WisPaper is an academic search product with an unusually specific complaint about the tools it wants to replace. Keyword search, its makers argue, can tell you that a paper contains the string “LLM”, but it cannot tell you whether the paper actually addresses your question. So the product puts a trained agent in the middle: it reads your question, writes its own search queries, then reads the candidate papers and decides which ones genuinely answer you.
That is a bigger claim than “faster search”, and it comes with numbers attached. The homepage advertises 95% literature retrieval accuracy over 500 million metadata records and 130 million full-text articles, across 32 disciplines, refreshed by 50,000 new items a day. There is a peer-reviewable paper behind it, four paid tiers, and a Singapore company called ATOM INFINITE PTE. LTD. on the footer.
This article works through all of it: what the three modules do, how the underlying model was trained and what it actually scores, every price and credit allowance on the pricing page, the arithmetic those allowances imply, and a piece of original analysis nobody else has published — what the product’s own sitemap says about how it is being marketed. If you are choosing research tooling for a team, the pricing and limits sections are the ones that decide it, and the AI strategy question sits underneath the whole exercise.
Table of contents
- What WisPaper Actually Is
- The Three Modules Inside WisPaper
- WisModel: The Agent Behind Deep Search
- What the WisPaper Benchmarks Actually Show
- WisPaper Pricing: Free, Plus, Pro and Max
- The Credit Arithmetic Before You Subscribe
- What WisPaper’s Own Sitemap Reveals
- TrueCite and the Citation Trust Problem
- WisPaper Against Google Scholar and Zotero
- Who Should Use WisPaper
- The Limits Worth Knowing Before You Rely on WisPaper
- References and Further Reading
What WisPaper Actually Is
The short description is a research workspace that tries to own the whole path from question to manuscript, rather than one step of it.
An agent, not a search box
The vendor’s own framing is “Your AI Academic Agent” with the tagline “Read Less Get More”. The FAQ draws the distinction plainly: normal search matches a keyword like “LLM”, while what it calls intent validation interprets the underlying research logic — the example given is “failure cases of LLMs in healthcare” — and then pre-reads and filters the results on your behalf.
It is desktop-first and workspace-shaped
The product ships as a downloadable application with a hosted workspace behind it, and the navigation names five features: Search, Library, Agent, Feeds and QA. The pitch is that notes, citations and experimental records accumulate in a personal knowledge base that is reusable across papers and projects, rather than evaporating when a session ends.
The company behind it is small and Singaporean
The footer credits ATOM INFINITE PTE. LTD. and the only published contact is [email protected]. There is no investor page, no team page and no customer list on the site. For a tool that wants to sit at the centre of a research workflow, that thinness is worth noting before anything else.
| Stated figure | Value | What it is measuring |
|---|---|---|
| Literature retrieval accuracy | 95% | Unqualified on the homepage; no benchmark named |
| Academic literature metadata | 500M+ | Records, not full papers |
| Full-text articles | 130M+ | The subset it can actually read |
| Daily literature update | 50K+ | New items ingested per day |
| Primary disciplines covered | 32 | Breadth claim; the paper evaluates on 10 |
| Operating company | ATOM INFINITE PTE. LTD. | Singapore, footer copyright 2025 |
The Three Modules Inside WisPaper
The published paper describes a closed loop of three modules that share one paper database, and the loop is the actual product design rather than marketing structure.
Scholar Search runs in two speeds
Quick Search is conventional BM25 keyword retrieval on Elasticsearch, and the paper claims sub-20-millisecond latency for it. Deep Search is the expensive one: the agent decomposes your natural-language question into two to four Boolean sub-queries plus one to four weighted validation criteria, runs the retrieval, then judges every candidate paper against those criteria before showing you anything.
Library is a reference manager with embeddings
The Library module handles hierarchical folders with unlimited nesting, extracts metadata from PDFs automatically — title, authors, affiliations, abstract, venue, date, DOI — and adds embedding-driven semantic search over your own collection. It exports to BibTeX, RIS and EndNote, and synchronises with Zotero and Mendeley, which matters because nobody abandons a decade-old reference library on a whim.
AI Feeds is the retention mechanism
Feeds asks you to pick broad arXiv categories, then applies semantic matching against a profile vector built from your stated interests. It also learns implicitly: papers you save to the Library are treated as relevance signals and fed back through a pairwise learning-to-rank update. This is the module that turns a tool you visit into a tool that visits you.
| Module | Mechanism | Costs credits? |
|---|---|---|
| Quick Search | BM25 on Elasticsearch, sub-20ms | No |
| Deep Search | Agent writes Boolean queries plus weighted criteria | Yes |
| Library | PDF metadata extraction, embeddings, Zotero sync | Storage-bound |
| AI Feeds | Category filter plus profile-vector matching | No |
| QA | Grounded question answering over papers | Yes |
| Survey / Agent | Review drafting and multi-step research tasks | Yes |
WisModel: The Agent Behind Deep Search
The interesting engineering is a purpose-trained model rather than a prompt wrapped around a general-purpose one.
