This post was written for the Astera Essay Competition and published initially as a google doc. We are posting this in LessWrong as we felt that this community can provide us with critical feedback and help improve the ideas presented in the essay.

Disclaimer: LLMs were used to help frame and write parts of this essay. We did a major rewrite afterwards to better articulate our thoughts, significantly editing out the initial LLM parts of the essay.

Introduction

Scientific articles today take on a predictable form: an introduction summarising the field of study, methods and results describing the process and findings, and a discussion placing the work in the bigger picture. As useful as it has been, this form of publishing cannot scale with the sheer volume of scientific data being generated. In fact, we believe that traditional publishing will only increase noise, add pressure on scientists, increase inequality and inaccessibility of scientific knowledge, reduce generalization across fields, and devalue corrective mechanisms like peer review and replication studies.

Fortunately, there is a growing awareness of these problems and the need for creative solutions. Platforms like PREreview allow scientists to crowd-source peer review, while open access movements like eLife and Lifecycle Journal aim to change publishing completely. Registered reports, protocols, and negative results are also staking their claims as valuable research components.

These solutions, however, are bottlenecked by the scaling problem, in turn constrained by the traditional perception of a scientific article. In this essay, we argue for a ground-up reworking of the scientific information ecosystem, with a focus on scalability, transparency, open ownership, and engagement. We explore a new conceptual framework that reduces a paper into unique components and links them to make a holistic picture. The proposed system will have clear benefits over the current one: removing barriers to access, explicitly erasing retracted papers and their influences, clearly marking knowledge gaps, and importantly, shifting incentives from publishing papers to building knowledge. Finally, we develop experiments to test feasibility.

Distilling a scientific article into unique nodes

A scientific paper contains a set of empirical evidence (results) obtained through some methodology (methods), and an associated high-level theory (discussion). To build on current knowledge and strengthen confidence in the evidence and theories described, the paper references (builds on) other similar studies.

When viewed from this perspective, it is clear that the current form of scientific publishing is riddled with redundancies (Fig. 1). They become visible when one thinks of the unique aspect of a paper, be it a methodology, an empirical result, or a high-level theory, as a node in scientific knowledge. Each paper contains some unique nodes but also reuses pre-existing nodes. Results are usually unique to each paper, whereas broader theories repeat in the literature (Fig. 1).

We argue that a scientific paper broadly contains two types of information. The first is statements associated with methodology, empirical evidence, and high-level theories. The second is directional references that build, improve, or contradict existing knowledge. When reduced in this manner, scientific papers emerge as textualization of a knowledge graph, with statements being nodes (or vertices) and references being links (or edges). A graph-based representation automatically compresses information as nodes are no longer repeated (Fig. 1 → Fig. 2). Directionality of graphs combined with weighted links between nodes allows one to evaluate the confidence of a node. Additionally, hypotheses/claims can be refreshed based on new evidence (while preserving history), making the graph always up-to-date.

Representing scientific information as knowledge graphs

In the 1950s, German sociologist Niklas Luhmann pioneered a physical knowledge graph system called zettelkasten. In that sense, reimagining scientific information as knowledge graphs is not a novel idea. However, scaling it from personal knowledge management to representing the entirety of scientific knowledge has not been possible yet. Some of the reasons for this limitation are:

1. The current framework has a large historical momentum. Shifting to another system would require converting a century's worth of information stored as textual papers.
2. Building a knowledge graph requires balancing information density with readability. For example, how much information should be added to each node?
3. The closed and fragmented nature of journals makes it difficult to source and maintain even a simplistic knowledge graph, with papers as nodes and references as edges.

Most of these problems have now been solved by time and technology:

1. History: Digitization and the development of large language models (LLMs) allow us to convert textual papers and seriously consider changing the scientific publishing system.
2. Information density–readability tradeoff: The ability to maintain and navigate massive multi-modal databases will allow us to create accessible, yet information-rich nodes.
3. Distributed systems: Technologies like blockchain allow new ways of building and storing scientific knowledge in graphs that are open and distributed.
4. Incentives: Researchers have consistently demonstrated the need and desire for an open and transparent way of communicating science.

