
└── POLYMATH_CORPUS.md # Cross-domain paper collection
git clone https://github.com/vanbelkummax/bioreasoner.git
cd bioreasoner
pip install -r requirements.txtfrom transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("vanbelkummax/bioreasoner-2.1")
tokenizer = AutoTokenizer.from_pretrained("vanbelkummax/bioreasoner-2.1")
prompt = """Given these three papers:
1. [Paper on spatial transcriptomics methodology]
2. [Paper on graph neural networks]
3. [Paper on tumor microenvironment]
Generate a novel hypothesis that bridges these domains."""
# Model outputs include <think> blocks for transparent reasoning
output = model.generate(tokenizer(prompt, return_tensors="pt").input_ids)We believe that "Scientific Alignment"—making models that are both novel and rigorously grounded—should be a shared community resource, not a proprietary secret.
We will release:
- Model weights (Hugging Face)
- Scientific Tribunal evaluation scripts
- Synthetic reasoning datasets
- Training pipelines and configurations
Being based at Vanderbilt University Medical Center allows us to validate BioReasoner's outputs against:
- Real-world clinical data
- Expert pathology feedback
- Ongoing research collaborations
This ensures the model's "polymathic" bridges are grounded in biological reality.
- SFT on 751 paper triplets
- DPO v2.0 with 280 preference pairs
- DPO v2.1 with force-prefix (100% format adherence)
- Polymath corpus: 15,294 papers across 25+ domains
- Scale to 15,000+ triplets for cross-domain emergence
- Release model weights on Hugging Face
- Publish Scientific Tribunal framework
- Clinical validation studies
@software{bioreasoner2026,
author = {Van Belkum, Max},
title = {BioReasoner: Training Language Models for Grounded Scientific Reasoning},
year = {2026},
url = {https://github.com/vanbelkummax/bioreasoner},
institution = {Vanderbilt University Medical Center}
}This project is licensed under the Apache 2.0 License - see the LICENSE file for details.
Building AI that reasons like a scientist, not a search engine.