Two objectives, trained together
WisModel is fine-tuned on two jobs at once. The first is turning a messy research question into structured Boolean queries and weighted criteria. The second is judging a paper against those criteria, using a four-way taxonomy: support, somewhat_support, reject, or insufficient_information. Both produce structured JSON, which is what makes the agent loop programmable rather than conversational.
Supervised fine-tuning, then reinforcement learning
Training runs in two stages. Stage one is supervised fine-tuning on both objectives simultaneously. Stage two applies Group Relative Policy Optimization to the judging objective alone, with a blunt binary reward — one if the verdict matches the human label, zero otherwise. That reinforcement learning pass is aimed squarely at the hardest part of the problem, which is the middle of the scale rather than the extremes.
The evaluation set is genuinely substantial
The team assembled 2,777 authentic queries across ten disciplines in both English and Chinese, and had domain-expert PhD students hand-annotate 5,879 validation criteria against them. Whatever you make of the results, that is a real annotation effort, not a hundred cherry-picked examples.
What the WisPaper Benchmarks Actually Show
This is where the marketing and the paper stop agreeing, and the gap is the single most useful thing to understand before buying.
The judging accuracy is excellent
On paper-criteria matching, WisModel scores 93.70% overall against a field of frontier models that do markedly worse: Gemini3-Pro at 73.23%, Qwen3-Max at 72.82%, GPT-5.1 at 70.81% and DeepSeek-V3.2 at 66.82%. The margin is not subtle, and it widens dramatically on the partial-credit class — WisModel gets somewhat_support right 91.82% of the time, while Gemini3-Pro manages 15.9%.
The retrieval recall is the number to read twice
The same paper reports that on TaxoBench — a benchmark of 72 highly cited surveys averaging 354.5 citations each, drawn from eight computer science subfields — WisPaper achieves 22.26% recall, against an O3 baseline of 20.92%. Those are the honest end-to-end numbers for “did it find the literature”, and 22.26% is a long way from 95%.
The two numbers are not the same measurement
To be fair to the vendor, these measure different things. The 93.70% is how often the agent judges a paper correctly once it has the paper; the 22.26% is how much of an expert-built survey taxonomy the whole pipeline recovers from scratch. Both can be true. But the homepage’s unqualified “95% literature retrieval accuracy” sits much closer to the judging figure than to the recall figure, while using the word retrieval. Read it as a validation score, not a promise that you can stop looking elsewhere.
The improvement over the baseline is real but small
WisPaper’s 22.26% beats O3’s 20.92% by 1.34 percentage points, which is a relative improvement of about 6.4%. That is a genuine result on a hard benchmark and worth having. It is not a category break, and no amount of homepage styling makes it one.
WisPaper Pricing: Free, Plus, Pro and Max
The pricing page is refreshingly concrete: four paid tiers, a free tier, and every allowance stated in numbers.
Four tiers plus a free plan
Plus is $4.20 a month or $50 a year with 10,000 monthly credits and 10 GB of high-speed storage. Pro, badged Most Popular, is $15 a month or $180 a year for 200,000 credits and 50 GB. Max comes in X2 and X5 variants; Max X5 is $75 a month or $900 a year for 1,000,000 credits and 250 GB. The free plan gets 1 GB and earns credits through sign-in and onboarding tasks.
Credits are the real currency
Credits are consumed by Deep Search, Survey, QA and Agent tasks — everything the agent does. Quick Search and AI Feeds do not appear to draw on them. Critically, the page states that unused balance is cleared when the cycle ends, so credits do not roll over. A top-up pack of 10,000 credits is available if you run dry, at a stated discount of 12% or 50% depending on tier.
The free tier deliberately withholds the headline feature
The FAQ is explicit that free accounts can use most core features such as search and QA, but do not get access to the newest capabilities like AI Survey. That is the feature the marketing leans on hardest, so the free plan functions as a demonstration of the search engine rather than a trial of the product.