Overall, we think the time is ripe for a complete overhaul of the scientific publishing system. A knowledge graph hastens the communication of scientific results by removing the need to write long papers and instead, simply adding nodes to the graph. LLMs can fill the gaps by providing the context, caveats, and comparisons needed to read the knowledge graph. They can also add historical papers and make the system future-proof by allowing nodes to be converted back to traditional papers (Fig. 3).

Advantages of knowledge graphs

Open and reproducible science

With knowledge graphs, scientific contributions are no longer constrained by traditional publishing formats. Once a method is perfected or an experiment is completed, researchers can immediately “publish” it as a node and keep changing or improving it. Replication studies become a major advantage, with every independent verification strengthening the relationship between nodes. This exercise, often viewed as “low value” within traditional publishing, becomes an important feature within our proposed system. We foresee a similar shift for peer review. A one-time gatekeeping event becomes continuous, traceable acts of reviewing nodes, proposing new connections, challenging existing ones, and assigning confidence levels.

Confidence as a systemic property

In a knowledge graph, confidence is not a binary stamp (accepted/rejected) but a derived property. In a hypothesis/claim node, confidence arises from consensus of linked result nodes, which depend on the rigor of method and peer reviewer nodes. If a method node is found to be flawed or a result node retracted, that signal propagates through the graph and recalibrates downstream confidence. Fraudulent or erroneous work doesn't just disappear, it is explicitly removed and its influence on dependent hypotheses/claims becomes visible.

Living repository of knowledge

Unlike the traditional system, where connections are established only when a paper is cited, a graph allows connections to be proposed and updated continuously. Every node carries version histories, laying a track record of how scientific beliefs evolve (Fig. 3). A knowledge graph facilitates links between seemingly disparate fields and flags unresolved debates by revealing contradictory findings. The evolving nature of graphs also spur discovery. Sparse regions—with few nodes, weak confidence, or unresolved conflicts—become indicators of unexplored scientific questions.

Clear knowledge gaps for grant proposals

A structurally elegant property of the graph system is the clear indication of knowledge gaps. In today’s system, grant applicants justify the need for funding by building a narrative about what is known and not known, and why the gap matters. In a mature knowledge graph, this narrative already exists. Sparse or low-confidence subgraphs, unresolved contradictions, and hypotheses/claims without supporting results are all legible gaps. Funding bodies can directly query the graph and proposals can be anchored to specific nodes or missing connections. This makes funding requests rational and verifiable rather than rhetorical.

Decentralization and equity through open infrastructure

A knowledge graph can be built on distributed infrastructure like blockchain or similar append-only, consensus-verified ledgers. This ensures that no single body controls scientific knowledge. Every participant holds a copy, yet no copy is silently altered. A system like this would go a long way in addressing inequities in research capabilities. For instance, researchers in resource-poor institutions or countries no longer face paywalls or fees. Researchers at any level can contribute to the record and be acknowledged for the same. Such a distributed system also provides resilience, making knowledge less vulnerable to institutional collapse, political interference, or commercial decisions by publishers.

Incentive realignment

Perhaps the biggest advantage of this system is a shake up of incentives. The current system rewards novelty over rigor. Within the graph system, strengthening nodes and links is as valuable as, if not more than, adding new nodes. With a new emphasis on rigor and confidence over continuous novelty, career incentives may shift. Funding bodies and hiring committees could, in principle, evaluate a researcher's full contribution profile and not just their paper count.

Concerns

Converting existing papers

Centuries of scientific knowledge are currently locked in closed, unstructured publications like PDFs, paywalled journals, and physical archives. Transitioning this corpus into a structured knowledge graph is a formidable challenge along multiple dimensions. Financially, large-scale digitization and structuring would require sustained investment. Technically, extracting structured nodes and relationships from papers, especially older ones with inconsistent formatting, is a hard natural language processing problem, although LLMs have made this quite tractable. Legally, much of the existing literature is owned by publishers, meaning that even automated extraction could be hindered by copyright barriers.

Necessity of maintenance and oversight

An operational challenge for a distributed knowledge graph, especially one of this scale and openness, is ensuring that the graph does not get cluttered with redundant nodes, spurious hypotheses, and unfounded links.^[1] [2] For instance, duplicate nodes, spurious connections, and reviewer nodes can all be directed to signal confidence in favored results and hypotheses. Some technical solutions to this are blockchain immutability to preserve records of changes, semantic scoring to detect duplicate nodes, and endorsement of connections to prevent spurious links. Effective maintenance, however, requires humans in the loop. Curators can drive maintenance by merging, pruning, and organizing the graph, but this introduces hierarchy and editorial authority. We believe that while this is necessary and required, the goal is to have clear, transparent, and contestable criteria for graph maintenance.

