This is where the real-time thinking happens: arXiv category coverage, eval design, prior-art hunting, and neurodivergence-inspired research on reasoning architectures — shared with the ADHD community. Got opinions on isolation discipline, corpus gaps, or just want to argue about the next eval problem — come argue with us live.
👉 Join the community → as a contributor, maintainer, early adopter, or just a member. One short form. We coordinate category-taxonomy contributions, eval problems, integrations, and adopter onboarding there.
Before Claude designs something new, it checks arXiv first.
Real papers, fetched over real HTTP, read in isolation so no source anchors another, converged into ONE recommendation — cited, with a first step and the known ways this has already gone wrong for somebody else. Not a search wrapper: search finds you sources, NeuroArxiv forces a decision grounded in them.
Reach for it before committing to non-trivial architecture, algorithms, or systems techniques — anywhere real prior art plausibly exists and the cost of guessing wrong is a rebuild, not a typo.
👤 Author: Udit Akhouri — github.com/UditAkhourii
Not "research vs. no research" — that's an easy win for anyone. The real question is whether NeuroArxiv's isolate-then-converge discipline beats a plain, capable agent with normal web + arXiv access and no special process. Same model, same 5 cross-domain problems (physics, applied math, quantitative biology, ML, statistics), three conditions, run independently. A sample of the web-search condition's citations was verified against the real arXiv API before anything below was scored — it's a genuinely grounded condition, not a strawman.
The headline isn't a ratio. Every transcript was re-read for one specific pattern: does the answer name a source it just cited and flag a real limitation in that source's own claim — not a generic risk, a documented weakness in the specific paper.
| Cold | Web + arXiv (undisciplined) | NeuroArxiv | |
|---|---|---|---|
| Problems with a source-skepticism flag | 0/5 | 0/5 | 5/5 |
| Total flags | 0 | 0 | 7 |
Zero vs. zero vs. seven, out of five problems each. NeuroArxiv caught a withdrawn proof it had cited and declined to rely on it. It caught a benchmark result validated at only one context length and flagged it before recommending the approach. Cold and web-search both produced real, reasonable answers — neither produced that.
On raw answer quality (specificity, risk quality), NeuroArxiv beat the
web-search condition by a modest 1.1x–1.3x, and lost on citation
breadth in 2 of 5 problems — arXiv-only search has a narrower net than
general web search, and that's reported, not hidden. Full scorecard,
per-problem transcripts, and every honest limitation:
EVALS.md · raw data: bench/deep-tech-eval-transcripts.md.
One line, no clone, no build step of your own:
npx github:UditAkhourii/neuroarxiv installWorks for Claude Code and Codex CLI. Both discover skills at
<agent home>/skills/<name>/SKILL.md and read the same frontmatter, so one
bundled skill serves both — with no flag, install picks whichever agents are
actually present on your machine.
| Target | Flag | Lands in |
|---|---|---|
| Claude Code | --claude |
~/.claude/skills/neuroarxiv |
| Codex CLI | --codex |
~/.codex/skills/neuroarxiv |
| Both | --all |
both, whether or not they're detected |
Pin a target explicitly when you want it installed regardless of detection:
npx github:UditAkhourii/neuroarxiv install --codexCLAUDE_CONFIG_DIR and CODEX_HOME are honoured if you keep those directories
somewhere non-default.
Restart the agent (or start a new session) and NeuroArxiv is live — in Claude
Code as /neuroarxiv "<problem>", in Codex by asking for it by name.
Prefer a full local checkout (editing the engine, running the CLI directly, contributing)?
git clone https://github.com/UditAkhourii/neuroarxiv.git
cd neuroarxiv
npm install
npm run build
node dist/cli.js installneuroarxiv "cache LLM completions across requests without serving stale answers"
neuroarxiv "leader election for a queue with flaky nodes" --papers 6Inside Claude Code, no install is required to try it once — the skill in
skills/neuroarxiv/SKILL.md runs the same
loop using WebFetch against arXiv's export API directly. Full flag
reference: neuroarxiv --help.
PROBLEM
│
▼
0. CATEGORIZE — map the problem onto 3-5 arXiv categories + search terms
│
▼
1. FETCH — real HTTP against export.arxiv.org, category by category
│ (no LLM call — deterministic, courtesy-rate-limited)
▼
2. DIVERGE — one isolated LLM read per paper, in parallel
│ (each sees ONE abstract, never the others)
▼
3. SCORE — relevance / practicality / rigor, per paper
+ CLUSTER — group by underlying architectural angle
│
▼
4. CONVERGE — pick ONE cluster as the recommended path, synthesize,
cite, name the first step, name the risk, list pitfalls
pulled from EVERY paper's limitation — not just the winner's
Convergence is the deliberate departure from open-ended research tools: NeuroArxiv doesn't hand back "here are 4 papers, you decide." It commits to one recommendation, states why the runner-ups lost, and names what to watch for even in the paths not taken.
Every claim traces to a fetched abstract — papers, ids, and links are real arXiv metadata, never invented. The read prompt is explicitly forbidden from quoting more than a few words verbatim, and the skill's anti-patterns section calls out hallucinated citations as the failure mode to watch for.
MIT
Skill: skills/neuroarxiv/SKILL.md ·
Eval methodology: EVALS.md