Atlas logoAtlas
Log inStart your workspace
All reports

Chat With a Paper

A comparison of ChatGPT, NotebookLM, Anara, Jenni and Atlas across 10 paper-reading capabilities.

Abstract

This report compares ChatGPT, NotebookLM, Anara, Jenni and Atlas across 10 capabilities for working with research papers. It examines how the products connect conversation to source evidence, support interaction with text and images, organize concepts, and preserve annotations and references. The comparison records shared support for several reading workflows and identifies Atlas as covering all 10 criteria. Its recorded coverage brings citation explanations, paragraph selection, knowledge maps and agent-created annotations into the same paper workflow.

Contents

Introduction

Reading a research paper involves following an argument, understanding its methods, examining its evidence and deciding how its findings relate to a question. AI-assisted reading connects those activities to a conversation: a reader asks for an explanation, checks the supporting passage and develops a further question.

The interface determines how that conversation relates to the paper. Citations locate evidence, selection tools bring passages and figures into focus, maps represent relationships, and annotations preserve the reader’s developing interpretation. This report examines 10 capabilities across those activities and discusses what the recorded results mean for the reading workflow.

Comparison Framework

The comparison covers ChatGPT, NotebookLM, Anara, Jenni and Atlas using 10 criteria defined around working with a focal research paper. The capability matrix is the source for the product results discussed in this report. Each section explains the capability, its practical significance and the corresponding result.

Table 1. Research Paper Capabilities Across Five Products

✅ indicates support; ❌ indicates that the criterion is unmet in this comparison.

CriterionChatGPTNotebookLMAnaraJenniAtlas
1. Citations open PDF section✅❌✅✅✅
2. Citations explain support❌❌❌❌✅
3. Paragraph hover-to-chat❌❌❌❌✅
4. Integrated concept maps❌❌❌❌✅
5. Add cited papers to project❌❌✅✅✅
6. Add figures and tables directly to chat❌❌❌✅✅
7. Understand images and image-only papers✅✅❌❌✅
8. Read PDF alongside chat✅❌✅✅✅
9. Add highlights and annotations to the PDF❌❌✅❌✅
10. Agentically highlight and annotate the PDF❌❌❌❌✅

1. Citations Open PDF Sections

A PDF citation connects an answer to a specific location in the original paper. Clicking it opens the supporting section with its surrounding paragraphs, figures and page layout. For a claim about an experiment, the reader can return directly to the passage describing the results.

This connection makes source inspection part of the conversation. The reader can examine qualifications, sample descriptions and adjacent evidence while considering the answer. Returning to the original context is particularly useful when a short quotation captures only part of an argument.

ChatGPT, Anara, Jenni and Atlas support this capability in the comparison. These products provide a direct route from the answer to the paper, allowing verification to continue in the original document.

2. Citations Explain Their Support

An explanatory citation describes how the cited evidence supports a particular claim. For example, a statement about a treatment effect can be accompanied by reasoning that connects the reported comparison, outcome and study conditions to that statement.

The explanation makes the relationship between evidence and interpretation available for inspection. Citation research treats the quality of source support as an evaluation concern: the ALCE benchmark examines citation quality alongside answer quality. This provides a useful basis for asking whether the cited passage actually supports the associated statement.[1]

Atlas is the product marked as supporting explanatory citations. In this workflow, the reader receives both the source destination and an account of its relevance, bringing evidence inspection and interpretation into the same citation experience.

3. Paragraph Hover-to-Chat

Paragraph hover-to-chat exposes an action when the pointer moves over a passage in the PDF. Selecting that action adds the paragraph to the conversation, where the reader can ask about its terminology, reasoning or implications.

The selected passage gives a question an explicit subject. A reader working through a difficult methods section can bring the relevant paragraph into the conversation at the point where a question arises. This keeps the interaction connected to the reading position and the wording under discussion.

Atlas supports this interaction in the comparison. Its paragraph selection connects reading and questioning through an action on the paper itself, making passage-level discussion part of the PDF workflow.

4. Integrated Concept Maps

A concept map presents concepts and the relationships connecting them. In a paper, a map can organize a central claim, its supporting ideas and the conditions under which it applies. The resulting representation gives the reader a way to examine the structure of the argument.

Novak and Cañas describe concept maps as representations of knowledge built from concepts and meaningful relationships, including connections across different parts of a map. Applied to paper reading, this structure can make links between ideas explicit and give the reader an overview to revisit while studying individual passages.[2]

Atlas meets the integrated concept-map criterion recorded in the table through its knowledge maps. This result identifies a visual route through the paper’s ideas alongside the text-based reading and chat workflow.