Accessibility and equity paradox

The proposed system is more equitable than traditional publishing—distributed ownership, no paywalls, and no submission fees. But building and querying a sophisticated knowledge graph requires computational infrastructure, technological literacy, and reliable internet access, all of which remain unevenly distributed across the world. The practical barrier to contribution simply shifts from "Can you afford journal fees?" to "Do you have the technological capacity to participate?" Ensuring genuine equity requires deliberate investment in accessible interfaces, multilingual support, and capacity-building in under-resourced research communities.

Key experiments

Experiment 1: Node consensus among scientists

Can many scientists, working independently on the same set of papers, converge on a consistent set of nodes and connections? If there is high variance in how different researchers distill the same paper into graph components, the entire system becomes unreliable. To test this, one can recruit researchers across career stages and disciplines, give them a common set of papers, and ask them to independently produce nodes and connections. The key measures would be inter-rater agreement on node identity (do two researchers produce the same method and result nodes for the same paper?), agreement on connection validity (do they agree on which results support which hypotheses?), and the nature of disagreements, whether terminological (same concept, different label) or substantive (different interpretations of a given study claim).

Experiment 2: LLM-based graph read/write fidelity

Can LLMs be used to generate and read graphs? Two complementary capabilities need to be evaluated:

Paper → graph (extraction; Fig. 1 → Fig. 2): Given a set of scientific articles, can an LLM reliably identify methods, results, and hypothesis nodes and propose accurate directional connections? This requires a ground-truth benchmark—ideally, the graph produced in Exp. 1—against which the LLM output can be scored for node specificity (did it find all the nodes?), precision (did it hallucinate nodes?), and connection accuracy.

Graph → paper (synthesis; Fig. 3): Given a set of nodes and connections, can an LLM reconstruct a coherent scientific narrative? This will test whether the graph representation is genuinely lossless, i.e., does the structured form capture everything a paper communicates or are nuances and context lost?

These two directions form a compression–decompression test. If paper → graph → paper produces a reconstructed paper that scientists consider equivalent to the original, then the graph is sufficiently expressive. If important content is lost, then that loss needs to be characterized: is it recoverable or does it reflect a fundamental limitation of the system?

Experiment 3: A pilot graph on open-access literature

How can one gauge the benefits of a knowledge graph-based publishing system? One approach would be to construct a pilot graph from a bounded corpus, like a single active research area, using validated LLM extraction. Examples include small sub-fields with open-access papers, or companies like Arcadia that publish results in a specialized sub-field. The pilot graph would be used actively for a moderate period (6 months to a year) while data is gathered on how the graph is used by researchers. Some questions to ask are:

1. Knowledge synthesis: Is there a measurable reduction in the time required to synthesize relevant prior work and identify gaps? This practical metric would be compelling to funding bodies.
2. Creativity: When interacting with the graph, do scientists generate hypotheses that they believe would not have arisen from traditional literature scans? A well-connected graph will allow interdisciplinary syntheses that keyword-based literature scans obscure.
3. Behavioral novelty: Do scientists use the system in unanticipated ways? Unexpected usage patterns are a strong signal of genuine value. E.g., using a graph to find the "weakest link" in a hypothesis chain, or using confidence gradients to identify high-uncertainty areas ripe for experimentation.
4. Contribution diversity: Does the pilot attract contributions from people who do not publish prolifically, like early-career researchers, those in low-resource institutions, or from adjacent fields? If the graph truly lowers barriers to meaningful participation, then it should be evident in contributor demographics.

Acknowledgements: Ideas in this essay were catalyzed by conversations with Emma Lieker and Ben Iorio, with feedback from George Schaible.  

1. ^{^}

   This problem is common in open collaborative systems. For example, GitHub repositories with large contributor bases routinely accumulate duplicate issues, requiring dedicated maintainers just to keep the tracker coherent. In a global scientific knowledge graph, the problem would be orders of magnitude larger.

2. ^{^}

   Arguably, these problems are prevalent in the current scientific publishing system as well.