Papira

From research papers to research decisions.

Research Landscape Understand the field
Opportunity Report Find evidence-backed gaps
Contribution Positioning Know where your work fits
Survey Foundation Export a survey-ready structure

Input Your idea + related-work PDFs

Related-work PDFs

PDFattention.pdf

Attention Is All You Need

Ashish Vaswani, Noam Shazeer, Niki Parmar, et al.

1. Introductionrecurrent models are hard to parallelize

3. Model Architecturemulti-head self-attention

5. ResultsBLEU 28.4 on WMT 2014 EN–DE

ProblemSeq2seq bottlenecks

ApproachSelf-attention

BenchmarkWMT EN–DE

ResultBLEU 28.4

PDFbert.pdf

BERT: Pre-training of Deep Bidirectional Transformers

Jacob Devlin, Ming-Wei Chang, Kenton Lee, Kristina Toutanova

1. Introductionunidirectional LMs limit context

3. BERTmasked language modeling

4. Experiments

ResultsGLUE 80.5 average

ProblemShallow bidirectional context

ApproachBERT pre-training

BenchmarkGLUE

Result80.5 GLUE avg

PDFresnet.pdf

Deep Residual Learning for Image Recognition

Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun

Abstractdeeper nets degrade training

1. Introductionresidual / skip connections

4. ExperimentsImageNet top-5 3.57%

ProblemDegradation in deep nets

ApproachResidual blocks

BenchmarkImageNet

Result3.57% top-5 error

PDFgpt3.pdf

Language Models are Few-Shot Learners

Tom B. Brown, Benjamin Mann, Nick Ryder, et al.

Abstractscaling unlocks few-shot learning

1. Introductionin-context learning without fine-tunes

2. Approach175B autoregressive LM

ProblemTask-specific fine-tuning

ApproachFew-shot prompting

Method175B transformer

DatasetCommon Crawl mix

Related work — stop collecting papers. Start seeing the landscape.

Columns Against what have these approaches been evaluated?

Rows What problems are these papers trying to solve?

Cells What approaches target these problems?

Evaluated onMachine translationWMT EN–DE

Evaluated onLanguage understandingGLUE

Evaluated onImage recognitionImageNet

ProblemSeq2seq bottlenecks

ApproachSelf-attentionBLEU 28.4

?Open

?Your opening

ProblemShallow bidirectional context

?Open

ApproachBERT pre-trainingGLUE 80.5

?Open

ProblemDegradation in deep nets

?Open

ApproachResidual blockstop-5 3.57%

ProblemTask-specific fine-tuning

?Open

ApproachFew-shot promptingin-context

?Open

Every cell is grounded — filled = evidence from a PDF · empty = a Research Gap

Path

Input · Idea + PDFs Output · interactive Coverage Matrix

Empty cells = Research Gaps

LongMemEval

Temporal

DMR

Long-term memory

Graphiti71%

Mem039%

Gap

Knowledge retrieval

RAG-Fusion79%

Gap

Self-RAG64%

Memory editing

Gap

ROME73%

Not ChatGPT

~~A reading list~~ → An interactive Coverage Matrix

~~Generic lit review~~ → Relative to your idea

~~“I think there’s a gap”~~ → Empty cells with evidence

You leave with

Landscape

Problems × competencies, interactive

Opportunities

Gaps ranked by evidence

Positioning

Where your idea fits

Survey export

Structure ready to write

Pay once · first free

Stop collecting papers. Start seeing the landscape.

Related-work PDFs · interactive Coverage Matrix · Gaps grounded in evidence