What if AI stopped summarizing scientific knowledge—and started helping science become cumulative?
We built increasingly powerful machines for generating text.
We built search engines for finding papers.
We built RAG systems for retrieving evidence.
We built AI assistants for summarizing literature.
But there is a deeper problem.
Science is not simply a collection of papers.
It is a constantly evolving structure of claims, evidence, contradictions, failed replications, hidden conditions, mechanisms, hypotheses, and experiments.
And most AI systems still treat it as a document-retrieval problem.
That is the problem ZHANG is designed to attack.
From Literature Search to Scientific Reasoning
A conventional literature assistant might answer:
“Here are 30 papers related to your question.”
A better system might summarize them.
ZHANG asks a different question:
What can we actually infer from the structure of the evidence?
Its conceptual pipeline is:
Research Question
↓
Evidence
↓
Claims
↓
Knowledge Graph
↓
Contradictions
↓
Reproducibility
↓
Cross-Domain Transfer
↓
Hypotheses
↓
Falsification Experiments
↓
Scientific Action Brief
The objective is not to generate another polished literature review.
The objective is to transform fragmented scientific knowledge into something actionable, inspectable, and falsifiable.
The Core Thesis
The central idea behind ZHANG is simple:
Scientific knowledge becomes powerful when relationships between evidence become explicit.
Consider two papers.
Paper A reports:
Method X improves outcome Y.
Paper B reports:
Method X does not improve outcome Y.
A conventional AI system may average the conclusions or summarize both.
ZHANG treats the disagreement as a potential source of knowledge.
It asks:
What changed?
Population?
Dataset?
Protocol?
Temperature?
Measurement?
Sample size?
Time horizon?
Intervention?
Experimental conditions?
The contradiction may not mean that one paper is wrong.
It may reveal a moderator.
And that moderator may be more scientifically valuable than either conclusion alone.
Contradictions Are Not Noise
Modern AI systems are optimized heavily for coherence.
But scientific progress often begins with incoherence.
A contradiction can reveal:
- a hidden variable
- a boundary condition
- a measurement artifact
- a population difference
- an experimental limitation
- an incomplete theory
- or an entirely new research direction
So ZHANG treats disagreement as a first-class object.
Not:
“Which paper is correct?”
But:
“Under what conditions can both observations be true?”
That is a fundamentally different reasoning strategy.
The Evidence Ledger
One of the directions we consider essential for ZHANG is an explicit Evidence Ledger.
Every important claim should be traceable:
CLAIM
│
├── Source papers
├── Supporting evidence
├── Contradicting evidence
├── Evidence type
├── Inference level
├── Reproducibility
├── Transfer assumptions
└── Uncertainty
This creates a crucial property:
Every important conclusion should be able to answer: “Why should I believe this?”
That question becomes increasingly important as AI systems move from generating text to influencing research decisions.
Reproducibility Is Part of Knowledge
A citation alone does not tell us how trustworthy a scientific result is.
Two papers may make similar claims while having radically different:
- methodological transparency
- sample sizes
- protocols
- replication histories
- experimental conditions
- statistical strength
ZHANG therefore treats reproducibility as part of the knowledge structure rather than as an afterthought.
The goal is not to produce a magical “truth score.”
It is to expose the structure of confidence and uncertainty.
The Most Interesting Layer: Cross-Domain Transfer
Scientific breakthroughs frequently emerge when ideas move between disciplines.
A mechanism discovered in one field may illuminate an apparently unrelated problem somewhere else.
Imagine:
Battery degradation
↓
Fatigue mechanics
↓
Crack propagation
↓
Analogical mechanism
↓
New hypothesis
↓
Falsification experiment
This is where ZHANG becomes more than a literature assistant.
It starts approaching a different question:
Can machines systematically discover useful scientific analogies?
If they can, literature stops being merely an archive of what humanity already knows.
It becomes a searchable space of latent possibilities.
From Answer Generation to Falsification
There is another principle we consider critical:
ZHANG should not simply generate confident hypotheses.
It should help scientists try to break them.
A useful output should therefore contain:
Hypothesis
Evidence supporting it
Evidence against it
Assumptions
Uncertainty
Potential confounders
Suggested falsification experiment
The objective is not:
“The AI discovered the truth.”
