Material and results · Cited reports
A cited AI report: which sentence should you check first?
AstaBrief turns research questions and paper excerpts into reports. One published training example shows why people, dates and conditions still matter when you follow a citation.
Follow published material into a report, then return from an important sentence to its paper. The report here comes from training data; BitShovel checked the sources and product entry, without running AstaBrief.
Connecting a question to concrete cases
A media student asks how to find opportunities in start-up projects. In Ai2's published training material, the question comes with 26 paper excerpts. One report organizes academic-industry connections, mentorship and portfolios into sections. Its case section retains a company acquisition and a student's job-search experience, with a reference to their original paper.
This report was generated by Claude 3.7-powered ScholarQA and is part of AstaBrief's training material. AstaBrief-8B, announced October 2, composes a cited report from a question and retrieved excerpts in one pass and powers Asta's Fast mode. Ai2 reports full-pipeline averages of 51.1 seconds for Fast and 178.5 for Thinking, with development evaluations mainly from 2025. These describe a route to an earlier draft, rather than a current model ranking or BitShovel timing test.
Sources and further reading
- Ai2: the AstaBrief announcementOfficial announcement · October 2, 2026The model's purpose, the roles of Fast and Thinking, and training and evaluation conducted mainly in 2025.
- Ai2: where the published training reports come fromOfficial dataset card · Pinned revision c8c9357Research queries collected through June 2025, reports from different models, judge preferences and the CC BY-NC 4.0 licence.
- The media-student question: original input and two reportsOfficial training data · First recordRecord dddca3a4-a844-5803-83fc-2af47dd8510c. This reading examines the Claude 3.7 / ScholarQA report, not AstaBrief output. The complete data file is large.
The people, dates and conditions behind a citation
The cited 2021 Music Den paper by McGee and colleagues supplies both cases. In 2018, a company refined its software with mentorship and industry connections and was acquired six months later. A student used a music magazine in her portfolio to help secure a job. The paper documents that incubator's practice. The report selects understandable examples; the reader still decides which claim they support.
| The fact found in the original material | What it helps explain |
|---|---|
| In 2018, one company received mentorship and industry connections and was acquired six months later. | How contacts and product revisions connected to this outcome, rather than a six-month success rate for start-ups. |
| One student used her music magazine in a portfolio that helped her secure a job. | How a demonstrable work became job-search material, rather than the employment effect for all participants. |
Sources and further reading
- The media-student question: original input and two reportsOfficial training data · First recordRecord dddca3a4-a844-5803-83fc-2af47dd8510c. This reading examines the Claude 3.7 / ScholarQA report, not AstaBrief output. The complete data file is large.
- McGee and colleagues: the Music Den practice accountOriginal paper · Published online March 3, 2021The 2018 company acquisition and a student's portfolio experience, documented as a specific practice account.
A report can contain different kinds of support
The same training report marks two later items about student projects and interdisciplinary labs as LLM Memory: model knowledge, rather than the supplied excerpts. The label does not prove them wrong, but points to a different verification task. Start with the sentence that matters to you and identify whether it rests on a paper, model knowledge or further interpretation.
Ai2 also cites a 2025 summarization study. Its Table 3 shows an abstract describing experimental participants after exposure to climate disinformation, while an older ChatGPT-4 summary shifts to present-tense claims about individuals in general. The scope changes with the people and tense. This is not an AstaBrief test; it illustrates why following a citation means checking the studied people and setting as well as the topic.
Sources and further reading
- Ai2: the AstaBrief announcementOfficial announcement · October 2, 2026The model's purpose, the roles of Fast and Thinking, and training and evaluation conducted mainly in 2025.
- The media-student question: original input and two reportsOfficial training data · First recordRecord dddca3a4-a844-5803-83fc-2af47dd8510c. This reading examines the Claude 3.7 / ScholarQA report, not AstaBrief output. The complete data file is large.
- Peters and Chin-Yee: broader claims in summariesOriginal study · April 30, 2025Table 3 compares original abstracts with older-model summaries, rather than testing AstaBrief or 2026 models.
Starting with your own material
For a new subject, start with a set of papers you are willing to revisit and a bounded question, such as: What roles do mentors and student-created works play in these papers? This is an editorial suggestion, not a query tested in Asta. Check the people, dates and conditions behind a key sentence before carrying it into your notes, writing or work.
The October 3 interface check found Thinking, powered by Claude, and Fast, powered by Ai2's open model, under Generate a report; signing in saves work. BitShovel inspected the menu without generating a report or establishing current prices, limits or Chinese performance. Developers can download the English-tagged Apache 2.0 weights, prepare excerpts in the expected prompt format and adapt the code. Training data uses CC BY-NC 4.0; this is not a finished one-click PDF product.
Sources and further reading
- AstaBrief-8B: model cardOfficial model card · Pinned revision 3a4e553Question-and-excerpt input, the English language tag, Apache 2.0 weights and access information.
- Ai2: where the published training reports come fromOfficial dataset card · Pinned revision c8c9357Research queries collected through June 2025, reports from different models, judge preferences and the CC BY-NC 4.0 licence.
- Asta: the online report entryProduct interface · Checked October 3The report entry and Fast/Thinking menu were inspected anonymously, without signing in or submitting a query. Current fees and limits were not established.
- ScholarQA Lite: report-generation implementationOfficial code · Pinned revision a962328One-pass generation, prompt preparation and output parsing for developers to adapt, rather than a finished one-click PDF product.
See how these changes connect
Connect the cited report to other materials: how do recordings, photos and papers become results you can keep using?
Relevant papers come first: explore how retrieval, selection and report writing share the work.
