Research assistant for PDFs
Read less.
Understand more.
Drop a PDF and get the key insights, source citations, and research directions — in seconds, not hours.
Upload your PDF
Drag and drop PDF here
or click to browse50 MB limit · Documents stay private
Example output
3 insightsTransformer attention scales quadratically
The paper demonstrates that self-attention complexity grows as O(n²) with sequence length, making long-document processing prohibitively expensive without approximation techniques.
Page 4 — Complexity Analysis
Sparse attention achieves 94% of full accuracy
Experiments on the PG-19 benchmark show that local-global sparse patterns retain near-full performance while reducing memory from 16 GB to 2.1 GB at 8k tokens.
Page 12 — Results
How do sparse patterns interact with retrieval-augmented generation?
Sub-quadratic methods plateau beyond 32k context
Linear-attention variants exhibit diminishing returns past 32k tokens, suggesting a fundamental information-density bottleneck unrelated to compute.
Page 19 — Discussion