It is:
“Here is a hypothesis worth testing, and here is how it could fail.”
That distinction matters.
Why ZHANG Is Not Just Another RAG System
RAG primarily answers:
What information is relevant?
ZHANG is designed to ask:
How are the pieces of knowledge related?
That distinction can be expressed as:
| Conventional RAG | ZHANG |
|---|---|
| Retrieve documents | Structure evidence |
| Summarize papers | Model claims |
| Optimize relevance | Analyze relationships |
| Prefer coherent answers | Surface contradictions |
| Generate conclusions | Expose assumptions |
| Cite sources | Track evidence |
| Answer questions | Generate testable hypotheses |
| End with text | Produce an Action Brief |
The difference is not a larger language model.
It is a different representation of knowledge.
The Architecture
ZHANG is intentionally modular:
Retrieval
↓
Ingestion
↓
Extraction
↓
Knowledge Graph
↓
Reasoning
├── Contradiction Detection
├── Reproducibility Analysis
├── Transfer Hypotheses
└── Experiment Generation
↓
Synthesis
↓
Scientific Action Brief
The architecture is designed around one important principle:
The schema is the asset.
Models will change.
Retrieval systems will change.
Data sources will change.
But a robust representation of:
claims → evidence → relationships → uncertainty → hypotheses → experiments
can survive those changes.
That makes the knowledge schema potentially more valuable than any individual model.
What Exists Today?
ZHANG is deliberately transparent about its current stage.
The MVP runs fully offline on a synthetic battery-research corpus.
The current implementation focuses on validating the architecture and reasoning pipeline rather than pretending to already be a production-scale scientific search engine.
That distinction matters.
The next step is connecting the architecture to real scientific ecosystems and evaluating it against real-world research tasks.
The Real Benchmark
The interesting benchmark for ZHANG should not simply be:
“Can the AI summarize 100 papers?”
That problem is largely solved.
A more meaningful benchmark would ask:
Can it identify hidden contradictions?
Can it correctly ground claims?
Can it distinguish evidence from inference?
Can it detect reproducibility weaknesses?
Can it discover meaningful cross-domain analogies?
Can it generate experimentally falsifiable hypotheses?
Can scientists verify its reasoning?
Those are much harder problems.
And potentially much more valuable ones.
Toward an Epistemic Layer for Science
This leads to the larger vision.
Today, scientific infrastructure is optimized around:
papers, journals, citations, databases, and search.
The next layer could be optimized around:
claims, evidence, contradictions, mechanisms, uncertainty, hypotheses, and experiments.
In that world, AI does not replace scientists.
It becomes infrastructure for making scientific reasoning more explicit.
The ultimate architecture could look like:
Literature
↓
Evidence
↓
Knowledge Graph
↓
Machine Reasoning
↓
Hypothesis
↓
Human Scientist
↓
Experiment
↓
New Evidence
↓
Updated Knowledge
↺
That final loop is the real destination.
ZHANG Is Not Trying to Read Science Faster
It is trying to make science more cumulative.
If ZHANG only helps researchers read papers faster, it is a useful research tool.
But if it can reliably connect:
literature → evidence → contradiction → hypothesis → experiment → new evidence
then it becomes something much more ambitious:
A machine-assisted infrastructure for cumulative science.
That is the direction we are exploring.
And perhaps the most important question is no longer:
Can AI understand scientific literature?
It is:
Can AI help humanity discover what the literature does not yet know?
Every Field Has Its Zhang
The name is intentional.
Scientific literature is filled with:
Zhang et al.
The name represents the millions of researchers whose individual contributions collectively form the scientific record.
ZHANG is not named after a country, institution, or individual.
It represents the anonymous pattern of scientific contribution.
Every field has its Zhang.
And perhaps the next generation of scientific AI should learn how to connect them.
Explore the ZHANG architecture and code on GitHub
ZHANG — The Science-to-Action Engine
From papers to evidence.
From evidence to hypotheses.
From hypotheses to experiments.
From experiments to new knowledge.
created by Seyed Alireza Alhosseini Almodarresieh
Top comments (0)