The annual discount is already baked into the monthly figure
The page advertises Save 15% for annual billing, and the comparison table labels it -15%. But the arithmetic is worth doing: $15 × 12 is exactly $180, and $75 × 12 is exactly $900. The monthly figure shown is the annual rate divided by twelve, not a separate month-to-month price. Paying monthly will cost more than the number on the card; the discount is displayed as the default rather than as an extra saving.
| Tier | Per month | Per year | Monthly credits | Storage |
|---|---|---|---|---|
| Free | $0 | $0 | Earned via tasks | 1 GB |
| Plus | $4.20 | $50 | 10,000 | 10 GB |
| Pro | $15 | $180 | 200,000 | 50 GB |
| Max X2 | Not shown | Not shown | Not shown | 100 GB |
| Max X5 | $75 | $900 | 1,000,000 | 250 GB |
The Credit Arithmetic Before You Subscribe
Nobody does this sum before subscribing, and it changes which tier is rational.
Credits per dollar exposes the real ladder
Divide the stated allowance by the stated price. Plus buys 10,000 credits for $4.20, which is 2,381 credits per dollar. Pro buys 200,000 for $15, or 13,333 per dollar. Max X5 buys 1,000,000 for $75 — also 13,333 per dollar, to the same rounding. Pro and Max X5 are priced at identical value; Plus is worse by a factor of 5.6.
Plus is a storage plan wearing a credit plan’s clothes
If credits are what you came for, Plus is close to indefensible: paying 3.5 times more per month gets you twenty times the credits. Plus makes sense only if you want the paid storage tier and barely touch the agent. Anyone who intends to run Deep Search or Survey regularly should skip straight to Pro.
Storage tells the same story more gently
Storage per dollar runs 2.38 GB on Plus, 3.33 GB on Pro and 3.33 GB on Max X5 — the same flat rate above Plus. The ladder is consistent: one cheap entry tier priced for casual use, then a single linear rate across everything above it. That is a defensible structure, just not the one the tier names imply.
What WisPaper's Own Sitemap Reveals
Vendor pages tell you what a company wants to be. Its sitemap tells you what it has actually been doing, and this one is unusually talkative.
The site is 17,233 URLs deep
The sitemap index lists eleven sub-sitemaps, mirrored across English and Chinese. Counted on 7 September 2026, the English side holds 8,620 URLs and the Chinese side 8,613, for a total of 17,233. For a product whose editorial output is 48 hand-written guides, that is a striking ratio — roughly 359 machine-generated pages for every page a human appears to have shaped.
The FAQ corpus is 4,043 templated pages per language
The largest section is not the blog but the FAQ: 4,043 URLs in each language. It arrived in two bulk drops rather than organically — 1,994 pages last-modified on 30 October 2025 and 1,999 more on 20 April 2026, plus 50 on 10 April. The templating is visible in the slugs. Of 4,043 questions, 3,595 begin with the word “how”, which is 88.9% of the corpus.
The combinatorics are legible from the URLs alone
Count the opening fragments and you can see the generator: how-to-use appears 244 times, how-can-i 206, how-to-find 181, how-can-ai 160, can-ai-help 110. Then count the endings and the second axis appears — in-a-specific-field 55 times, for-non-native-speakers 50, for-a-grant-proposal 44, for-early-career-researchers 30. It is a stem list crossed with a modifier list.
The blog ran hard for ten weeks, then stopped
The 2,065 blog pages are paper summaries with dates in their slugs, and the dates cluster into 37 active days between 28 February and 11 May 2026 — a mean of 55.5 posts per day, a median of 53, and a peak of 122 on 20 March. Then it ends. Nothing has been published there since 11 May, almost four months before this was written. The research section shows related strain: 2,368 pages resolve to 2,328 distinct base slugs, with 40 pages that are numbered duplicates of an existing one.
The 48 newest pages are a different animal entirely
The learn section is the outlier. All 48 pages carry a last-modified date of 3 September 2026, they are written for students rather than crawlers, and eight of them are about a single anxiety: “are chatgpt citations real”, “how to verify ai-generated citations before you submit”, “why ai invents academic citations and how to catch them”. Published three days after the TrueCite announcement, that is not a coincidence — it is a funnel being built.
TrueCite and the Citation Trust Problem
The most recent addition is also the most defensible thing the company has shipped.
What TrueCite does
Announced from Singapore on 31 August 2026, TrueCite takes a BibTeX file, validates every reference in it against real academic databases, and flags the ones that cannot be found. That is a narrow, checkable job with an unambiguous output, and it addresses a problem that is measurably getting worse as more manuscripts pass through a generative model on the way to submission.
Why the problem is structural, not careless
The company’s own explainer is unusually clear-headed about the mechanism. A language model generates a citation token by token from training patterns rather than looking it up, which produces three failure modes: complete inventions, mashups of real authors with titles they never wrote, and real papers cited for findings they do not contain. The third is the dangerous one, because it survives a casual existence check.
