DeepPaperNote
single-paper deep reading workflow
A dense AI paper with objectives, attention maps, ablations, and claims
ALGORITHM · POLICY UPDATE
Attn(Q,K,V)=softmax(
QKT
√dk
+ B)V
𝓛(θ)=−∑t=1Tlog pθ(yt|x,y<t)+λ KL(πθ||πref)
sequence objective · policy gradient · regularized alignment
resolve
source
plan
results
review
Evidence before writing
Results become tables
Usable figures become images
domain
paper
status
A structured Obsidian note
Core metadata
Evidence-backed claims
Compact result table
Image-first figures
Deep analysis
Fig. 2 · training dynamics
Key formula
Âi = ri - mean(r)
∇θJ = 𝔼[Âi∇logπθ]
∇θJ = 𝔼[Âi∇logπθ]
paper.md + images/
DeepPaperNote
Turn a complex paper into an Obsidian note you will actually want to keep.
Created by Huashu-Design