Hand Talos one arXiv link and it delivers a 3,000-character deep read in 90 seconds, covering method, experiments and critical discussion with cited evidence.
Project background
For researchers and graduate students who need to keep up with top conferences, close reading is a weekly routine. This run reads *Attention Is All You Need*, the foundational paper on attention mechanisms (arXiv 1706.03762, NIPS 2017). Reading a 15-page English paper closely and turning it into reusable notes usually takes TBD minutes. With Talos, the same job produced a roughly 3,000-character Chinese deep read in 1 minute 30 seconds.
The challenge
Close reading is hard in three ways: the English original is slow to get through, method details are scattered between the main text and the appendix, and numbers are easy to misread. General chat tools can produce a summary, but the output usually stays inside the conversation as a spoken recap. The reader gets no source they can trace back, and no notes they can file away. The goal here was a deep read in which every number can be traced to the original paper, and which can be kept.
Approach
Step 1: Hand the paper link to Talos
The reading capability came from a single natural-language install. Talos finds paper-reading skills in public repositories, assembles one, and it is ready to use; after that, each paper needs only a link. This run used two sentences of input:
I want to read papers now. Please search for good skills that let me upload a PDF paper and start analyzing it right away so I can read it. If you find one, install it directly.
https://arxiv.org/abs/1706.03762This flow has three key elements:
- A natural entry point: no commands to memorize; describe the capability you want in one sentence.
- Minimal input: the only trigger for a deep read is a link; there is no need to prepare an abstract or figures.
- Selectable mode: Talos first confirms how to read (deep read / engineering breakdown / multi-paper comparison / evidence audit). This run used deep read.
Step 2: Talos reads the paper automatically
After receiving the link, Talos runs the following steps automatically:
- Source retrieval: fetch the full-text rendering of arXiv v7 of the paper, and download the PDF of the same version for cross-checking.
- Material statement: state first what was read, which version, and which parts could not be read; training memory is never passed off as having read the paper.
- Section-by-section breakdown: work through keywords, a one-paragraph summary, basic paper information, the core problem and contributions, in-depth method analysis, experiments and results, and critical discussion.
- Evidence anchoring: tag every number and conclusion with a table or figure number, or a section location; give formulas their full derivation context.
- Cross-checking: compare numbers in the body against table values in both directions, and flag mismatches on the spot.
- Output and filing: the result can be saved directly as a deliverable file, and can also be output as a visual HTML long-form piece for archiving when needed.
End to end took about 1 minute 30 seconds.
Step 3: Structure of the output
The finished piece runs to about 3,000 characters in seven parts:
| Part | Content |
|---|---|
| Material scope | Which parts were readable and which were not, the paper version, and whether the venue was verified separately |
| Keywords and summary | 5 keywords plus a single-paragraph summary of 150 characters or less |
| Basic paper information | Title, venue and year, link, task domain, paper type |
| Core problem and contributions | Where the problem comes from, and the three contributions the paper states |
| In-depth method analysis | Overall skeleton, the scaled dot-product attention formula and the motivation for the √d_k scaling, multi-head attention, its three uses, positional encoding, the complexity comparison table for four layer types, and training configuration |
| Experiments and results | Main results table, ablation table, syntax-analysis transfer results, with inference settings |
| Critical discussion | Paper claims / evidence-supported / reasonable inference / open questions, kept in four separate layers |
Key numbers stay in the prose: the main results table gives EN-DE BLEU of 27.3 for base and 28.4 for big, and EN-FR BLEU of 38.1 and 41.8; the ablation section reports that single-head attention scores 0.9 BLEU below the best setting, and that reducing d_k to 16 drops the score to 25.1.
Results
| Metric | Manual close reading | Talos | Change |
|---|---|---|---|
| Time to draft | Similar to Talos (user-tested comparison, exact value TBD) | 1 minute 30 seconds | — |
| Output length | TBD | About 3,000 characters | — |
| Evidence anchors | TBD | Every item tagged with table/figure number and section location | Traceable |
| Evidence layering | TBD | Claims / evidence-supported / reasonable inference / open questions | — |
| Output format | Manually organized note document | Draft in the session, fileable as a deliverable and as a visual HTML long-form piece | Editable and archivable |
| Reusability | Reorganized for every paper | Install the skill once, then just drop in a link | High |
Talos produced the roughly 3,000-character Chinese deep read in 1 minute 30 seconds, covering method, experiments and critical discussion. User testing feedback: with a general AI chat, drafting speed was similar; the difference lay in output quality and delivery format.
Business impact
- Per paper: the path from opening the original to holding a readable Chinese deep read compresses to 1 minute 30 seconds.
- Repeat use: the reading capability is installed once and stays available; after that each paper needs only a link, so the marginal cost approaches zero.
- Information quality: every number carries an anchor to the original, so citations can return to a specific table and section; places where the paper contradicts itself are flagged proactively.
- Format flexibility: the draft can be filed as a deliverable file, or output as a visual HTML long-form piece for easier archiving and sharing.
What you can reuse
- Install the capability first, then drop in a link: set up paper reading with one natural-language sentence; after that a deep read needs only an arXiv link.
- State the reading mode up front: deep read, engineering breakdown, multi-paper comparison and evidence audit differ substantially; say which one you want before starting.
- Ask for sources: put "numbers must trace back to the original tables and sections" in the request; only then is the draft verifiable.
- File the draft: have Talos save the deep read as a deliverable file and build up your own paper library over time.
FAQ
Q: Do I need to download the PDF myself first?
A: No. Give Talos the arXiv link and it retrieves the full text itself, and downloads the PDF of the same version for cross-checking. If you already have a PDF, you can also upload it directly.
Q: How long does one paper deep read take?
A: In this case, 1 minute 30 seconds for about 3,000 characters. Length and time depend on the length of the paper and the reading mode you choose.
Q: Are the numbers in the deep read reliable?
A: Talos requires every number to be anchored to a table or figure number, or a section location, in the original. Where something cannot be read, it states that no direct evidence was seen rather than filling in a number. Contradictions inside the paper (in this case, 41.8 in the abstract versus 41.0 in the body) are flagged.
Q: Can this method be reused on other papers?
A: Yes. The core is "install the reading capability once + drop in a paper link". It works for arXiv papers in any field, or other public PDFs, and you can switch to engineering-breakdown or evidence-audit mode when needed.