Verification is a better product than generation
There is a strategic read here. Anyone can add a chat box to a search engine; very few tools can tell you their output is checkable. Pointing the roadmap at verification rather than more generation is the right instinct, and it sits closer to the data science discipline of validating a result than to the content-generation race most academic AI tools have joined.
WisPaper Against Google Scholar and Zotero
The honest comparison is not against one incumbent but against the two-tool habit almost every researcher already has.
What it replaces
For discovery, the agent genuinely does something Google Scholar does not: it filters on whether a paper addresses your question rather than whether it contains your words. For reference management, the Library covers the Zotero basics — folders, metadata extraction, BibTeX and RIS export — and syncs with Zotero and Mendeley rather than demanding you abandon them.
What it does not replace
Institutional database access, interlibrary loan, and the 370 million metadata records that sit outside the 130 million full-text articles it can actually read. Nor does it replace subject-librarian judgement on an unfamiliar field, which is precisely where a 22.26% recall figure should make you cautious.
The switching cost is lower than usual
Because the Library syncs with the incumbents rather than replacing them, the risk of trying it is mostly the subscription and the time. That is a deliberately low bar to clear, and it is the strongest argument for evaluating it rather than reading about it.
| Job | Incumbent | Does the agent improve on it? |
|---|---|---|
| Finding papers by topic | Google Scholar | Yes — intent filtering beats string matching |
| Exhaustive systematic search | Scopus, Web of Science | No — recall is the weak number |
| Reference management | Zotero, Mendeley | Parity, and it syncs rather than competes |
| Staying current | arXiv alerts, RSS | Yes — implicit learning from saved papers |
| Checking a bibliography | Manual DOI lookup | Yes — TrueCite automates it |
| Paywalled full text | Institutional access | No — not a licensing route |
Who Should Use WisPaper
The tool has a clear centre of gravity, and it is narrower than the marketing suggests.
Researchers doing exploratory scoping
If your task is “what does the literature say about X” and you are early enough that you do not yet know the right search terms, the agent’s query-writing step earns its keep. This is the use case the whole design points at, and the one where a mediocre recall figure matters least.
Anyone who reads across languages
The bilingual English and Chinese evaluation set is not a footnote. A tool trained and measured on Chinese-language research queries is genuinely differentiated, and for anyone tracking work published in both languages that is a real reason to look.
Teams already worried about citation integrity
If your group is drafting with AI assistance, the TrueCite check is worth more than the search engine. Building bibliography verification into a submission checklist is the kind of unglamorous control that prevents an ugly correction later, and it fits naturally alongside existing intelligent automation in a research workflow.
The Limits Worth Knowing Before You Rely on WisPaper
None of these are disqualifying. All of them should be priced into the decision.
The headline accuracy figure is not the retrieval figure
Say it once more, because it is the thing most likely to mislead: 95% is not the number the paper reports for finding literature. On the benchmark built for exactly that job, the system recovers 22.26% of an expert taxonomy. Treat it as a strong first pass, never as a completed systematic review.
Credits expire, and the cheapest tier is a trap
Unused credits are cleared at the end of every cycle, so a quiet month is money gone. And the entry tier delivers 2,381 credits per dollar against Pro’s 13,333. If you subscribe at all, the arithmetic says subscribe at Pro.
The company is thin and the content operation has stalled
A Singapore private limited with a support email and no named team is a real dependency risk for something that holds your research library. The stalled blog — 2,065 pages, then silence since 11 May 2026 — is not proof of trouble on its own, but it is the kind of signal worth watching before you migrate a decade of references.
Your questions and your library leave your machine
Deep Search, QA and Agent tasks run server-side, which means your research questions and the papers you save transit a third party. The Fair Use Policy also warns that the service is designed for individual human use and that automated scripts or excessive volume may trigger temporary limits — worth reading if you had scripted access in mind.
The verdict
WisPaper is a genuinely interesting piece of engineering with an honest paper behind it and a homepage that oversells it. The agent architecture is real, the judging accuracy is excellent, the pricing is legible once you do the division, and the pivot toward citation verification is smart. Evaluate it on the free tier, subscribe to Pro if it sticks, and hold the 22.26% figure firmly in mind while you do. It belongs in the same shortlist as the other tools in our AI models and tools hub, and like all of them it rewards testing over trusting.
References and Further Reading
WisPaper — Your AI Academic Agent
WisPaper Pricing Plans and Credit Allowances
WisPaper: Your AI Scholar Search Engine (arXiv 2512.06879)
WisPaper: AI Academic Retrieval Platform
WisPaper Introduces TrueCite to Help Researchers Verify AI-Generated Academic References
WisPaper: The Leading AI Tool for Literature Review in 2025
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.