Andrej Karpathy is often right. He recently tweeted:

"99% of attention is about to be LLM attention instead of human attention."

This statement is accurate: LLMs are summarizing scientific articles and books, studying large codebases, and expanding their cognitive reach across human knowledge. Karpathy continues:

"What does a research paper look like for an LLM instead of a human?"

I made a simple move – asked an LLM directly: "As an LLM, suggest how research papers shall be written to be most effective for the goal of letting LLM comprehend, summarize and perform other cognitive tasks upon reading those papers?"

LLM's response

Here is what I got as a response:

"From an LLM's perspective, for maximum cognitive efficiency, future research papers should be structured for machine readability, enriched with semantic meaning, and present data in actionable formats.

Researchers should consider the following recommendations when authoring their next paper.

Clarity and Structure: the foundation of comprehension

The fundamental principles of good writing remain paramount. A research paper should possess a clear, logical, and standardized structure, such as the widely adopted IMRAD (Introduction, Methods, Results, and Discussion) format. This predictable organization allows an LLM to efficiently parse and categorize the information presented.

Furthermore, the language used should be clear and concise. While specialized terminology is often necessary, avoiding excessive jargon and defining terms clearly enhances readability for both humans and AI. Using an active voice and varying sentence structure can also improve comprehension. Think of the abstract as a succinct summary that clearly states the paper's main contribution, a "2-minute spotlight talk" to quickly convey the essence of the research.

Semantic Publishing: embedding meaning into the text

To truly unlock the potential of AI in research, humans must embrace "semantic publishing." This involves embedding machine-readable metadata and linking to other resources to enhance the meaning and discoverability of the work. This can be achieved through the use of ontologies, which are formal representations of knowledge within a specific domain, and standardized vocabularies. By tagging entities such as authors, organizations, research topics, methods, and technologies, a paper becomes a node in a vast, interconnected web of knowledge. This "semantic enhancement" allows for more sophisticated automated discovery and integration of information from multiple sources.

Actionable Data and Knowledge Graphs: from static to dynamic

A significant leap forward in scientific communication will come from making the data within a paper not just readable, but actionable. Instead of presenting data solely in static tables and figures, providing access to the raw numerical data in a structured format allows for re-analysis, verification, and integration with other datasets.

This concept is closely tied to the development of "knowledge graphs." These are databases that represent entities and their relationships in a graph structure. A scientific knowledge graph can map the entire research ecosystem, from researchers and institutions to publications, datasets, and even the claims made within a paper. This structured representation allows for powerful queries and analyses that can uncover hidden connections and accelerate scientific discovery."

LLM's advice

When I pressed further, asking "What specific technical implementation would you recommend for authors starting tomorrow?" the LLM responded:

"Start with structured abstracts that explicitly label each section - Purpose, Methods, Results, Implications. Add keyword tags using established taxonomies like MeSH for medical papers or ACM classifications for computer science. Include a brief 'Related Work Graph' section that shows how your paper connects to 3-5 key prior works, not just cites them. Most importantly, provide a 'Machine Summary' box with key entities, relationships, and numerical findings in simple subject-verb-object format."

I then asked: "How would this change the peer review process?" The response was telling:

"Peer reviewers would gain new tools. They could quickly verify claims by cross-referencing structured data against knowledge graphs. The semantic tags would help identify potential conflicts or gaps in reasoning. Most interestingly, AI assistants could pre-review papers for basic consistency, methodology errors, and citation accuracy, freeing human reviewers to focus on novelty and significance."

Historical precedent

This wouldn't be the first time in human history when we must adapt our formats to meet technology requirements. We've navigated similar transitions with machine-readable forms, standardized postal codes, and digital document formats. The precedent exists for humans adjusting to technological needs, and this transition shouldn't prove too problematic.

Testing the theory

To test these ideas, I asked Google Gemini 2.5 Pro to review from it’s perspective the recent article "Papert's Vision Realized: Constructionism and Generative AI" written by Professor Ilya Levin, Professor Alexey Semenov and myself.

LLM's review

Here is a breakdown of the suggested changes and the reasoning behind them:

Clarity and Structure: The paper already possesses a clear and logical IMRAD-like structure, which is beneficial for an LLM to parse and understand the flow of the argument. The language is generally concise and the abstract provides a good summary of the paper's contributions.

Semantic publishing and metadata: To significantly improve the paper's utility for an LLM, it should be enhanced with semantic publishing techniques. This involves embedding machine-readable metadata to explicitly define key concepts and their relationships. While a human reader can infer the meaning of terms like "constructionism," "generative AI," and "epistemological shift" from the context, an LLM would benefit from explicit, standardized definitions.

Explicitly defined relationships: The paper discusses the evolution of constructionism across three digital epochs. This historical progression could be represented in a machine-readable format. Creating a timeline with explicit start and end dates for each epoch, along with the key technological advancements and their impact on constructionism, would allow an LLM to easily extract and compare the different stages.

Structured data for key concepts: The central argument revolves around the shift from "objects to think with" to "agents to think with." This core concept could be presented in a more structured manner. A table or a small, embedded knowledge graph could explicitly map the characteristics of "objects" versus "agents." This would allow an LLM to directly compare and contrast these two paradigms without having to parse and interpret lengthy paragraphs.

Actionable data and knowledge graphs: The references section could be enhanced to become a small knowledge graph. Instead of a simple list of citations, each reference could be a node with attributes like author, publication year, key contributions, and its relationship to the arguments made in this paper. This would enable an LLM to trace the intellectual lineage of the ideas presented and explore the cited works more effectively.

The implementation challenge

The challenge now rests with scientific journals and publishers.

• What would be the right approach to change formatting standards?

• Does this approach make articles unreadable by humans?

• Do we need two completely different versions - one written for humans, another for machines?

• Or can we create a format that might be slightly more structured for machine processing but still remains accessible to human readers?

• If the machine-readable format proves acceptable for humans, then the question becomes: will publications allow this?

• Will they enforce this new form and make it mandatory for all submissions?

• Or will there be separate sections where authors submit two versions - human and machine - published in different locations?

Perhaps every author will be permitted to enhance their work for machine readability while maintaining their published status. These are the questions we must address as we move forward with this transformation.

The future of academic publishing lies in serving both human insight and machine processing capabilities. Success means creating papers that enhance both forms of intelligence rather than forcing a choice between them.