LLM-optimized research articles

Summary: Assume there are several years when AI assistants (“LLMs”) have super-human ability to understand and apply known ideas, but sub-human ability to generate genuinely new ones. In this window, research communication should shift substantially. Rather than writing papers optimized for human readers, we should expect researchers will explain ideas conversationally to an LLM, which produces a written artifact, which audience members then consume through their own LLM. This combines the adaptability of one-on-one communication with the scalability of text. The written artifact probably remains human-readable—for training stability, peer review, and keeping humans in the loop—but its conventions would drift toward an LLM-optimized style: high information density, negligible review of existing knowledge, greater length variability, non-linear organization, exhaustive referencing, and explicit uncertainty markers.

Premise

A cartoon history of communication might use audience size as a key organizing metric:

1. One-to-one (e.g., conversation): The oldest format. Even today, this is generally the most efficient way to communicate an idea to a specific person because it is adaptable. The listener can ask the speaker to speed up or slow down, or to explain points they don’t understand. But it doesn’t scale; the process must be repeated for each new listener.
2. One-to-few (e.g., lectures, speeches): These formats scale better, but as the audience grows, individual listeners can ask fewer clarifying questions, and speakers can less easily tailor the message to what the audience already knows. Personal tutors are substantially more effective than classroom teachers.
3. One-to-many (e.g., written documents, recorded video): These formats scale nearly perfectly because copying information is ~free, but efficiency falls substantially: there is no interactive feedback. (Readers can contact the author, but this reduces to one-on-one communication and doesn’t scale.)

This will be disrupted by modern AI assistants. Here are two disputable claims about them that I won’t defend:

• They are making incredibly rapid progress at understanding and applying ideas that are already understood, and it may not be long until they understand such things as well as any human. Certainly, their knowledge is vastly broader, and they probably know most widely known things better than most humans.
• Relative to this, they have struggled to create new understanding (“insight”, “ideas”).I will concentrate on “new ideas” within academic-style research, but I think this applies more broadly.^a For instance, their ability to discuss math and physics is stunning when they operate within well-understood areas, but they fail predictably when they move outside those areas—which becomes acute when you ask them to do something genuinely new.

AI assistants are a complicated package of tooling extending beyond the base large language model, but here I’ll just call them LLMs for short.

It is reasonable to suspect that producing new insight is just the next capability LLMs will develop, since they have acquired one stunning ability after another over the last few years. But I want to consider an alternative scenario:An implicit assumption for any of this to make sense is that “new idea” is a coherent concept, even if I can’t define it precisely.^b

1. LLMs will become superhuman at understanding and applying known ideas; but
2. Producing new ideas is a meaningful unsolved research problem that will not be solved through near-term scaling and will remain a real limitation for at least several years.

This window, though possibly brief, arguably enables a next step in the cartoon history of communication: LLMs serve as intermediaries for—and maybe replace—written documents as the primary medium for broadly communicating new ideas. As long as humans are a bottleneck for something important, you still need to get new ideas from one human mind to many others. But in this vision, the human who develops new ideas (“the researcher”) explains them to an LLM, and then another LLM transmits the ideas to many other humans (“the audience”, often including other researchers). By hypothesis, the LLM can understand the idea nearly as well as, or better than, any given human. This form of idea dissemination combines the efficiency of one-on-one communication with the scalability of written communication: each audience member can converse with their personal LLM, which can explain the idea rapidly and adaptively.

What probably happens

In particular, this framing suggests we move away from having researchers record their insights as written documents intended to be read by other researchers (academic papers), and instead have researchers communicate ideas to an LLM as efficiently as possible. Plausibly, this looks like a mostly traditional conversation, and it can be much faster than writing a document. Not only does it enjoy all the advantages of one-on-one communication, the researcher can also assume their student (the LLM) is infinitely patient and knows essentially all recorded human knowledge.

Likewise, when an audience member wants to learn the idea, they ask an LLM (which has consumed the idea and understood it deeply) to explain it. This likely would be even more efficient than having the original researcher explain the idea one-on-one, since the LLM will have more patience, more context about the audience member’s background knowledge, and more ability to connect the new idea to existing concepts.Here is some discussion of someone using LLMs to improve the write-up of a math paper even while holding the core technical result (the theorem proof) fixed.^c

So the idea leaves the researcher’s brain through an LLM, and enters the audience member’s brain through a potentially different LLM. What does the central link look like? For the foreseeable future, I suspect the intermediary will be some sort of written artifact at least vaguely similar to a human-written document. A few reasons:

• LLMs are currently trained on text originally intended for humans. I don’t think we know yet how to stably train them recursively at scale on artifacts produced by other LLMs.
• The raw conversation transcript with the researcher seems suboptimal for training new LLMs. Some sort of condensation should take place before pre-training on it or loading it into context.
• Even if it were possible, having “neuralese documents” be the core artifact on which other LLMs are trained seems bad for keeping humans in the loop. Likewise for massive databases of conversation transcripts, which would technically be human-readable but too information-sparse for people to feasibly consume.

