Research Paper Reader (ArXiv + HF Papers)

Reading papers is mostly downloading PDFs, getting lost in LaTeX, and losing context when you switch between papers. The agent fixes this: fetch a paper by ID, browse its sections, ask questions about specific parts, compare two papers side-by-side, queue interesting ones for later — without ever opening a PDF viewer.

What it does

Fetch arXiv papers by ID — clean text, LaTeX flattened, sections enumerated
Browse Hugging Face Papers (today's trending or by keyword)
Section-by-section reading: "explain section 3.2 of <paper>"
Cross-paper comparisons: "compare results in <A> and <B>"
Personal reading queue with priority and notes
Daily digest of papers in your watched topics

What you'll need

Web fetch for arXiv + HF Papers HTTP endpoints
Memory for reading queue + paper notes
Cron for daily digests
A list of topics you follow

Setup

1. Configure your topics

Send to Flowly

Remember: Research topics I follow (memory tag "research-topics"): - AI agents (multi-agent, plugins, tool use) - LLM evaluation (especially long-context benchmarks) - Embeddings & retrieval - Speech-to-text (low-latency, on-device) Skip topics I do NOT want surfaced: - Pure RL theory papers - Robotics - Cryptography

2. ArXiv fetch convention

Tell the agent the patterns once — it'll reuse them:

Send to Flowly

For any "fetch arxiv <id>" or "read paper <id>": 1. Fetch https://arxiv.org/abs/<id> — extract title, authors, abstract, sections 2. Fetch https://arxiv.org/pdf/<id> if PDF is needed; else use HTML render 3. Save to memory tagged "paper:<id>" with full text + structure 4. Reply with: title, authors, abstract, section list. Wait for me to ask for a specific section. For "compare papers <id1> <id2>": 1. Ensure both are loaded (fetch if missing) 2. Output a side-by-side table: problem, method, dataset, key result, limitations

3. HF Papers daily digest

Send to Flowly

Cron "papers-daily" weekdays at 9 AM: 1. Fetch https://huggingface.co/papers (today's trending) 2. For each paper, extract title + abstract 3. Score against my research-topics: - High match (3+ topic keywords): score 9 - Medium (1-2 keywords): 6 - No match: skip 4. Take top 5 5. For each, also include: arXiv ID, GitHub repo if mentioned, key claim 6. Send Telegram message: 📚 Today's papers 1. <Title> — score 9 — arxiv:1234.5678 <2-line summary, what's new> Repo: github.com/... ...

4. Reading queue

Send to Flowly

Conventions: - "queue arxiv <id>" — adds to my reading-queue with date + notes - "show queue" — list pending, sorted by priority + age - "done with <id>" — mark read, ask for a 1-line takeaway - "search my notes for <topic>" — memory_search across paper:* tags

Example session

"Fetch arxiv 2403.12345"

→ Title, authors, abstract, sections list arrive.

"Read section 4 of that paper"

→ Section 4 displayed clean (no LaTeX cruft), agent answers questions inline.

"Compare it to arxiv 2401.99999 — focus on data efficiency"

→ Side-by-side comparison table.

"Queue this one with note: relevant to plugin sandboxing"

→ Saved.

Tips

Section-first reading. Don't dump full paper text into context. Browse sections, read the relevant ones. Saves tokens, keeps focus.
Comparisons are where this shines. AI gets lost on long single reads but excels at "what's different between X and Y."
Queue with notes. Future-you with no context will thank you. "queue arxiv 2403.12345 — promising for case-tracker semantic search" is way more useful than "queue 2403.12345".
HF trending updates fast. Daily 9 AM digest catches the wave. Hourly is overkill; weekly misses the conversation.
Set a cap on the queue. Once it hits 30, the queue becomes guilt. Friday cron: prune anything older than 14 days unless starred.