A paper no longer has to enter context as one long document. Paper Schema organizes problems, methods, experiments, and evidence into a structure agents can retrieve, compare, and open on demand.
Agents do not need more text. They need an argument structure they can keep working with
Papers are still written for continuous human reading. Titles, abstracts, methods, experiments, and discussions form a long narrative, while readers must reconstruct the actual research question, core contribution, method components, and evidence relationships for themselves.
Full-text chunks answer where something appears, but not what problem the paper solves, what the method contains, or how the experiments support the claims. Research agents repeat this reconstruction during every close reading, comparison, and review, consuming context each time.
Paper Schema is not another summary. It represents a paper as an intermediate structure that can be retrieved, composed, and verified while preserving the path back to source paragraphs, tables, and citations.
“Turn understanding a paper from one-off contextual reasoning into a reusable data contract.”
It sits between bibliographic metadata and full text
Each paper becomes five connected object types. An upstream agent no longer has to infer hidden structure from a long document; it can select the material required for the task.
- Paper
- The topic, research question, core contribution, and main results.
- Entity
- Scientific objects such as Problem, Component, Finding, Measure, and Resource.
- Relation
- Within-paper relationships such as uses_component, evaluates, supports, and resolves.
- Evidence
- Equations, tables, experimental results, comparisons, code, and data resources.
- Provenance
- Precise links from structured objects and evidence back to their source locations.
Paper Schema is neither a replacement for full text nor a black-box writing system. It supplies atomic research material that agents can compose for a task.
Do not compress the full text. Let the agent decide what to open next
An agent can begin with the smallest useful representation of a paper and add context only when the question requires it.
- 01Start with the paper cardTitle, topic, research question, core contribution, and main results.
- 02Open structured objectsRead only the Problem, Method, Dataset, Measure, or Finding relevant to the task.
- 03Inspect relations and evidenceSee how methods address problems, experiments evaluate methods, and results support claims.
- 04Return to selected source textFetch the supporting paragraph and nearby context only for evidence that will be used.
Discover papers, read their structure, verify evidence, and prepare research material
The public contract currently contains 18 operations, organized into six capabilities through nine agent-oriented tools.
- Discover papers
- Find papers with completed Schema extraction by keyword, DOI, author, venue, year, or resource status.
- Read objects
- Inspect Problem, Component, Finding, Measure, and Resource objects within a paper.
- Understand relations
- Query logical and experimental relationships among objects within a paper.
- Verify evidence
- Search equations, tables, results, and resources, then use provenance to open a limited amount of source text.
- Inspect citations
- Retain the complete reference list and expand a bounded citation graph over resolved papers.
- Prepare material
- Build agent-ready material packages for survey, benchmark, method, and reproduction tasks.
The value appears when structure powers the next step
These four scenarios are not additional generative endpoints. They are outcomes produced when different agent workflows consume Paper Schema objects, evidence, and material packages.
Structured close reading of one paper
A researcher wants to quickly understand the problem, method, experimental design, and main conclusions of a paper.
The agent reads the paper overview, selects Entities relevant to the current question, follows Relations across methods, experiments, and conclusions, and fetches source text only for decisive Evidence. The result has a stable structure, so follow-up questions do not require reloading the entire paper.
- Select objects for the task
- Follow the argument through relations
- Trace conclusions to source evidence
Cross-paper method comparison and reproduction
Papers describe methods, data, metrics, and experimental conditions in different ways, making direct comparison inconsistent.
Method and reproduction packages place Problem, Method, Component, Dataset, Setup, Measure, and Evidence in one comparison frame. Missing code, data, hyperparameters, or experimental context becomes an explicit gap that can feed a reproduction checklist.
- Align comparison dimensions
- Separate methods from conditions
- Expose reproduction gaps
Write a research review with Claude Science, InternLM Scholar, or another agent platform
A researcher starts with a question and needs a well-structured review with reasonable coverage and verifiable evidence.
On an agent platform such as InternLM Scholar, install the Paper Schema skill. Broad metadata or semantic retrieval discovers candidate papers; Paper Schema handles structured reading, related-work expansion, survey material assembly, and evidence location; the research agent then organizes method families, disagreements, evidence strength, and research gaps.
- Broad retrieval provides recall
- Paper Schema provides structure and evidence
- Research agents synthesize and write
Build an evidence-backed topic research graph with FrontierLens
A researcher starts with a question. Rather than generating a global knowledge graph in one step, FrontierLens builds a bounded session graph by retrieving seed papers and progressively opening paper structure, cross-paper connections, and traceable evidence.
FrontierLens calls Paper Schema to search papers and read Paper, Entity, Relation, Evidence, and resolved Citation objects. Users first identify research paths in the topic graph, then open a paper's structure and relations, and finally follow Evidence provenance back to source passages. Sources and edge semantics stay visible across discovery, understanding, and verification.
Shannon4Science/sciverse-frontier-lensStart with a research question and generate a session reading guide
After entering “What methods improve Transformer inference speed?”, FrontierLens plans the search, retrieves seed papers, and organizes a card-based reading path. The researcher can see graph size, guided-paper count, and topic boundaries together before deciding whether the session covers the intended scope.
Inspect paper communities and research paths in the relation graph
The relation view organizes papers and research objects into a filterable network. Users can zoom, control category visibility, inspect the source behind a node or edge, and export the current session subgraph to identify method communities and connecting paths.
Open a paper and read its structure before its full text
After selecting a paper, FrontierLens reads its Entity and Relation objects from Paper Schema and places method components, core contributions, experimental setups, key findings, and measures in one paper-level structure graph. The guide highlights what matters while the graph supports deeper inspection.
Follow Evidence and provenance back to the source
Selecting an Entity or Evidence opens the source panel and uses paragraph or marker provenance to locate the supporting passage. When exact provenance is unavailable, the product explicitly labels approximate search or related-version status instead of presenting inference as quoted fact.
FrontierLens builds a bounded session graph around the current research question. It is not a globally normalized, conflict-governed, and continuously maintained domain knowledge base. Related Suggestions are discovery leads rather than factual relations, and the Citation Graph contains resolved paper edges only.
Trustworthy use begins with knowing what it does not do
- 01Coverage currently centers on 1M+ AI conference papers with completed Schema extraction; a miss does not mean the research does not exist.
- 02Cross-paper Entity discovery indicates similarity or relevance, not completed global identity resolution.
- 03The Citation Graph contains resolved paper edges only and does not replace complete citation counts.
- 04Materials are bounded, task-oriented packages; consumers must inspect returned, total, and truncated.
- 05A complete systematic review still requires broader retrieval, while a domain knowledge graph requires additional normalization, governance, and quality control.