5. Add Cited Papers to Projects

Adding cited papers to a project turns selected bibliography entries into sources available for further research. A reader encountering an unfamiliar method can add the paper that introduced it and keep both sources in the same research collection.

References often indicate where an argument’s definitions, methods or evidence originated. Collecting those papers creates a practical route for following that context and returning to it later. Keeping the sources together also preserves the material needed for subsequent questions about the research topic.

Anara, Jenni and Atlas support adding cited papers to projects. These products connect the focal paper’s reference list to an expanding source collection, supporting the transition from reading one paper to investigating its background.

6. Add Figures and Tables to Chat

Direct figure and table selection brings a complete visual object from the PDF into the conversation. A reader can select a results table or chart and ask about a comparison while retaining the headers, labels and structure that give its values meaning.

Scientific results are often organized spatially. A value belongs to a particular row and column; a plotted trend depends on its axes and legend. Selecting the whole object keeps those relationships available when formulating a question, especially when the relevant evidence spans several cells or visual elements.

Jenni and Atlas support direct figure and table selection in the comparison. Their workflows make the object itself the unit of discussion, connecting questions about results to the evidence displayed in the paper.

7. Understand Images and Image-Only Papers

Image understanding enables questions about information presented visually in a PDF. Image-only papers extend this requirement to scanned pages, where the document’s contents are stored as images. Questions may concern a chart, a diagram or text visible on a scanned page.

This capability matters when the information needed for an answer resides in the page image. Document visual question answering studies this problem directly by asking systems to answer questions from document images. The DocVQA research establishes document images as a distinct question-answering setting.[3]

ChatGPT, NotebookLM and Atlas meet the image-understanding criterion in the recorded comparison. This coverage supports a reading workflow that includes visual evidence and image-only source material.

8. Read PDF Alongside Chat

A side-by-side reading interface keeps the original PDF and the conversation visible together. The reader can inspect a passage while composing a question, then compare the answer with the page as the discussion develops.

Reading and interpretation involve repeated movement between source material and notes or explanations. Research on active reading examines annotation, navigation and spatial layout as parts of that activity. Keeping the paper visible provides a stable reference while the conversation changes.[4]

ChatGPT, Anara, Jenni and Atlas support reading the PDF alongside chat. In these products, discussion and source inspection can proceed within the same visible workspace.

9. Add Highlights and Annotations

PDF highlights mark passages of interest, while annotations attach a reader’s interpretation or question to a source location. A reader can highlight an assumption in the methods section and leave a comment explaining why it matters for their own work.

Persistent annotations retain both the observation and its context for a later reading session. They turn a paper into a record of the reader’s engagement, including unresolved questions and passages worth revisiting. Annotation is also one of the activities examined in research on active reading.[4]

Anara and Atlas support persistent highlights and annotations in the comparison. Their PDF workflows let readers return to earlier observations at the relevant locations in the paper.

10. Agent-Created Highlights and Annotations

Agent-created annotation applies a reading request directly to the PDF. A user might ask for key findings to be highlighted and their limitations annotated; the agent selects the relevant passages and places the resulting marks and comments in the document.

This connects an instruction with a set of source-anchored reading artifacts. The reader can inspect the proposed interpretation where the evidence appears and revisit the annotations during subsequent work. Its practical significance is the movement from a conversational request to persistent work on the paper.

Atlas supports agent-created highlights and annotations in the comparison. Combined with its manual annotation and citation capabilities, this provides a workflow for requesting, inspecting and retaining an annotated reading of the source.

Conclusion

The comparison shows shared support for core reading activities. ChatGPT, Anara, Jenni and Atlas connect citations to the PDF and display the paper alongside chat. Anara and Jenni join Atlas in collecting cited papers; Jenni also supports direct figure and table selection, while Anara supports manual annotations. ChatGPT and NotebookLM share the image-understanding criterion with Atlas.

Atlas covers all 10 criteria in the matrix. Its recorded combination connects source navigation, citation reasoning, passage and object selection, visual understanding, knowledge maps and persistent annotations. For the paper-reading workflow examined here, this breadth supports a continuous progression from asking a question to inspecting evidence, organizing an interpretation and retaining work on the source.

References

1. Gao, T., Yen, H., Yu, J., and Chen, D. (2023). Enabling Large Language Models To Generate Text With Citations. EMNLP.

2. Novak, J. D., and Cañas, A. J. (2008). The Theory Underlying Concept Maps And How To Construct And Use Them. IHMC technical report.

3. Mathew, M., Karatzas, D., and Jawahar, C. V. (2021). DocVQA: A Dataset For VQA On Document Images. WACV.

4. Active-Reading Study. Research paper hosted by Microsoft Research, examining reading, annotation, navigation and document layout.