AI research workspace·v0.9 beta

Your research
library, made
intelligent[ai].

Upload your papers once and get a personal research assistant that has read every one of them. Every answer is grounded in your sources — every claim, traceable to a page.

Start your library →See how it works

$0 to startReads formulas, tables & figuresLibrary stays private

Global Chat

📚 entire library · 351 papers

Q Summarize the disagreements between these five papers.

Disagreement clusters around two interpretations of the MACE-reduction effect: Suzuki^[4] and Hernández^[7] argue a dose-dependent biological mechanism; Patel^[18] and Lin^[22] attribute the gap to cohort confounding (age, baseline BMI, prior MI). Lincoff^[11] is consistent with both readings — its stratified subgroup analysis is the place to look next.

/Trusted by researchers from

StanfordMITETH ZürichCambridgeMax PlanckUC Berkeley

/01 · WHAT YOU CAN DO

Everything you do
with papers, in one place.

Upload PDFs

Drag in papers. Organize in folders. Sync from Zotero or Mendeley.

We read every one

Full text, figures, tables, formulas, citations — all indexed and understood.

Ask anything

Plain language in. Grounded, cited answers out. Click any citation to verify.

SIX THINGS YOU CAN DO ONCE YOUR LIBRARY IS READY

[lib]

Chat with your whole library

Ask one question across hundreds of papers. Get a synthesis with citations to every source.

[fol]

Chat with a folder

Narrow to one project or topic. Same chat, focused answers from just those papers.

[pdf]

Chat with one paper

Open any PDF and ask about claims, methods, or that one figure you keep re-reading.

[lit]

Draft a literature review

Litericaai writes a structured draft, finds what's missing, and proposes the next paper.

[net]

See the citation network

How your papers reference each other — seminal works, clusters, outliers at a glance.

[team]

Share folders with your team

Invite collaborators. Ask questions together. Pin answers. Leave inline notes.

/02 · COMPREHENSION

Reads what other
AIs skip.

Formulas. Tables. Figures. Every layer of a paper — parsed, understood, explained.

4 · Method · Policy-Gradient Updates

p. 7 / 24

4.2 Objective

∇_θ J(θ) = 𝔼_π [ ∇_θ log π_θ(a | s) · A^π(s, a) ](4.2)

5.1 Benchmark results

Table 2. Top-1 / top-5 accuracy on ImageNet-21k.
Model	Acc@1	Acc@5	Params	FLOPs
Baseline	78.1	93.4	21M	4.0G
Ours-S	84.2	96.1	22M	4.1G
Ours-M	85.7	97.0	56M	12.3G
Ours-L	87.4	97.8	124M	28.9G

Figure 3. Attention rollout across layers 1–4. Concentration on tokens 5–6 in layers 1–2, dispersal by layer 4.

READING · 1 page · 412 tokens · 1 equation · 1 table · 1 figure

✓EQUATION 4.2 · UNDERSTOOD

Policy-gradient objective

Maximizes expected reward by adjusting policy parameters θ in the direction of higher-advantage actions. Variance-reduced via the advantage function A^π in place of raw returns.

type objectivefamily policy gradientcf. REINFORCE, A2C

✓TABLE 2 · PARSED

Best: Ours-L 87.4% top-1 @ 124M params

Linear accuracy gain with model size; diminishing returns past 56M. 7× FLOPs cost from S→L for +3.2 pts top-1.

rows 4cols 5best row Ours-Lvs. baseline +9.3 pt

✓FIGURE 3 · CAPTIONED

Attention concentrates on tokens 5–6 early, disperses by layer 4

Heatmap shows sharp peaks in layers 1–2 for content tokens, with broadening receptive field by layer 4. Consistent with finding in §4.2.

peak L2 · tok 5–6spread increasing w/ depth

/03 · LITERATURE REVIEW

A draft review,
written from your papers.

Pick a folder. Litericaai reads every paper, finds the through-lines, and writes a structured review — with every claim cited to a page.

← ReviewsAttention mechanisms in modern transformer architecturesGenerating✓ Completed·3 sources·~312 words↓ Export

Attention mechanisms in modern transformer architectures

Introduction

Attention reshaped how sequence models capture long-range dependencies3. The field consolidated around self-attention after1 showed recurrence could be discarded; pretraining work2 then proved attention layers transfer across NLP tasks.

Themes

Three themes emerge. Architectural minimalism1 prioritizes parallelism. Bidirectional context2 reframes pretraining as denoising. Soft alignment3 anticipated both, reframing translation as attention-weighted lookup.

Methods

All three works rely on dot-product or additive attention13, scale to large parallel training, and report ablations on heads and alignment depth. Evaluation centers on translation BLEU13 and downstream NLP benchmarks2.

Gaps

Two gaps stand out. The corpus is silent on the compute–quality trade-off — none report wall-clock or energy budgets. Second, the analysis of failure modes (attention sinks, head collapse) has not yet entered this slice of the literature.

Conclusion

The arc runs from soft alignment as an addition to RNNs3, to attention as the only mechanism1, to attention as substrate for transferable pretraining2. Future reviews should incorporate efficiency- and analysis-oriented work.

/04 · CITATION NETWORK

See how your sources connect.

Visualize how papers in your library reference each other. Seminal works pull to the center; clusters self-organize by topic; outliers float free. Hover any node to see what it cites and what cites it.

Seminal

Cluster member

Outlier

[Every answer is grounded in your library. Every claim is traceable to a specific page. Every workflow — from initial reading to final review — happens in one place.]

— THE LITERICA[AI] PROMISE

/05 · WHY IT MATTERS

Sits in the gap nothing else fills.

GENERIC AI CHATKnows the internet. Not your library.

×Hallucinates plausible citations

×No grounding in your sources

×Can't open the PDF you read last Tuesday

REFERENCE MANAGERSStore files. Don't read them.

×Metadata-only search

×No synthesis, no Q&A

×Citation networks are static

SEARCH TOOLSFind papers. Don't synthesize them.

×Per-paper, not per-library

×Returns lists, not answers

×Doesn't draft your review

LITERICA[AI]Grounded answers from your corpus.

✓Every claim links to a page

✓Synthesizes across hundreds of papers

✓Drafts reviews, maps citations

✓Your library stays private

/06 · BUILT FOR

Anyone who reads a lot.

[01]

Graduate students

Lit reviews, thesis chapters, qualifying exams. Start writing on day one instead of day ninety.

[02]

Academic researchers

Keep up with a fast-moving field. Synthesize across hundreds of sources, in plain language.

[03]

R&D and competitive intel

Technical and patent literature is a daily input. Make every PDF interrogable.

[04]

Journalists, analysts, consultants

Defensible, citation-backed answers from a corpus you curated yourself.

/07 · PRICING

Start free. Pay when
your library scales.

[free]

Starter

Try the workspace.

$0/ forever

+50 papers in your library
+100 chat messages per month
+1 literature review per month
+Citation network & comprehension
+Solo workspace

Get started

Pro

For serious readers.

$19/ month, per user

+Unlimited papers & folders
+Unlimited chat & literature reviews
+Frontier-tier reasoning models
+Zotero & Mendeley sync
+Export reviews to Word, LaTeX, BibTeX
+Priority processing

Start Pro →