There are further pressures to do this in the near future, where LLMs are not fully reliable:

• We need some form of feedback that the LLM has actually received and understood the idea, to avoid telephone-game degradation. It might be useful to “test” the LLM by having the researcher probe it with questions, as a teacher would test a student, but a more straightforward and reliable method is to have the LLM express the idea in its own words as compactly and concisely as possible, and have the researcher review it for errors and completeness; this plausibly looks like a written document.Of course, the simplest way for the LLM to describe the idea to the original researcher to prove that it understands might differ substantially from how the LLM would describe it to another LLM, since in the latter case one could draw on ideas/concepts/theorems that the human researcher does not know.^d
• Likewise, for the foreseeable future we will want a way for other humans to review the idea and decide if it’s good and worth disseminating (peer review). For that, they need something—ideally human-readable—that they can point to as the thing they are reviewing, even if talking to an LLM about the document, rather than reading it directly, is their primary interface.

OK, so then, despite my abstract framing at the beginning of this post, I am really just suggesting a desirable and maybe inevitable evolution of existing research papers:

• They are written by an LLM.
• They are intended to be read primarily by LLMs, but remain human-readable.
• They are usually consumed by humans indirectly through the reader’s personalized LLM.

LLM-optimized style

What specific differences from traditional scientific papers might we expect? Some possibilities:

1. A more compact writing style that assumes familiarity with essentially all existing recorded knowledge. No need to review standard arguments, define standard notation, etc.
2. Much less repetition of new concepts within the document, reserving it essentially only for disambiguation (error correction).
3. Much less “scientific storytelling” and intro-results-conclusion structure.
   • Note: I do not mean “no motivation”, “no intuition”, or “just the facts, ma’am”. The arguments for caring about the facts in a paper are important!
4. Freedom from writing conventions adapted to human working memory and tool access.
   • For instance, after displaying a formula for a certain quantity, a normal scientific paper might plug in particular values as an example and compute an illustrative number, but this is unnecessary for an LLM with instant calculator access.
5. A greater range of lengths: both massive synthesizing documents and bite-size notes (e.g., a single lemma). In particular:
   • Length need not be upper bounded by human attention span.
   • Length need not be lower bounded by the overhead of being processed by an academic journal or newspaper editor (like tweets).
6. Non-linear organization: Traditional papers are linear because that’s how humans read; an LLM-mediated document could be more graph-like, with the LLM choosing a traversal order based on what the reader already knows. There might be an explicit dependency graph.
7. More notational freedom. Idiosyncratic or optimized notation is fine because LLMs are much more capable than humans at translating between languages and styles.
8. More tolerance for ambiguity where context resolves it (LLMs are good at disambiguation), but more precision where it is useful.
   • For example, there may be more explicit code or formal math in lieu of pseudo-code or natural language descriptions; for the most part, the latter are aids to human consumption rather than the cleanest way to record the ideas.
   • Likewise, there may be more embedded raw data or detailed calculations as appendix-like material the LLM can query.
9. Explicit confidence/uncertainty markers on claims, since LLMs can track and propagate these in a way that would be burdensome for humans.
10. More negative results and failed approaches (useful for other researchers, lower overhead to include).
11. Exhaustive and detailed referencing each timeOf course, this only need be done for ideas that another LLM would not already know the origin.^e an existing nontrivial idea is invoked (e.g., pointing to a specific paragraph rather than an entire paper), which would be too laborious for humans to write and too overwhelming to read.

How close are we to seeing this? I suspect the process has already begun. Researchers are definitely using LLMs to produce and consume research papers; it seems inevitable that this will cause writing norms to drift: Papers will get longer and more detailed as the burden of writing goes down, and reviewers will naturally push back less against writing that is long and harder to read directly as long as their LLM assistant can understand it clearly.

Journals will resist or embrace this to varying degrees. There will be growing pains as procedures and incentives that rely on out-of-date assumptions are reformed. I’m pretty confident the benefits will greatly outweigh the costs.