Scientific Discovery Outside the Language Model
A language model can write a convincing paragraph about a compound it has never reasoned about. Discovery needs reasoning that persists, can be checked and improves with every experiment.
A language model can write a convincing paragraph about a compound it has never reasoned about. That is the difficulty with using it for discovery: the output reads like science, and nothing inside the model can tell you which steps of it are true.
The short answer: scientific discovery needs reasoning that persists, that can be checked step by step, and that gets better with every experiment. A language model predicts the next plausible sentence from what it was trained on. GitMir builds the reasoning outside the model — an explicit, inspectable structure of claims, evidence and links between them — and uses the model as one component inside it. That combination produces results a language model on its own cannot: conclusions that hold across thousands of sources at once, every step traceable to where it came from, and the same answer every time the same check is run.
This article is about why discovery is a poor fit for a language model alone, what changes when the reasoning lives outside it, and where that applies. Two companion pieces go deeper into specific uses: finding the contradictions no single paper states and hypotheses you can check step by step.
What a language model is actually doing
A language model is very good at one thing: given text, produce the text most likely to follow. Trained on enough of the scientific literature, it produces text that sounds like the scientific literature. It can summarise a field, suggest a mechanism, propose an experiment.
What it cannot do, by construction:
- Hold a field at once. A model sees what fits in its context window. A field is hundreds of thousands of papers, datasets and failed experiments. Whatever is outside the window does not exist for that answer.
- Keep what it concluded. Every conversation starts from the same weights. A conclusion reached on Monday is not a premise on Tuesday unless someone pastes it back in.
- Show which steps are true. The output is one fluent piece of text. There is no seam between the claim that came from a measured result and the claim that was simply the most likely continuation.
- Give the same answer twice. Sampling makes the output vary. A chain of reasoning that cannot be reproduced cannot be reviewed in the way science reviews things.
None of this is a defect to be fixed in the next release. It is what a model that predicts text is. The problem is asking it to be the whole system.
What changes when the reasoning lives outside the model
GitMir's work starts from a different arrangement. The reasoning is not a paragraph the model writes; it is a structure the system builds and keeps:
- Claims, each with its source. A finding, a measurement, a mechanism, a negative result — each one held separately, each one linked to the paper, dataset or experiment it came from.
- Links that mean something. "Inhibits", "contradicts", "was measured under", "depends on" — relationships between claims, not proximity in a text.
- Reasoning as explicit steps. A conclusion is a path through those claims and links. Every step can be inspected on its own.
- Deterministic checks. The same structure and the same check produce the same result, every time. A reviewer can rerun it.
- Memory that grows. What was concluded, tested and refuted stays in the structure and becomes the starting point for the next question.
The language model still has a job: reading unstructured text, proposing candidates, turning a structured result back into language a scientist can read. It is one component inside a larger system, not the intelligence itself. We describe the effect of this split in our research on reasoning outside the AI.
Results a language model cannot produce on its own
The difference shows up as outputs, not as a better-written answer.
A conclusion that holds across the whole field. Because the claims of thousands of sources sit in one structure, a question can be answered against all of them at once — not against the dozen that happened to be retrieved.
A contradiction no single paper states. Two results that cannot both be true often live in different journals, years apart, in different vocabularies. Nobody contradicts anyone in either paper. The contradiction only exists between them. See the contradictions no paper states.
A hypothesis with its evidence attached. Not "compound X may help with Y", but the path: which measured effect, from which source, links to which mechanism, which links to which target — and which step is the weakest. See hypotheses you can check step by step.
The next experiment chosen on purpose. When competing hypotheses are explicit, the system can pick the experiment whose outcome would separate them most, instead of the one that is easiest to run.
A result that can be replayed. Months later, the reasoning that produced a conclusion can be rerun on the same inputs and arrive at the same place — or show exactly which new evidence moved it.
Where this applies
The approach is not specific to one science. It applies wherever knowledge is spread across many sources and a wrong step is expensive:
- Drug discovery and repurposing. Linking mechanisms, targets and compounds across literature and databases into candidate paths that a chemist can check link by link. We describe our work with the pharmaceutical industry on the GitMir Science page.
- Materials science. Connecting synthesis conditions, structure and measured properties reported across thousands of studies to find the combinations nobody has tried.
- Biology and medicine. Reconciling results from studies that used different models, assays and definitions, and marking where they genuinely disagree.
- Engineering research. Holding test results, failure analyses and design constraints together so that a new design is checked against everything already learned.
- Autonomous R&D. Closing the loop between hypothesis, experiment and result so that a system keeps learning between runs — covered in autonomous R&D that remembers what it learned.
What we publish and what we do not
GitMir Science publishes effects and measurements, not the internal representation that produces them. Scientific discovery is a research direction we are actively working on; the use cases above describe what the approach is built to do and where we are applying it, not a list of discoveries already made. When we have results, they will be published the way our benchmark is: with method, scope and limitations, so that they can be checked without us.
If you work on discovery problems where the reasoning has to be checkable — in pharma, materials, biology or engineering — we would like to hear from you through GitMir Science.
For background on why fluent but unsupported output is the core risk of language models in research, the overview of AI hallucination) is a useful starting point.
More on this, with the numbers behind it
What business logic is, and where it lives →Features, rules, customer paths and how they connect — why nothing a team already owns holds them in one place, and what that costs.Benchmark 001, in full →One repository, fifty questions, every system, the same AI model — and how to repeat it.Ask a published model your own question →A real product, read in full. Ask it what a change reaches and check the answer against the source it came from.Context engineering for coding agents →Why the window fills and the agent still misses the rule — and what to give it instead of files.GitMir Intelligence →Ask what a change reaches, what a rule holds and which customer path meets it — with the part of the product each answer came from.Put a number on your own workflow
Bring us the recurring workflow where AI still needs expensive people to supervise, review and correct. We baseline what it costs and put the target in writing before we build. Or connect a repository and see it on your own code first — 10,000 credits free, no card.
Build My Business Case → Start free →Frequently asked questions
Can a large language model make scientific discoveries on its own?
It can suggest ideas, summarise literature and propose experiments, but on its own it cannot hold a whole field at once, keep its conclusions between sessions, show which steps of its reasoning are supported, or give the same answer twice. Discovery needs all four, which is why GitMir builds the reasoning outside the model and uses the model as one component.
What does "reasoning outside the AI" mean?
The reasoning is an explicit structure of claims, sources and links that the system builds and keeps, rather than a paragraph a model writes. Each step can be inspected, the same check gives the same result every time, and what was learned stays for the next question.
Is the language model still used?
Yes. It reads unstructured text, proposes candidates and turns structured results back into readable language. It is a component of the system, not the whole intelligence.
Has GitMir already made a discovery with this?
Scientific discovery is a research direction we are actively working on. The use cases describe what the approach is built for and where we apply it. Results will be published with method, scope and limitations, like our benchmark.
Related articles
The Contradictions No Paper States
The most valuable finding in a field is often a contradiction no paper states. Why summaries hide it, and how holding every claim with its conditions brings it out.
Hypotheses You Can Check Step by Step
A hypothesis is only as good as its weakest step. Built as an explicit path, every link sourced, it can be reviewed in minutes, and the next experiment chosen on purpose.
Autonomous R&D That Remembers What It Learned
A research loop that starts every cycle from a blank state is automation, not research. What makes R&D autonomous, why agents alone fall short, and where it applies.